The geoid for the Baltic countries

determined by the least squares

modification of Stokes’ formula

Artu Ellmann

Doctoral Dissertation in Geodesy No. 1061

Royal Institute of Technology (KTH)

Department of Infrastructure
100 44 Stockholm, Sweden

March 2004
 TRITA-INFRA 04-013
ISSN 1651-0216
ISRN KTH/INFRA/--04/013-SE
ISBN 91-7323-080-4

© Artu Ellmann 2004

Printed by
Universitetsservice US AB
Stockholm, Sweden, 2004
______________________________________________________________________
Abstract

Precise knowledge of the geoid contributes to the studies of the Earth’s interior, the
long-term geophysical processes and to oceanography. An accurate regional geoid
model, in particular, enables the user in many cases to replace the traditional height
determination techniques by faster and more cost-effective GPS-levelling.
In regional gravimetric geoid determination, it has become customary to utilize
the modified Stokes formula, which combines local terrestrial data with a global
geopotential model. The Dissertation is devoted to the determination of a high-
resolution geoid model for the three Baltic countries – Estonia, Latvia and Lithuania.
Six different deterministic and stochastic modification methods are tested. These are:
Wong and Gore (1969), Vincent and Marsh (1974), Vaníček and Kleusberg (1987) and
the biased, unbiased and optimum least squares modifications by Sjöberg (1984b, 1991,
2003d). Three former methods employ originally the residual anomaly in Stokes’
integral. For the sake of comparison these methods are expressed such that the full
gravity anomaly is utilised in all the six methods.
The contribution of different error sources for geoid modelling is studied by
means of the expected global mean square error (MSE). The least squares methods
attempt to minimise all relevant error sources in geoid modelling by specially
determined modification parameters. Part of the present study contributes to some
important computational aspects of the least squares parameters sn.
This study employs the new geopotential model GGM01s, which is compiled
from data of the GRACE twin-satellites. Three sets (one from each country) of GPS-
levelling points were used for an independent evaluation of computed geoid models.
Generally, the post-fit residuals from the least squares modifications are slightly smaller
(up to 1 cm) than the respective values of deterministic methods. This could indicate
that the efforts put into minimization of the global MSE have been advantageous.
The geoid model computed by the unbiased LS modification provides the “best”
post-fit statistics and it is thus preferred as the final representation of the joint Baltic
geoid. The modification parameters of this model are calculated from the following
initial conditions: (1) upper limit of the GGM01s and the modification degree of Stokes’
function are both set to 67, (2) terrestrial anomaly error variance and correlation length
are set to 1 mGal2 and 0.1°, respectively, (3) integration cap size is 2°. This approximate
geoid model is supplemented by separately computed additive corrections (the
combined topographic and atmospheric effects and ellipsoidal correction), which
completes the geoid modelling procedures. The new geoid model for the Baltic
countries is named BALTgeoid-04. The RMS of the GPS-levelling post-fit residuals are
as follows: 5.3 cm for the joint Baltic geoid model and 2.8, 5.6 and 4.2 cm for Estonia,
Latvia and Lithuania, respectively. This fit indicates the suitability of the new geoid
model for many practical applications.

Key words: geoid, Stokes’ formula, deterministic and stochastic modifications, least
squares, additive corrections, GRACE, Baltic.

i
______________________________________________________________________
Acknowledgements

I wish to express my gratitude to all who have contributed to the completion of this
dissertation.
First and foremost, I am grateful to Professor Lars E. Sjöberg, my Supervisor, for
his inspiring guidance, intellectual leadership and for creating a very stimulating
research “environment”. This dissertation is the culmination of four years of my PhD
studies at KTH. Lars has been a true mentor to me during this period. I am indebted to
him for innumerable discussions, his valuable comments, professional advice and help
through the process of developing the idea for this dissertation and making it a reality.
In particular, I also have to thank him for the effort he put into reading this dissertation
under significant time constraints; for thoughtful criticisms and suggestions, which
made this a more coherent reading.
I am grateful to Kami Forskningsstiftelse for the financial support they made
available for my graduate studies. This research could not be initiated and completed
without this support. This funding played a crucial role by allowing me to concentrate
my full effort on the studies. I am very thankful to Professor Krister Källström for the
interest he showed in my progress and to whom I always could turn for advice and help.
I would like to extend my gratitude to all members of Group of Geodesy for their
support. Special thanks are to my fellow PhD student Jonas Ågren, who read a draft
version of this document and provided constructive comments and suggestions. Many
impromptu discussions with Jonas contributed greatly to this work.
I would like to express my gratitude to my past and present roommates at the
KTH (in the order of appearance): Jonas Ågren, Addisu Hunegnaw, Marek Rannala and
Johan Vium Andersson, who enabled me to focus on my work and were always willing
to help me in practical problems. Many other colleagues provided useful advice through
informal conversations. I would like to thank: Señor Erick Asenjo, Drs.: Tomas
Egeltoft, Huaan Fan, Hans Hauska, Milan Horemuz, Ming Pan; and Mr. Ramin
Kiamehr.
Many thanks go to Gabriel Strykowsky and Rene Forsberg, Danish National
Survey and Cadastre (KMS) for providing the gravity data used in this work. I wish to
express my thanks to Priit Pihlak, Estonian Land Board; Janis Kaminskis, State Land
Service, Latvia; Eimuntas Parseliunas, Vilnius Technical University, Lithuania; for
providing the GPS-levelling data.

Last but not least, I would like to thank my nearest relatives and friends for their patient
and persistent encouragement over the course of my PhD studies.

ii
______________________________________________________________________
Table of Contents

Abstract i
Acknowledgements ii
Table of Contents iii
List of Figures vi
List of Tables vii
List of Acronyms and Abbreviations viii

PART ONE – The geoid for the Baltic countries 1

1. INTRODUCTION 3

1.1 Preamble 3
1.2 Purpose of the study and author’s contribution 4
1.3 Outline of Part One 6
1.3.1 Outline of Chapter 2 (PAPERS A, B and C) 7
1.3.2 Outline of Chapter 3 (PAPER D) 8
1.3.3 Outline of Chapter 4 (PAPERS E, F and G) 8

2. MODIFIED STOKES’ FORMULA WITH APPLICATION TO THE BALTIC 9

GEOID

2.1 Modification of Stokes’ formula 9

2.1.1 Wong and Gore (1969) modification method 12
2.1.2 Vaníček and Kleusberg (1987) modification method 12
2.1.3 Integration with the original gravity anomaly 13
2.1.4 Deterministic and stochastic modification methods 15
2.2 Expected global mean square error and modification parameters 16
2.3 Data and modification criteria 18
2.3.1 Geopotential models 18
2.3.2 Modification limits 20
2.3.3 Target area 22
2.3.4 Previous geoid models in the target area 23
2.3.5 Terrestrial data 24
2.4 Evaluation of the expected global mean square error 26
2.4.1 Deterministic methods (M = 67) 28
2.4.2 Simple modification method in conjunction with the full expansion of
EGM96 28

iii
 2.4.3 LS modification methods (GGM01s, L = M = 67) 30
2.4.4 Deterministic and LS modification methods (GGM01s, M = 95) 31
2.4.5 The expected global MSE, discussion 31
2.5. Numerical investigations 31
2.5.1 Discrepancies between the geoid models 33
2.5.2 Comparisons between regional geoid models utilising different GGM 34
2.5.3 Verification with GPS-levelling data 35
2.5.4 Statistics of a four-parameter fitting 36
2.5.5 Preferred modification method 38
2.5.6 Comparison with the NKG results 39
2.5.7 Assessment of the geopotential models 40
2.5.8 Practical considerations 40
2.6 Conclusions 41
Appendices 43
Appendix A. Equality from Sjöberg and Hunegnaw (2000, Eq. 3). 43
Appendix B. Models for gravity anomaly and data error degree variances 44
Appendix C. Fitting of geometric and gravimetric geoid models 46

3. DETERMINATION OF THE LEAST SQUARES MODIFICATION

PARAMETERS 49

3.1 Expressions of the least squares modification methods 49

3.2 Solving the LS parameters 50
3.3 Regularisation methods 51
3.3.1 Singular value decomposition 51
3.3.2 Truncated SVD 52
3.3.3 Tikhonov regularization 52
3.3.4 Discussion 55
3.4 Numerical investigations 56
3.5 Conclusions 58

4. ADDITIVE CORRECTIONS TO THE GEOID MODELS 59

4.1 General 59
4.1.1 Topographical elevations 60
4.2 Combined topographic effect 62
4.3 Total atmospheric effect 63
4.4 Correction for downward continuation of gravity anomaly 65
4.5 Ellipsoidal corrections 66
4.6 Conclusions 70

5. CONCLUSIONS, DISCUSSION AND FUTURE INVESTIGATIONS 71

REFERENCES 75

PART TWO – PUBLICATIONS 81

iv
 List of Publications 82

PAPER A:
Ellmann A (submitted) Testing of some deterministic and stochastic modifications of
Stokes’ formula: a case study for the Baltic countries. Submitted to Journal of Geodesy.

PAPER B:
Ellmann A (in print a) Effect of GRACE satellite mission to gravity field studies in
Fennoscandia and the Baltic Sea region. Proc. Estonian Acad. Sci. Geol., 53 (No. 2).

PAPER C:
Ellmann A (2002) An improved gravity anomaly grid and a geoid model for Estonia.
Proc. Estonian Acad. Sci. Geol., 51: 199-214.

PAPER D:
Ellmann A (in print b) On the numerical solution of parameters of the least squares
modification of Stokes’ formula. IAG Symp. Series. IUGG 2003. Springer Verlag.

PAPER E:
Ellmann A, Sjöberg LE (2002) Combined topographic effect applied to the biased type
of the modified Stokes formula. Boll. Geod. Sci. Aff., 61: 207-226.

PAPER F:
Ellmann A (in print c) A numerical comparison of different ellipsoidal corrections to
Stokes’ formula. IAG Symp. Series. IUGG 2003. Springer Verlag.

PAPER G:
Ellmann A, Sjöberg LE (in print) Ellipsoidal correction for the modified Stokes
formula. Boll. Geod. Sci. Aff., 63 (No. 3).

v
 ______________________________________________________________________
List of Figures

Fig. 2.1. Location of the target area. 22

Fig. 2.2. Distribution of the data points. 24

Fig. 2.3. Free-air gravity anomalies in the data area. 24

Fig. 2.4. Cumulative sum of the global MSE components. 29

Fig. 2.5. The Baltic gravimetric geoid model BALTgeoid-04. 32

Fig. 2.6. Discrepancies between the unbiased LS and Vaníček -Kleusberg geoid
models. 33

Fig. 2.7. Discrepancies between the unbiased and biased LS modifications. 33

Fig. 2.8. Discrepancies between the Vaníček-Kleusberg and Wong-Gore geoid

models. 34

Fig. 2.9. Discrepancies between the Wong-Gore geoid models utilising different
geopotential models (EGM96 is subtracted from GGM01s). 34

Fig. 3.1. Typical behaviour of the L-curve when solving the LS parameters. 54

Fig. 3.2. Least squares modification parameters bn versus the Molodenskii’s

truncation coefficients Qn. 57

Fig. 3.3. Behaviour of the LS modified Stokes function SL(ψ) versus the original
Stokes function S(ψ). 57

Fig. 4.1. Topographical elevations in the spherical harmonic representation. 61

Fig. 4.2. Combined topographic effect. 63

Fig. 4.3. Atmospheric effect for the unbiased LS modification method. 64

Fig. 4.4. Distribution of the ellipsoidal correction for the unbiased LS modification
method. 69

vi
______________________________________________________________________
List of Tables

Table 2.1. Main parameters of the deterministic and stochastic modifications of

the Stokes formula. 18

Table 2.2. The expected global mean square error for six different modification 27
methods of Stokes’ formula.

Table 2.3. Numerical statistics (STD and mean) of the comparison of the geoid
models and national GPS-levelling points. 36

Table 2.4. The statistics of verification of the geoid models with the GPS-
levelling data, after the four-parameter fit is applied. 37

Table 3.1. Expressions of the least squares modifications of Stokes’ formula. 50

vii
______________________________________________________________________
List of Acronyms and Abbreviations

BALTgeoid-04 Geoid for the Baltic countries, developed at the KTH, year 2004
BSL Baltic Sea Level
CHAMP CHAllenging Minisatellite Payload
ETRS89 European Terrestrial Reference System, year 1989
EGM96 Earth Geopotential Model (degree/order 360/360)
GGM01c GRACE Gravity Model GGM01c (degree/order 200/200)
GGM01s GRACE Gravity Model GGM01s (degree/order 120/120)
GOCE Gravity and Ocean Circulation Explorer
GPS Global Positioning System
GRACE Gravity Recovery and Climate Experiment
GRS-80 Geodetic Reference System, year 1980
IAG International Association of Geodesy
KMS Danish National Survey and Cadastre
LS Least Squares
MES Mean Earth Sphere
MSE Mean Square Error
MSL Mean Sea Level
NASA National Aeronautics and Space Administration, USA
NKG Nordic Geodetic Commission
NKG96 Nordic-Baltic geoid model, year 1996
r-c-r Remote-compute-restore
RMS Root Mean Square Error
T-SVD Truncated Singular Value Decomposition
TEG (University of) Texas Earth Gravity Model
VK Vaníček and Kleusberg (1987) modification method
WG Wong and Gore (1969) modification method

viii
Part ONE
______________________________________________________________________

The geoid for the Baltic countries

1
(this page is left intentionally blank)

2
Chapter 1
______________________________________________________________________
Introduction

1.1 Preamble
This section is meant for those not familiar with the research field of physical geodesy.
It presents a brief scientific background and motivation to the Dissertation.
Geodesy is the comprehensive term for studies of the size and shape of the Earth,
the precise location of positions on the Earth’s surface and the determination of the
Earth’s gravity field. Large-scale geodetic parameters, such as the spherical shape of the
Earth, were known long ago thanks to observations of various celestial bodies - the Sun,
the Moon, the planets and the stars. Together with astronomy, geodesy is among the
oldest sciences; it is doubtless the oldest geoscience. A thorough outline of its long
history is given by Vaníček and Krakiwsky (1986), covering the development of
geodesy from the accomplishment of the oldest civilizations - Sumerian, Egyptian,
Chinese, Indian and ancient Greek - to the impressive accuracy of up-to-date space-
borne techniques.
The expression “figure of the Earth” has different meanings in geodesy. The
oblate ellipsoid is one such figure of the Earth and the actual topography of the
continents is another. However, the topographical surface is rather irregular and
therefore cannot be described without certain mathematical approximations. Since
oceans cover about 70% of the Earth’s surface, it is natural to choose the mean sea level
as a representation of the Earth’s figure. This leads to the introduction of one of the
most important definitions in the geosciences – the geoid. The geoid is defined as an
equipotential surface of the Earth’s gravity field, (generally) inside the topographical
masses on land and more or less coinciding with mean sea level at sea. Due to
irregularities in mass distributions inside the Earth, the geoidal heights undulate with
respect to the geocentric reference ellipsoid. However, the deviation of the two surfaces
does not exceed ± 100 m, globally. Precise knowledge of the geoid contributes to the
studies of the Earth’s interior, the long-term geophysical processes (post-glacial
rebound, plate tectonics, mantle convection, etc.) and to oceanography.
The main aim of physical geodesy may be formulated as the determination of the
level surfaces of the Earth’s gravity field, the geoid in particular. Prior to the “space
era”, the geoid was poorly resolved from land surveying efforts and gravity mapping. In
particular, artificial satellites are very useful for detecting the long wavelength
component of the Earth’s gravitational potential. New developments and advances in
gravity field determination from satellite tracking have taken place in the last few years.
Two recent satellite gravity missions, such as CHAMP (CHAllenging Minisatellite

3
Payload) and GRACE (Gravity Recovery and Climate Experiment), have provided
global scale and very accurate gravity data. Up-to-date accuracy of the global geoid
modelling is a few decimetres. However, for many scientific and practical applications
more accurate geoid models with higher resolution are necessary.
The geoid plays an essential role in the geodetic infrastructure, as the topographic
heights and the depths of the seas are reckoned from it. Thus, many applications in
geodesy, geophysics, oceanography and engineering require geoid-related heights.
Traditionally, spirit levelling has been applied for accurate height determination. During
the last two decades, the increased need for refined geoid models has been driven by
demands of using the Global Positioning System (GPS) for height determination. More
specifically, GPS-derived geodetic heights (reckoned from a reference ellipsoid) must
be transformed into traditional heights, in order to make them compatible with the local
vertical datum. At discrete points, a traditional height is obtained by algebraically
subtracting the value of the geoidal height from the geodetic height. Consequently, for
the conversion and combination of these fundamentally different height systems, the
geoid model must be known to an accuracy comparable to the accuracy of GPS and
traditional levelling, i.e. a few centimetres.
Regional improvements of the global geoid models can be obtained by applying
Stokes’ formula. This formula, published by G. Stokes already in 1849, is one of the
fundamental relationships in physical geodesy. It enables the determination of the
separation between the geoid and geocentric reference ellipsoid from a global coverage
of gravity anomalies. For practical considerations, however, the area of computation is
often limited to a spatial domain (e.g. spherical cap) around the computation point.
Proposed originally by Molodenskii et al. (1962), the truncation error of the remote
zone can be reduced by introducing a modification of Stokes’ formula, which combines
the terrestrial gravity anomalies and the satellite-derived low-frequency component of
the geoid. This combination is very prosperous since some recent advances in
technology and geodetic theory have created preconditions for achieving 1-cm accuracy
for a regional geoid model.

1.2 Purpose of the study and author’s contribution

The main objective of the Dissertation is to compute a joint geoid model for the three
Baltic countries. In approaching this goal a number of different modifications of Stokes’
formula are tested and many essential aspects of geoid modelling are investigated.
Emphasis in this study is given to the stochastic modification methods. More
specifically, three least squares modifications by Sjöberg (1984b, 1991, 2003d) are
studied in great detail. This is accompanied with several numerical verifications and
analyses of state-of-the-art approaches, leading to some new insights and solutions.
Even though only one geographical region is treated here, an attempt is made to present
the subjects under study in a way, that could be considered as a guideline for application
of said methods in any given region.
Over the course of the present study the computational challenge to define the LS
parameters sn from an ill-conditioned linear system (As = b) was solved. Chapter 3
describes the applied methods, the computing codes will be made available to the public
in a user-friendly form.

4
 The author had the opportunity to work with Prof. Lars E. Sjöberg whose
proposals on studying and applying the least squares modifications are extensively
employed in the Dissertation. This study includes some new geoid modelling
procedures hitherto numerically unproved in practice. These methods are: the optimum
least squares modification of Stokes’ formula (Sjöberg 2003d) and the ellipsoidal
correction for the original and modified Stokes’ formula (Sjöberg 2003b and in print a,
respectively). Additionally, based on the ideas of Sjöberg (1993 and 1994), the formulas
of the combined topographic effects are derived (and numerically illustrated) for the
biased type of the modified Stokes formula (Sjöberg 1984b).
Most of the research contributions described herein have already been published,
or manuscripts are either “in print” or submitted for publication in peer-reviewed
scientific journals. Subsequently, the Dissertation consists of two parts: this review (Part
One) and seven publications, which comprise Part Two of the Dissertation. The
publications will be referred to as PAPERS A-G as follows:

PAPER A:
Ellmann A (submitted) Testing of some deterministic and stochastic modifications of
Stokes’ formula: a case study for the Baltic countries. Submitted to Journal of Geodesy.

PAPER B:
Ellmann A (in print a) Effect of GRACE satellite mission to gravity field studies in
Fennoscandia and the Baltic Sea region. Proc. Estonian Acad. Sci. Geol., 53 (No. 2).

PAPER C:
Ellmann A (2002) An improved gravity anomaly grid and a geoid model for Estonia.
Proc. Estonian Acad. Sci. Geol., 51: 199-214.

PAPER D:
Ellmann A (in print b) On the numerical solution of parameters of the least squares
modification of Stokes’ formula. IAG Symp. Series. IUGG 2003. Springer Verlag.

PAPER E:
Ellmann A, Sjöberg LE (2002) Combined topographic effect applied to the biased type
of the modified Stokes formula. Boll. Geod. Sci. Aff., 61: 207-226.

PAPER F:
Ellmann A (in print c) A numerical comparison of different ellipsoidal corrections to
Stokes’ formula. IAG Symp. Series. IUGG 2003. Springer Verlag.

PAPER G:
Ellmann A, Sjöberg LE (in print) Ellipsoidal correction for the modified Stokes
formula. Boll. Geod. Sci. Aff., 63 (No. 3).

Most of the aforementioned papers were conceived and written by the author of this
dissertation alone. Nevertheless, fruitful discussions, valuable comments and advise

5
from Prof. L. Sjöberg have stimulated the content of all the papers. PAPERS E and G
consist some developments and more details of the ideas, which are originally proposed
by Prof. L. Sjöberg. The first author described the study results and formulated some
suggestions, the validity of those was ensured and specified by Prof. L. Sjöberg.
The investigations included in the Dissertation are based on the experience gained
from work performed during several years. The following list is intended to give the
reader an overview of related works not included in the Dissertation:

Sacher M, Ihde J, Celms A, Ellmann A (2000) The First UELN Stage is Achieved,
Further Steps are Planned. In Gubler E and Hornik H (Eds): Report on the Symposium
of the IAG Subcommission for the European Reference Frame (EUREF) held in Prague,
June 1999, Veröffenlichungen der Bayerischen Kommission für die Internationale
Erdmessung, München. Heft No. 60, pp. 87-94.

Ellmann A (2001) Least squares modification of Stokes’ formula with application to the
Estonian geoid. Licentiate thesis, Geodesy Report No 1056, ISBN 91-7283-212-6;
viii+98 pages, ill.; Royal Institute of Technology, Division of Geodesy and
Geoinformatics, Stockholm.

Ellmann A (2002 a) Estonian Geoid Model by the Least Squares Modification of

Stokes’ formula. In Poutanen M and Suurmäki H (Eds.): Proceedings of the 14th
General Meeting of the Nordic Geodetic Commission, pp. 138-147, Kirkonummi-
Helsinki.

Ellmann A (2002 b) LEO satellites for Earth’s gravity field recovery. Satellite
Application Department papers, International Space University Summer Session
Program 2002 at the California State Polytechnic University, USA, in CD-ROM.

1.3 Outline of Part One

Part One consists the following five chapters:

1. Introduction,
2. Modified Stokes’ formula with application to the Baltic geoid,
3. Determination of the least squares modification parameters,
4. Additive corrections to the geoid models,
5. Conclusions, discussion and future investigations.

The first chapter is meant as a general introduction of the Dissertation. Chapters 2-4 are
presented in a self-consistent framework, comprising an overview of the applied
methods, summaries of the most important results and conclusions. The general outline
for those chapters is given in Sections 1.3.1-3. Chapters 2-4 are based on the PAPERS
A-G. In the other hand these chapters may contain some additional information, not
included in the original papers (due to space limitations often imposed by publishers).
In particular, our research utilises several approaches, which are theoretically rather
distinct from the methods traditionally applied for geoid determination in the area of

6
interest. Previous works for the same region are reviewed in Section 2.3.4. Since every
researcher is interested in comparing own results with the ones of the previous or/and
alternative approaches, we treated the most often used alternative method in the context
of the present study. This necessitated for conducting a number of complementary
investigations and computations. In order to unify the new results and the contents of
PAPERS A-G a detailed discussion is evolved in Chapter 2. Accordingly, Chapter 2 is
the most important (and the most voluminous) part of the Dissertation.
Wherever appropriate we refer to the related PAPERS for more detailed
information. The study results are concluded in Chapter 5, which contains a general
summary of all topics, a discussion of the most important results and some
recommendations for future research. A bibliography can be found at the end of Part
One.

1.3.1 Outline of Chapter 2 (PAPERS A, B and C)

Chapter 2 presents the theory for modification of Stokes’ formula, the numerical
outcome of different methods and tests of their accuracy. PAPERS A, B and C are
related to those subjects.
The modification methods proposed in the geodetic literature can be divided into
two distinct classes - deterministic and stochastic approaches. The deterministic
approaches principally aim at reducing of the effect of the neglected integration zone
only. No attempt is made to incorporate the accuracy estimates of the geopotential
model harmonics and terrestrial data, although the errors of both datasets are
propagating into the estimator of the geoidal height. In contrast, the stochastic
modification methods also attempt to minimise the data errors in geoid modelling.
Three deterministic methods (Wong and Gore 1969, Vincent and Marsh 1974 and,
Vaníček and Kleusberg 1987) and three stochastic modification methods by Sjöberg
(1984b, 1991 and 2003d, respectively) are applied for modelling the geoid over the
Baltic countries. Accordingly, PAPER A could be considered as an executive summary
of the most important findings of Chapter 2, although it describes only five modification
methods.
As is well known, an appropriate global geopotential model (GGM) is essential in
determining the regional geoid model accurately. It is thus important to validate the
quality of such models in a regional scale. The new GRACE Gravity Model GGM01s
was released in July 2003 by the Centre for Space Research at the University of Texas.
The suitability of this model for regional geoid modelling is tested. In order to assess
the improvements due to the GRACE data, the GGM01s-related results are compared
with those based on the Earth Gravity Model EGM96 (Lemoine et al. 1998). A detailed
comparison of the two geopotential models can be found in PAPER B, which consists
also the assessment of the contribution from some selected spectral windows of the two
geopotential models. Two sets of high-precision GPS-levelling data from Estonia and
Sweden were used for some comparisons.
The focus of PAPER C is given to the gravity data to be used for the geoid
modelling. The perfectness of any theoretical method is diminished or does even
become meaningless with insufficient data quality and coverage and improperly
selected gridding procedures. Subsequently, PAPER C describes several aspects of
gravity data analysis.

7
 Chapter 2 culminates with the estimation of the most appropriate modification
method for elaborating the geoid over the Baltic countries. The final selection of the
best modification method is based on two measures. Internal accuracy is estimated by
means of the expected global mean square error, whereas the GPS-levelling data is
applied for an external evaluation of the accuracy of the computed geoid models. A
brief conclusion follows. Additionally, Chapter 2 comprises some Appendices, which
present some complementary information.

1.3.2 Outline of Chapter 3 (PAPER D)

Chapter 3 and PAPER D aim at contributing to the important computational aspects of

the LS modification parameters sn. A set of LS parameters is determined by solving a
system of linear equations, aiming at minimizing all relevant errors of the geoid
estimators. Regrettably, for certain LS methods some numerical instabilities may be
encountered when practically computing the modification parameters. Then, in order to
obtain a meaningful solution, the tools of mathematical regularization need to be
applied.
Paper D covers only one regularization method, namely Tikhonov (1963)
regularization. This technique and a complementary method are reviewed in Chapter 3.
Both methods are sufficient for solving different cases, which may occur when
computing the LS parameters sn. Typical features of the parameters sn are illustrated by
numerical investigations.

1.3.3 Outline of Chapter 4 (PAPERS E, F and G)

Chapter 4 considers a number of corrections, which have to be applied in the geoid

determination process.
Geoid determination by Stokes’ formula requires that the disturbing potential be
harmonic on this boundary surface. This yields that the masses outside the geoid must
be absent. Furthermore, the gravity anomalies need to be given on the geoid. In modern
context, however, the gravity anomalies are referred to the ground surface. Moreover,
Stokes’ formula is valid on the sphere, whereas the Earth can be approximated by an
ellipsoid. Consequently, the original and modified Stokes’ formulas should also
comprise some correction terms accounting for the Earth’s ellipticity, downward
continuation of gravity anomaly and the contribution of atmospheric and topographic
masses. PAPER E is devoted to the topographic effects, whereas PAPERS F and G
concern the computation of the ellipsoidal effect.
There are many ways to account for the contribution of the above effects. In this
study these problems will be solved by applying so-called additive corrections (Sjöberg
e.g. 1993, 1994, 2003a,b,c). The method of additive corrections allows us, instead of
computing the co-geoid (by correcting the gravity anomaly prior to Stokes’ integration),
to apply the additive corrections to the approximate geoidal height.

8
Chapter 2
_____________________________________________________________________________________
Modified Stokes’ formula with application to the Baltic geoid

2.1 Modification of Stokes’ formula

George Gabriel Stokes published his well-known formula in 1849. According to Stokes’
theory the geoid determination problem is formulated as a boundary value problem in
potential theory. Hence, the gravitational disturbing potential T can be computed as

R
T=
4π ∫∫σ S (ψ ) ∆gdσ , (2.1)

where R is the mean Earth radius, ψ is the geocentric angle, ∆g is gravity anomaly, dσ
is an infinitesimal surface element of the unit sphere σ and S(ψ) is the Stokes function.
The orthogonality relations between Legendre polynomials Pn(cosψ) over the sphere
allow to present S(ψ) as a series

2n + 1
∞
S (ψ ) = ∑ Pn (cosψ ) . (2.2)
n =2 n − 1

The disturbing potential is the difference between the actual gravity potential of
the Earth (on the geoid surface) and the normal potential associated with a rotating
equipotential ellipsoid. Considering another famous interrelation of physical geodesy,
Bruns’ formula (cf. Bruns 1878),

T
N= , (2.3)
γ

the disturbing potential can be converted via normal gravity (γ on the reference
ellipsoid) into separation (N) between the geoid and geocentric reference ellipsoid.
Shortly, Stokes’ formula enables the determination of the geoidal height from the global
coverage of the gravity anomalies:

R
N=
4πγ ∫∫σ S (ψ )∆gdσ . (2.4)

9
 The original Stokes formula suppresses the spherical harmonics of degrees zero
and one in the potential T and is therefore strictly valid if these terms are missing. The
zero- and first-degree effects may occur, if the mass of the adopted reference ellipsoid is
not equal to the mass of the Earth and the origin of the reference ellipsoid is not
coinciding with the Earth’s gravity centre. These effects are customarily neglected (incl.
this dissertation) when computing a regional geoid model. If necessary, the zero- and
first-degree effects can be accounted for separately; see e.g. Heiskanen and Moritz
(1967).
The surface integral in Eq. (2.4) has to be evaluated over the whole Earth. In
practice, however, the area of integration is often limited to a spherical cap around the
computation point. Proposed originally by Molodenskii et al. (1962), the truncation
error of the remote zone can be reduced by a modification of Stokes’ formula, which
combines the terrestrial gravity anomalies and the long wavelength (up to degree M)
contribution from a geopotential model. Assuming a cap of integration σ0 around the
computation point, a simple modification method can be presented (cf. Heiskanen and
Moritz 1967, Ch. 7-4):

M
R
N = ˆ σ + c ∑ Qn ∆gˆ n ,
∫∫ S (ψ )∆gd (2.5)
4πγ σ0 n =2

where c = R / (2γ) and Molodenskii’s truncation coefficients Qn can be presented by

π
Qn = ∫ψ S (ψ ) P (cosψ )sinψ dψ
n . (2.6)
0

Note that Eq. (2.5) applies the unmodified Stokes function S(ψ), i.e. Eq. (2.2), and due
to the existence of various errors, the terrestrial gravity anomalies ∆ĝ and the harmonics
∆ĝn are only estimates of their true values. It is also assumed that the gravity anomaly
can be expanded into a series of Laplace harmonics, i.e. ∆g = ∑ n =2 ∆g n . The
∞

harmonics ∆ĝn can be calculated from a GGM as (cf. Heiskanen and Moritz 1967, p. 89)
n +2
GM  a  n
∆gˆ n =   ( n − 1) ∑ CnmYnm , (2.7)
a2  r  m =− n

where a is the equatorial radius of a GGM, r is the geocentric radius of the computation
point, GM is the adopted gravitational constant, the coefficients Cnm are the fully
normalised harmonic coefficients of the disturbing potential and Ynm are the fully
normalized spherical harmonics (cf. Heiskanen and Moritz 1967, p. 31). Application of
erroneous data and adopted approximations (both theoretical and computational), yield
that instead of the geoid we arrive at its estimator, i.e. a geoid model. However, for the
sake of simplicity the term “geoid” is occasionally used in the sequel.
The generalised Stokes scheme by Vaníček and Sjöberg (1991) utilises the
modified Stokes function and a residual gravity anomaly in the integral. The geoid
estimator Ñ1 is then provided by

10
 R  M
 M
2
N 1 =
4πγ ∫∫σ S L
(ψ ) 

∆ ˆ
g − ∑
n = 2
∆ ˆ
g n

d σ + c ∑
n = 2 n − 1
∆gˆ n , (2.8)
0

where the upper limit L is arbitrary and generally not equal to M (see the discussion in
Section 2.3.2). The modified Stokes function is expressed as

L
2k + 1
S (ψ ) = S (ψ ) − ∑
L
sk Pk (cosψ ) . (2.9)
k =2 2

The modification parameters sn are selected by different criteria, which will be

explained later. Note that throughout this study all sets sn start from n = 2, i.e. s0 = s1 =0.
The main reason is that one of the modification methods under study can never have
parameter s1 (due to inadmissible division by zero), see Eq. (2.10). For some other
modification methods, the contribution of s0 and s1 is insignificant (Sjöberg 1991).
As already noted, the truncation error in Eqs. (2.5) and (2.8) occurs for neglecting
the high-frequency (n > M) contribution of gravity anomalies outside the integration
domain (ψ0 < ψ ≤ π). The primary objective of the kernel SL(ψ) modification is to
reduce the truncation error to a level, which is acceptable for modern geodetic
applications. The modification procedure of the low-degree Legendre polynomials
(2 ≤ n ≤ L) implies that in general ||SL(ψ)|| < ||S(ψ)||. In other words, the modified Stokes
kernel tapers off more rapidly than S(ψ), thus the contribution of distant gravity
anomalies become manageably small. [Note that in the geodetic literature the modified
Stokes function, SL(ψ), has been called also ”spheroidal” or “reduced” Stokes’
function.]
The estimator by Eq. (2.8) employs the high degree residual gravity anomalies,
which are obtained by the subtraction the long–wavelength contribution of gravity from
∆ĝ. Since the low degree gravity field is removed from the Stokes integration, these
effects are compensated (i.e. “restored”) by the second part of Eq. (2.8). The latter is
nothing but the ‘pure’ long wavelength contribution of the geoidal height, cf. also Eq.
(2.21). This method is commonly called a remove-compute-restore (r-c-r) technique,
which is often supplemented with the treatment of the higher frequency topographic
effects. The r-c-r method has its roots in Molodenskii et al. (1962), Moritz (1966 and
1980), Vincent and Marsh (1974). The r-c-r technique is frequently used in practical
geoid computations nowadays (see e.g. Forsberg 1993). For details of this technique the
interested reader is referred to the original sources. Two recent and rather explicit
reviews about the problems of the r-c-r schemes can be found in Sjöberg and Ågren
(2002) and Sjöberg (submitted).
In the literature the contribution of the GGM in the modified Stokes formula is
conveniently referred to as “reference field” or “reference model”. In the present study
the usage of these terms are avoided, since this may misleadingly denote that the GGM-
derived contribution is errorless. We think that any GGM should be treated as an
ordinary data-source, which unavoidably contains errors.
Rigorously, geoid determination by Stokes’ formula holds only on a spherical
boundary, assuming also the masses outside the geoid to be absent. Consequently, the
original and modified Stokes’ formulas should also comprise some correction terms
accounting for the Earth’s ellipticity and the contribution of topographic and

11
atmospheric masses. This study accounts for all these effects by means of additive
corrections (see e.g. Sjöberg 2003c). More details and the expressions of the additive
corrections will be given in Chapter 4. After all, the range of the additive corrections
will be almost the same for all the modification methods under study. Several
comparisons of Chapter 2 are produced by means of subtracting the geoid models from
each other, thus the additive corrections would be reduced anyway. These corrections
are neglected here, since they are not relevant to the main objectives (i.e. selection of an
appropriate modification method for a given target area) of Chapter 2.
Prior to introducing another modification scheme of Stokes’ formula, the details
of two specific modifications of Eq. (2.8) are presented in Sections 2.1.1-2. These
approaches are: the modification methods of Wong and Gore (1969) and Vaníček and
Kleusberg (1987). Since both approaches will be referred frequently in the sequel, for
the sake of simplicity we denote them as WG and VK, respectively.

2.1.1 Wong and Gore (1969) modification method

As already noted, the modification methods differ from each other by the choice of the
modification parameters sn in Eq. (2.9). In the WG approach the parameters sn are a
priori fixed to

2
sn = , ∀n ≥ 2 , (2.10)
n −1

which is equivalent to the case when the summation in Eq. (2.2) starts from L+1. As
will be explained later (see Section 2.3.2), the choice of the upper limits L and M in Eq.
(2.8) may be related to the integration cap radius ψ0. In this study a special case of the
WG modification is considered. De Witte (1966) found small truncation errors at the
zero-crossing of the original Stokes function S(ψ), Eq. (2.2). This yields that the
integration should be extended to the zeros of S(ψ) {or SL(ψ)}. However, a zero can be
achieved anywhere by simply subtracting a constant from the Stokes function (either the
original or modified). This principle was originally suggested by Meissl (1971). The
modified Meissl kernel is defined as Sme(ψ) = S(ψ) - S(ψ0), where ψ0 is the selected
truncation radius. If S(ψ) happens to have zero at ψ0, then no subtraction is necessary,
of course. Heck and Grüninger (1987) propose an alternative to the simple subtraction,
placing a constraint on the values that can be chosen, i.e. either for the parameters L or
ψ0. Accordingly, the modification degree L in the present study will be selected such
that the WG modified kernel, SL(ψ) by Eq. (2.9), is zero at certain radius ψ0.

2.1.2 Vaníček and Kleusberg (1987) modification method

VK modification strategy aims at minimising of the upper bound of the truncation error
in a least squares sense. For this the Stokes function is modified as

P
2k + 1
S P (ψ ) = S L (ψ ) − ∑ tk Pk (cosψ ) , (2.11)
k =2 2

12
where the coefficients tk are determined for pre-selected values of ψ0 and P and can be
evaluated from a system of linear equations

P
2k + 1 P
2k + 1
∑
k =2 2
tk enk = Qn − ∑
k =2 n − 1
enk , (2.12)

where the coefficients

π
enk = ∫ P (cosψ ) P (cosψ ) sinψ dψ
ψ0
n k , k≤n, (2.13)

are, similar to Qn, the functions of the integration cap radius ψ0. The coefficients Qn and
enk are usually computed using some recursive algorithms; see e.g. Paul (1973) and
Hagiwara (1976). The parameters tk minimise the truncation error according to
principles proposed by Molodenskii et al. (1962).
The VK coefficients tk are then applied in Eq. (2.11) to compute the modified
kernel. Generally, the upper summation limits P and L are not equal, i.e. P ≠ L, but the
special choice P = L is often applied in practise. This yields that the modified Stokes
function, Eq. (2.11), becomes

L
2k + 1  2 
S L (ψ ) = S (ψ ) − ∑  + tk  Pk (cosψ ) , (2.14)
k =2 2  k −1 

which is equivalent to Eq. (2.9), with the modification coefficients sn defined as

2
sn' = + tn , (2.15)
n −1

where a prime is used to distinct the parameters sn from those of Eq. (2.10). In other
words, the VK modification method with P = L corresponds implicitly to the
generalised scheme by Eq. (2.8). Featherstone et al. (1998) proposed that the principles
of the Meissl modification can also be applied to the VK method. As for the Heck and
Grüninger (1987) modification, the values of P (and L) can be chosen such that the
kernel by Eq. (2.11) passes through zero at pre-selected truncation radius ψ0. However,
this approach requires some iteration, because the VK coefficients tk are themselves
functions of ψ0.

2.1.3 Integration with the original gravity anomaly

The concept of the geoid determination by the r-c-r technique implies that low-
frequency gravity signals are removed from the Stokes integration. The WG and VK
estimators employ the residual gravity anomaly ( ∆gˆ − ∑ n =2 ∆gˆ n ) and a GGM.
M

Importantly, as shown in Sjöberg and Hunegnaw (2000, Eq. 3; also in Appendix A), the

13
general estimator by Eq. (2.8) can be expressed such that the original surface gravity
anomaly instead of the residual anomaly is exploited in the integral. According to (ibid.)
the estimator by Eq. (2.8) is theoretically equivalent to

M
R
N 1 = L
ˆ σ + c ∑ (QnL + sn ) ∆gˆ n
∫∫ S (ψ )∆gd , (2.16)
4πγ σ0 n =2

where the truncation coefficients QnL are calculated as

L
2k + 1
QnL = Qn − ∑ sk enk . (2.17)
k =2 2

Comparing Eqs. (2.8) and (2.16) we see no particular advantage of reducing ∆ĝ in

Eq. (2.8) to Eq. (2.16) that uses the original gravity anomaly (see also the discussion in
Section 2.4.2). Two recent studies, Sjöberg and Ågren (2002), and Sjöberg (submitted),
demonstrate that the r-c-r result is as sensitive to various biases as is the case when
Stokes’ formula with full anomaly is used. Also, it should be noted (and shown in
Chapter 4) that the expressions with full gravity anomaly, i.e. Eq. (2.16), are well-suited
for the additive corrections.
There is another practical value of Eq. (2.16). Gravity data ∆ĝ is often deposited
in regular blocks, which are related to the geographical coordinates. In the geoid
determination process several GGMs and different modification limits M are usually
tested. This yields that the integration result of the r-c-r scheme, Eq. (2.8), will vary
accordingly. Conversely, the integration result in Eq. (2.16) remains unchanged,
provided that the parameters sn and the limit L remain the same. All variations in a
geoidal height would be due to manipulations with the second part of Eq. (2.16). Even
though the capacities of computers are more or less satisfactory nowadays, the latter
scheme allows us to speed up computations and is convenient for comparing the results
of different experiments.
Recall that the modification coefficient set sn in Eqs. (2.16) and (2.17)
corresponds to the respective modification methods. As a matter of fact, even the simple
modification method by Eq. (2.5) is nothing but the special case of Eq. (2.16) with all
coefficients sn = 0. From Eq. (2.16) another, more general estimator, which includes
most methods applied today for modifying Stokes’ formula, can be expressed. The
geoidal height is provided by two sets of parameters (sn and bn) as follows (cf. Sjöberg
2003d, Eq. 7):

M
R
N 2 = ˆ σ + c ∑ bn ∆gˆ n .
∫∫ S (ψ )∆gd
L
(2.18)
4πγ σ0 n =2

Note that for the sake of the comparison all methods of this study (except the simple
modification) utilise the modified Stokes’ function and the original gravity anomaly in
the integral. The transformation of the WG and VK geoid estimators into an equivalent
expression by Eq. (2.18) is made possible by the special choice of the parameters
bn = sn+ QnL. Since the new expression is mathematically equivalent to the initial form,
Eq. (2.8), we continue to refer to the original sources in the sequel.

14
2.1.4 Deterministic and stochastic modification methods

In Eq. (2.8) the low degree harmonics are simply removed from the kernel of the
Stokes’ function, assuming thus that the global geopotential model is errorless to the
spherical harmonic degree M. However, the coefficients of the geopotential model have
been obtained via some estimation process from satellite tracking data, containing noise,
which unavoidably propagate into the computed geoid undulations. One should also
consider the erroneous terrestrial gravity data within the spherical integration cap. The
gravity data is usually presented as surface blocks. An additional error occurs thus due
to loss of short wavelength gravity information (so called discretization error) when
estimating the mean anomalies ∆ĝ from point gravity data. Consequently, the estimator
of Eq. (2.18) can be rewritten in the spectral form (cf. Sjöberg 2003d, Eq. 8)

∞ M
 2 
N 2 = c ∑  − QnL − sn*  ( ∆gˆ n + ε nT ) + c ∑ bn ( ∆gˆ n + ε nS ) , (2.19)
n =2  n − 1  n=2

where we have included the spectral errors ε nT and ε nS of the terrestrial and GGM-
derived gravity anomalies, respectively. The modification parameters are

 s , if 2 ≤ n ≤ L
sn* =  n . (2.20)
0 , otherwise

The main objective of the modification procedure is to minimise the geoid

estimator error. The modification methods proposed in geodetic literature can be
divided into two distinct classes: deterministic and stochastic approaches. The
deterministic approaches principally aim at reducing the effect of the neglected remote
zone (σ - σ0) making use of a set of low-degree geopotential coefficients. No attempt is
made to reduce the errors of the potential coefficients and terrestrial data, although
errors of both datasets are contributing to the total error budget. Nevertheless, it was
often assumed, that terrestrial data from large territories may somewhat reduce the
geoid errors. The most prominent deterministic approaches are Molodenskii et al.
(1962), Wong and Gore (1969), Meissl (1971), Vincent and Marsh (1974), Heck and
Grüninger (1987), Vaníček and Kleusberg (1987), Vaníček and Sjöberg (1991),
Featherstone et al. (1998). The reviews and comparisons of various aspects of the
deterministic modifications methods can be found, e.g. in Jekeli (1981), Vaníček and
Featherstone (1998), Featherstone et al. (1998), Evans and Featherstone (2000),
Featherstone (2003), just to name a few. For numerical studies and practical
applications, see Kearsley (1988), Vaníček et al. (1996), Forsberg et al. (1997),
Featherstone et al. (2001), Omang and Forsberg (2002).
In contrast, the methods of the least error variance solution and least squares
spectral combination (Sjöberg 1980 and 1981; Wenzel 1981 and 1983) aim at reducing
errors stemming from the GGM and the terrestrial gravity anomaly through a stochastic
kernel modification. Finally, the modification methods proposed by Sjöberg (1984a,b,
1991, 2003d) allow minimization of the truncation error, the influence of erroneous
gravity data and geopotential coefficients in the least squares (hereafter LS) sense.
Reviews, comparisons and some recent applications of these methods can be found in

15
Sjöberg (1986 and 2003d), Sjöberg and Hunegnaw (2000), Nahavandchi and Sjöberg
(2001), Ellmann (2001) and PAPERS A, B and C.
Generally, for minimising the errors in Eq. (2.19) the stochastic methods aim at an
optimal combination of the data sources (and their error estimates) by adopting a priori
or empirical stochastic models. Frequently it is argued that these spectral models of
errors are too poorly known to justify the application of stochastic geoid estimators. For
instance, the error variances of the global geopotential model are also global and so do
not necessarily represent the area under investigation. Indeed, the preference for the
stochastic or deterministic methods is basically a philosophical question: Either one
uses possibly doubtful stochastic information, or this additional information is
completely neglected (Heck and Grüninger 1987). Nevertheless, we believe (supported
by the results of this study) that it is better to use a coarse error model, than to assume
that data are without errors.

2.2 Expected global mean square error and modification parameters

The approaches by Sjöberg (1984b, 1991 and 2003d) aim at minimizing the errors of
the geoid estimator, Eq. (2.19), in the least squares sense. A brief review and relevant
expressions are presented in this section.
Based on the spectral form of the “true” geoidal undulation N (Heiskanen and
Moritz 1967, p. 97),

∞
2
N = c∑ ∆g n , (2.21)
n =2 n − 1

the expected global mean square error (MSE) of the geoid estimator Ñ2 can be written:

 1 ∞  
 − N ) 2 dσ  = c 2 ∑ b2 dc + c 2 ∑ ( b* − Q L − s* )2 c +  2 − Q L − s*  σ 2 
M 2

mN2 = E 
 4π
∫∫σ 2
( N n n

n n n n 
 n −1
n n

n
 n =2 n =2  
(2.22)

where E{} is the statistical expectation operator and

b , if 2 ≤ n ≤ M *  sn , if 2 ≤ n ≤ L
bn* =  n ; sn =  . (2.23)
 0 otherwise  0 otherwise

Eq. (2.22) is an important formula. Its validity for the modification methods utilising
either Eq. (2.8) or (2.18), follows from the equality in Sjöberg and Hunegnaw (2000).
Since all the data errors are assumed to be random and with expectations zero, the norm
of the total error is thus obtained by adding their partial contributions. The first term of
the right side of Eq. (2.22) represents the contribution due to errors of the geopotential
model. The middle term reflects the truncation error and the last term accounts for the
influence of erroneous terrestrial data.
Principally, possible correlation between the data-sets can also be considered in
Eq. (2.22). Even though this subject will be tackled in more details in Section 2.3.2,

16
some relevant aspects are provided here as well. Note that a high-degree GGM is
determined from a combination of satellite data and terrestrial gravity data. This
combination implies that the two datasets in Eqs. (2.19) and (2.22) could be correlated.
Rigorously, this feature could be accounted for by adding some terms into both
formulae. If one utilises the “satellite-only” harmonics, this correlation is avoided, of
course. The modification limits of this study are selected (more details in Section 2.3.2)
such that this feature is excluded. Correlation may appear, however, in a few side
experiments when a full expansion of EGM96 is utilised. Since the terrestrial data
information is comprised in the higher degrees of a GGM, but most of the geoid power
is in lower degrees, the correlation-related influence is most likely insignificant.
Therefore this correlation is completely neglected in this study. Naturally, one cannot
ignore the correlation in the cases when a high-degree expansion is combined with very
small integration cap. This case, however, will not occur here. See Sjöberg (1984b and
1991) for the full theory and the expressions accounting also for systematic errors and
correlated data-sets.
It is notable, that the global mean degree variances of gravity signal and noise are
applied in Eq. (2.22). The gravity anomaly degree variances are denoted by cn and can
be computed as

1
cn = ∫∫σ ∆g dσ
2
, (2.24)
4π
n

whereas the terrestrially measured and GGM-derived anomaly error degree variances
are denoted by σ n2 and dcn. These are computed by

 1 
∫∫σ (ε )
2
σ n2 = E  T
dσ  , (2.25)
 4π
n


and

 1 
∫∫σ (ε )
S 2
dcn = E  dσ  , (2.26)
 4π
n


respectively. Note that the global representation of the quantities by Eqs. (2.24) - (2.26)
is employed. Since the true values of the error components are unknown, their
estimation could be based on the standard approaches and stochastic models. The
stochastic models used in this study are presented and discussed in Appendix B.
Apparently, the computed geoid height is affected by the properties of the
modification procedure, i.e. the choice of the integration cap radius, upper modification
limits (L and M) and the coefficients sn and bn. The key factor to minimize mN2 is,
however, a suitable selection the parameters sn in Eq. (2.22). For obtaining the LS
modification parameters Eq. (2.22) is differentiated with respect to sn, i.e. ∂mN2 / ∂sn .
The resulting expression is then equated to zero and the modification parameters sn are

17
Table 2.1. Main parameters of the deterministic and stochastic modifications of the Stokes formula. The
notations refer to Eqs. (2.18) and (2.27).

Para- Stochastic (LS) modifications Deterministic modifications

meters Biased Optimum Unbiased Wong-Gore Vaníček-Kleusberg Simple
L
2 2
sn = ∑a
n=2
kn ⋅ sn = hk , k = 2, 3... L .
n −1 n −1
+ tn 0

bn = sn
(QL
n + sn ) cn
QnL + sn QnL + sn Qn
cn + dcn

solved in the least squares sense from the linear system of equations:

∑a
r =2
kr ⋅ sr = hk , k = 2,3,….L , (2.27)

where akr and hk are modification coefficients, which can be expressed via Qn , enk , cn ,
dcn and σ n2 . More details of the computation of the LS parameters sn will be given in
Chapter 3 (see also Table 3.1). After solving sn numerically the corresponding
coefficients bn are computed. Three LS methods are studied in this dissertation. These
methods are the biased, unbiased and optimum LS modifications (Sjöberg 1984b, 1991
and 2003d, respectively), which basically deviate from each other in the choice of
parameters bn; see Table 2.1. The parameters of the deterministic methods under study
(simple, WG and VK) are also presented in Table 2.1.
To illustrate the significance of selecting the parameters bn, note the appearance of
the middle term in the right side of Eq. (2.22). Obviously, with bn = sn + QnL the
truncation error is completely reduced to degree M. This yields an unbiased (to degree
M) geoid estimator; hence the name for the unbiased LS method. The other two LS
approaches, biased and optimum, minimize the truncation error as well, but their
estimator is slightly biased (Sjöberg 2003d). Note, however, that if the GGM-related
errors dcn below selected limit M are small, then for the optimum LS method this bias is
very little. If the GGM-related error becomes significant (e.g. by increasing degree M)
the discrepancies between two methods increase, of course. Summarizing, the
modification methods by Sjöberg (1984b, 1991, 2003d) attempt via minimization of the
global MSE to reduce any error in geoid modelling. In these approaches, depending on
the local gravity data quality, the chosen radius of integration and the characteristics of
the used GGM, the modification parameters sn vary. In contrast, the coefficients sn for
the deterministic methods are a priori defined; see Eqs. (2.10) and (2.15).

2.3 Data and modification criteria

2.3.1 Geopotential models

For evaluating the global MSE all components of Eq. (2.22) need to be defined.
Obviously, the error degree variance dcn is dependent on the quality of the used GGM.
In the past 40 years, many geopotential models have been estimated, for a description of

18
existing models we refer to Wenzel (2000) and the references therein. For details of
development and analysis of geopotential models see Rapp and Pavlis (1990).
In the compilation of a GGM, the long-wavelength contribution of the Earth's
gravity field is recovered from the tracking data of artificial satellites. Improvements to
the Earth gravity models at medium and short wavelengths should come from the use of
satellite altimetry, terrestrial, marine or airborne gravity surveys - of varying epoch,
quality and geographic coverage. The accuracy of such models is quite dependent on the
geographic coverage of gravity data that goes into the solution. As the coverage
improves, so will the model.
Prior to the new satellite gravity missions CHAMP and GRACE the harmonic
degree errors of the GGMs were roughly divided into three frequency bands (cf.
Vaníček and Featherstone 1998):

1) Spherical harmonic degrees 2 ≤ n ≤ 20 offer a superior information source of the low

frequency component of the geoid. The estimation of the coefficients of these
degrees is exclusively based on the satellite-derived contribution.
2) Spherical harmonic degrees 20 < n ≤ 120, for which the GGM gives a reasonable
accuracy almost everywhere on Earth and where the terrestrial data may offer an
improvement in certain parts of the world, only if these data are of good quality.
3) Spherical harmonic degrees 120 < n ≤ 360, for which the geopotential model may not
be the best source of gravity field information and an improvement from terrestrial
data should thus be sought. The degradation of the GGM in this region can be seen
from the error degree variances, which are usually almost of the same magnitude as
the gravity signal.

In the present study the GRACE Gravity Model GGM01s is employed for
regional geoid studies. Additionally, the EGM96 (Lemoine et al. 1998) is utilised in
some computations. A few aspects of this choice need to be discussed.
The EGM96 harmonic coefficients are complete to degree and order 360. The
spatial resolution of the “satellite-only” solution is limited to about 600-700 km,
implying the highest harmonic degree as of 35.
Tracking data of the GRACE twin-satellites is the basis of the new geopotential
model GGM01s (www.csr.utexas.edu/grace/gravity), released in July 2003 by the
Centre for Space Research at the University of Texas, a group led by GRACE principal
investigator Dr. B. Tapley. The GGM01s field is developed to degree and order 120,
whereas the solution provides full-power results up to about harmonic degree 95. Due to
the global, homogeneous nature of GRACE data, the resulting geoid errors show no
discrimination between land and sea areas, as do previous gravity models. [This
criterion allows us to apply standard stochastic models without “tailoring” the
geopotential model errors after the area of interest, for more details see Appendix B].
The GRACE group estimated the GGM01s geoid accuracy to be accurate to
approximately 2 cm to degree/order 70 (300 km resolution) and 6 cm to degree/order 90
(200 km resolution). Furthermore, the GGM01s was combined with the Texas Earth
Gravity Model TEG-4 (recent update in Tapley et al. 2003) information equations
(created from historical multi-satellite tracking data and surface gravity data) to produce
a preliminary gravity model GGM01c, complete to degree and order 200. This model
will be used in one of our experiments, see Section 2.5.6. For more thorough review see

19
PAPER B, which investigates also the performance of GGM01s, GGM01c and EGM96
in the Baltic Sea region.

2.3.2 Modification limits

The selection of the upper limit of the geopotential model, M and the upper limit of the
harmonics to be modified in Stokes’ function, L {see Eq. (2.9)}, are of crucial
importance in the geoid modelling. The choice of the limit M is directly related to the
quality of the GGM to be used. In practice, due to restricted access to terrestrial data
(or/and for necessity to increase the computational efficiency) the integration cap is
often limited to a few hundred kilometres. This implies that a relatively high M should
countermeasure this limitation. On the other hand, the error in the GGM grows with
increasing degree, which provides a rationale for a compromise value of M.
Traditionally, a rather small M was favoured in the computations of the past geoid
models. For instance, the Canadian geoid models (Vaníček and Kleusberg 1987,
Vaníček et al. 1996) utilize M = L = P = 20. Consequently, in these computations even
the intermediate wavelength information from a GGM is prevented. In absolute
contrast, the computations of some geoid models (e.g. Forsberg et al. 1997) utilize a
high-degree field (i.e. complete expansion of GGM) in conjunction with unmodified
Stokes’ function S(ψ). Alternatively, a low degree modified integral kernel can be
combined with a high-degree expansion of GGM (e.g. Featherstone et al. 2001; Omang
and Forsberg 2002). Consequently, when choosing the limits L and M one needs to
consider many (often rather controversial) arguments.
As already pointed out in Section 2.2, by including the higher degree contribution
of a GGM the same data may be used twice. This necessitates to account for the
correlation relations, e.g. by adding some cumbersome terms (Sjöberg 1991) into Eqs.
(2.19) and (2.22). For the final geoid model we aim at selecting the modification limit
M such that the correlation between the GGM-derived and terrestrial datasets is
prevented.
As will be shown in Section 2.3.5, the limited extension of terrestrial data is the
most serious restriction for the present computations. In this study, similarly to some
earlier experiments in the same region, the integration cap radius 2° has been favoured
(as it produced good results in the past, see Ellmann 2001, PAPER C). We want the
modified Stokes function SL(ψ) to become zero at the edge of the truncation cap.
Therefore the choice ψ0 = 2° is the basis for determining the upper modification limit L
for the WG and VK methods.
We start with the VK modification method. It is customary (also advocated by the
authors of the original publication) to select L = P. Using ψ0 = 2° the VK modification
coefficients tk and thence the modified kernel SL(ψ) values are computed by Eqs. (2.12)
and (2.14). Some iterations are required to detect the modification degree L = P, which
enforces the kernel to zero at ψ0 = 2°. This condition is satisfied with L = P = 67. Limit
M should preferably not exceed the full-power harmonics (i.e. M ≤ 95) of the GGM01s.
Moreover, since the area of interest is gravimetrically well studied (see Section 2.3.5),
the terrestrial data is thus probably a better source of medium and short wavelength
geoid information. This suggests a smaller value for M. We select L = M = 67 for our
computations. In other words, the degree of the modified kernel SL(ψ) is the same as the
upper limit M of the geopotential harmonics to be used. This choice is also supported by

20
a circumstance that the error degree variances dcn of the GGM01s harmonics with
n ≤ 67 are smaller than the degree variances with n ≤ 20 (often selected for regional
geoid computations in the past) of any previous geopotential model. The kernel SL(ψ)
with the WG coefficients, 2/(n-1), becomes zero at ψ0 = 2° when L = 31 in Eq. (2.9).
For the sake of the comparison we choose M = 67 for both approaches.
Now it remains to define the parameters for the LS modifications. As it will be
shown in Chapter 3, the LS parameters “force” SL(ψ) to almost zero at the edge of any
pre-selected integration cap. Reasoning the comparison we thus select also for the LS
modifications the same upper limit as is favoured in the deterministic methods, i.e.
L = M = 67. This choice allows us to take full advantage of the new GGM01s, whereas
there is strictly no correlation with terrestrial data. On the other hand, if the terrestrial
gravity data in the area of interest is poor, there is no point to abandon erroneous high-
degree harmonics of a GGM. Remember, that the gain from high-degree harmonics may
be more rewarding than possible damage. A relevant matter is how to damp the errors
and find a correct balance (e.g. weights) between different data sources (i.e. GGM and
terrestrial gravity anomaly). The LS modifications are designed for aiming at this goal.
We refer to some numerical studies in Ellmann (2001, Figs. 5.6-7), which utilise
L = M = 360 of EGM96. It is demonstrated that the higher degree (say n > 70) spectral
contributions of EGM96 in the LS geoid modelling are rather insignificant. It can be
concluded thus, that the LS procedure is able to adjust the data in a way that the geoid
model becomes rather insensitive to the maximum degree of modification, because it
matches the different types of data in an optimum way. Conversely, the limit L should
be selected very carefully for the deterministic methods.
A study by Nahavandchi (1998) utilised the GPS-levelled heights of the Swedish
GPS Permanent GPS network stations for evaluation of the geoid modelling outcomes.
The following modification methods were considered: a LS, the WG, Vincent and
Marsh (1974) and a Molodenskii-type modification {i.e. Vaníček and Kleusberg
(1987)}. Nahavandchi (1998) concluded, that the LS modification method provides
superior results (note that the RMS values of post-fitting residuals remain unclear;
Nahavandchi, pers. comm. 2001), whereas the other methods provided much poorer
quality. In this context different aspects (e.g. extension of integration cap, topographic
corrections etc.) of those modification methods were studied. Although various values
for M (e.g. M = 20, 40, 60, 100, 180) were investigated, the full expansion (i.e.
M = 360) of EGM96 was applied in the final computations. Also it seems that for the
deterministic methods the degree of modification of Stokes’ function is taken equal to
the upper limit of EGM96, i.e. L = M = 360 (except the Vincent-Marsh method with
L = 0) as the final choice. Recall, that for the LS methods a high degree modification of
function SL(ψ) is not critical. Conversely, it is clear from the beginning that the
modified Stokes function for the deterministic methods will strongly oscillate due to
large L. As the limit L increases the oscillation increases (for an illustration see PAPER
G, Fig. 1) and thereby the geoid modelling results will be affected by said phenomena.
As pointed out by some authors (e.g. Featherstone 2003) the increased oscillation of the
kernel causes error in the numerical solution of the discretized integral argument (recall
that gravity data values are usually related to the centres of the surface blocks). If the
kernel varies hastily across the integration cap, the central value of the data blocks will
not be representative any more. Therefore a smooth integration kernel is required. This
may also explain the more scattered statistics attributed to the deterministic methods in
Nahavandchi (1998). In our study it is therefore preferred to use much lower L, which

21
forces the modified Stokes function of the deterministic methods to zero at the edge of
the integration cap. This allows a “fair” comparison of the deterministic and stochastic
methods. In other words, the demands of the deterministic methods are put on the first
place, although for the LS methods a higher L would be probably more optimal; for a
discussion see Section 2.4.
Shortly summarising, the limits L = M = 67 are used everywhere in the
computations for the VK and LS modification methods, whereas M = 67 and L = 31 is
adopted for the WG method. The limits L and M are essential for assessing the degree
variances cn and dcn by the approaches presented in Appendix B. Next we deal with the
estimation of the degree error variances of the terrestrial data σ n2 . This necessitates the
description of the target area and available datasets.

2.3.3 Target area

The present study aims at computing the geoid model for the three Baltic States –
Estonia, Latvia and Lithuania. These countries (the total area is ~175 000 km2) lay on
the eastern shore of the Baltic Sea, bordering with Russia, Belarus and Poland. The
target area is defined as the area with geographical boundaries from 53.8° to 60°
northern latitudes and from 20° to 28.5° eastern longitudes (approximate size 670 × 450
km2), see Fig. 2.1. The Baltic gravimetric geoid model comprises thus entirely or partly
- Estonia, Latvia, Lithuania, Russia, Belarus, Poland and Finland, together with a large
portion of the Baltic Sea surface.

Fig. 2.1. Location of the target area (enclosed by the bold rectangle). External rectangle
encloses the target area at a spherical distance of 2° (approx. 220 km) from its borders.

22
The elevation extremes are 0 m at shoreline and 318 m in southeast Estonia, whereas
most of the target area comprises sea and topography below 100 m (see also Fig. 4.1 in
Chapter 4). Treatment of topographical effects, together with quality and resolution of
terrestrial gravity data and geopotential models, is one of the most serious limits in the
computations of accurate gravimetric geoid models. Due to insignificant topography
these effects are rather small and almost constant over the whole target area. The Baltic
Sea region seems thus to be an attractive test area for different geoid modelling
approaches (see also Bilker et al. 2001).

2.3.4 Previous geoid models in the target area

It is appropriate to review some earlier geoid models for the same region. During the
last two decades the geoid determination for the whole Nordic region has been carried
out within the framework of the Nordic Geodetic Commission (NKG). Several NKG
geoid models were delivered, see e.g. reference list in Forsberg (2001). In 1990-ies the
NKG geoid models were extended to the Baltic countries. Access to new gravity data
from formerly classified sources and release of EGM96 resulted at achieving of better
than a dm-accuracy for the NKG96 model (Forsberg et al. 1997).
The national geoid solutions, either for Estonia, Latvia or Lithuania, were
published by Vermeer (1994), Kaminskis and Forsberg (1997), Forsberg (1998),
Parseliunas and Forsberg (1999), Jürgenson (2001 and 2003), Ellmann (2001).
Additionally, in connection with the Baltic Sea Level GPS-campaigns the BSL95A
geoid model was computed by Vermeer (1995). This geoid model includes most of the
target area of the present study. With a few exceptions (e.g. Vermeer 1994 and 1995,
Ellmann 2001) the computations have been carried out by the NKG software and
EGM96 is utilized in the computations. The NKG models are presented as quasi-geoidal
heights (i.e. models of height anomalies). It should be noted that the most recent NKG
model, Forsberg et al. (in print), was not yet published when conducting the present
study.
The NKG computational methodology has many similarities (though not exactly
the same, see the discussion in Sjöberg and Ågren 2002) with Vincent and Marsh
(1974) modification scheme

R  M
 M
2
N =
4πγ ∫∫σ S (ψ ) 

∆ ˆ
g − ∑
n =2
∆ ˆ
g n

d σ + c ∑
n =2 n − 1
∆gˆ n . (2.28)
0

Note that the unmodified Stokes function is utilized in conjunction with the residual
gravity anomaly in the integral. This modification method follows implicitly the
generalised Vaníček and Sjöberg (1991) scheme, Eq. (2.8), with parameters sn = 0 in the
integral. In the practical computations the complete expansion of a GGM is usually
employed (i.e. M = 360 for EGM96). However, as shown in Sjöberg (1986, also in
Sjöberg and Ågren 2002) that Eq. (2.28) is theoretically identical to the simple
modification method by Eq. (2.5), i.e.

23
 R  M
 M
2 R M
N =
4πγ ∫∫σ S (ψ ) 

∆gˆ − ∑
n =2
∆ ˆ
g n 

d σ + c ∑
n =2 n − 1
∆gˆ n =
4πγ
ˆ σ + c ∑ Qn ∆gˆ n . (2.29)
∫∫ S (ψ )∆gd n =2
0 σ0

Accordingly, the simple modification scheme by Eq. (2.5) is adopted in this study as a
very rough representation of the NKG approach. This allows us to compare the five
earlier reviewed modification methods with the general features of the NKG method.
Diversely from others, the integration kernel S(ψ) of the simple modification is not zero
at the edge of a small integration cap. It should be admitted, however, that the NKG
computations used a much larger integration area than the present study. Final
computations for the simple modification method utilise the full expansion of EGM96
(i.e. M = 360, recall that for other methods GGM01s with M = 67 is exploited). The
reasons behind such a choice will be explained in Section 2.5.6.

2.3.5 Terrestrial data

To accurately determine the geoid, the gravity data within a spherical distance of 2°
from each computation point are involved. The choice ψ0 = 2° implies that in addition
to the above listed countries also gravity datasets from certain parts of Sweden are
included; see Fig. 2.2. The gravity data used were obtained from the Danish National
Survey and Cadastre (KMS), the authorized holder of the Nordic-Baltic gravity
database. The total number of used gravity points is ~103 000.

62 62 25
0
0 0
-25
-25
- 25

0 -25
0
0
60 60 -50 0
-2 5 -25 0
0 -25
-2

-5 0
5

-25

-5 0
0
0

5
-2
58 58 5
0

-2
LATITUDE [°]
LATITUDE [°]

-2

-2 5 -50
5

-5 0 0
-25 2 5
-2

0
5

0
0

56 56 -2 5
0
0

0
25

0
0
0

0
54 54 -2 5
-2
0

5
-5
0

-25
0
-2 5

0 0
-5

52 52
0 25
0

0
0

16 20 25 30 16 20 25 30
LONGITU DE [°] LONGITU DE [°]

Fig. 2.2. Distribution of the data points. Target Fig. 2.3. Free-air gravity anomalies in the data
area is enclosed by the rectangle. area. Unit is mGal, contour interval is 25 mGal.

24
The treatment of the data, collected during several decades with different methods and
equipment and by different nations and specifications, requires careful study before
geoid computation. An extensive analysis of the used Nordic-Baltic gravity data-sets
can be found in Ellmann (2001, Ch. 4). Attention was focused on the following
subjects: transformation of the national gravity values into the IGSN71 gravity system,
divergence of time epochs, treatment of tidal effects, height and coordinate systems,
influence of the Fennoscandian postglacial land uplift etc. For instance, a vertical datum
difference of 15 cm (see PAPER B) causes a systematic gravity difference of ~0.05
mGal which propagates into the geoid as a long wavelength effect. The systematic
errors in Nordic-Baltic gravity datasets have also been noticed by other authors, see,
e.g., Omang and Forsberg (2002). In further computations, however, the data is used
without further corrections, hence, any possible systematic bias between the national
datasets is simply ignored both in earlier solutions and in this study. It should be
emphasised, however, that the gravity surveys of the three Baltic States are historically
related to the same datum. Subsequently, the impact of inherent systematic biases
should be theoretically reduced to insignificancy.
Since this study focuses on the three Baltic countries, we provide some relevant
information about the datasets of these countries. The Estonian gravity survey was
performed by the Institute of Geology at the Estonian Academy of Sciences in 1949-58
(Sildvee 1998). The total number of Estonian data points exceeds 4000, producing a
density of approximately one survey point per 10 km2. The mean standard error of the
gravity measurements is stated as 0.3 mGal. Heights of the survey points are related to
the Kronstadt tide-gauge, the accuracy for height determination is stated as 0.5…1.5 m.
A register of Latvian and Lithuanian gravity points is mainly reconstructed from the
1:200 000 scale paper maps (Parseliunas and Buga 1994; Kaminskis and Forsberg 1997;
Parseliunas and Forsberg 1999). It should be mentioned that the free-air anomalies for
most Latvian and Lithuanian gravity points are calculated via Bouguer anomalies
complemented with the respective heights from an older global terrain model. Hence,
the resulting free-air anomalies could contain remarkable errors.
The geoid models are also strongly dependent on the gravity data coverage. The
coverage within the three countries and in adjoined areas is more or less satisfactory,
except for the eastern and south-eastern parts, where only a small number of gravity
points is available; see Fig. 2.2. For obtaining a consistent anomaly grid the free-air
anomalies for such gaps were computed from EGM96 (PAPER B, Eq. 1). Thereafter, a
regular 1.5´ × 3´ (arc-minutes, approx. 2.7 × 2.7 km2) grid of free-air anomalies was
constructed; see Fig. 2.3. Within the whole data area (the target area and a 2° zone
outside from its borders) the anomalies vary from –80 to + 42 mGal with a mean of -8.8
mGal. The predicted anomalies are relatively smooth (standard deviation of a mean is
17.8 mGal), but it should also be noted that the eastern and south-eastern parts (and a
narrow strip at the southernmost edge of the data area), are almost entirely calculated
from EGM96. See PAPER C (also in Ellmann 2001) for more details of initial data,
gridding procedures and applied formulas.
The accuracy of gridded gravity anomalies can be determined empirically by
comparing the gridded values at known control points. For this purpose 240 points of
the Estonian gravity network were used. The precision of the free-air anomaly grid in
Estonia is estimated to 2.4 mGal. This information can be used for evaluating the
internal accuracy of the geoid models. According to Vermeer (1995) the following
empirical expression is useful for evaluating the short wavelength geoid error:

25
σN [mm] = 0.3 ⋅ d [km] ⋅ σ∆g [mGal]; where σ∆g is the accuracy of prediction of gravity
anomalies and d is an average point spacing. For Estonia we take d = 3 km and
σ∆g = 2.4 mGal. Thus the expected relative accuracy (with respect to the closest grid
points) of the geoid models is σN = 2.2 mm.
Due to lack of control points the accuracy of the gridded anomalies remain
unknown in the other parts of the target area. Therefore, different values of terrestrial
data error variance C(0) are tested in the Section 2.4. Note that degree variances σ n2 are
to be used for estimating the global MSE. In the other hand, regional geoid
computations involve only local data, which may drastically differ from the global
average. For this reason we will define such variance values, which correspond to the
data actually involved in our computations. Since gravity data accuracy within the target
area may be rather heterogeneous, we experiment with four different variance values
{C(0) = 0, 1, 4 and 9 mGal²}. Recall that the selection C(0) = 0 mGal² corresponds to
absolutely errorless data, 1 mGal² characterises the point data accuracy (i.e. without
discretization error) in Estonia, the variance values 4 and 9 mGal² correspond to the
gridded gravity anomaly accuracy in Estonia. The gravity anomaly degree variance, σ n2 ,
can be estimated from a covariance function C(ψ). In this study σ n2 is assumed to be of
reciprocal distance type (cf. Heiskanen and Moritz 1967, p. 256), which can be
estimated by a simple relationship in Sjöberg (1986). This implies that the covariance
function is isotropic and depends on only two parameters: the variance C(0) and the
correlation length. Similarly to some earlier studies at KTH (Nahavandchi 1998,
Ellmann 2001) we adopt the correlation length 0.1°. More details and references to
other σ n2 models are presented in Appendix B.
Due the variations both in coverage and accuracy of the terrestrial data the validity
of any model for σ n2 may be questioned. The outcomes of some other models may be
different, however, these are expected to be more or less proportional to the model used
in this study. After all, the possible inconsistencies between stochastic models are
probably less important for the discussion of terrestrial data errors in the sequel.
Naturally, if the model is inaccurate, then the absolute range of the global MSE is
affected. In this study, however, we focus on the relative differences between the
modification methods.

2.4 Evaluation of the expected global mean square error

The parameters sn are also present in Eq. (2.22). The computations of the deterministic
parameters are trivial (see Table 2.1), whereas the LS parameters are more complicated.
As was explained earlier, the LS modification parameters sn are to be calculated from
the following initial conditions: (1) the GGM01s model, (2) modification degrees
L = M = 67, (3) terrestrial data error variance C(0) = 0, 1, 4 and 9 mGal², (4)
integration cap size ψ0 = 2°. The modification coefficients sn are determined in a least
squares sense by solving the system of Eq. (2.27). More details and practical aspects of
computing the LS modification parameters can be found in Chapter 3, the stochastic
models for cn, dcn and σ n2 , are presented in Appendix B.

26
 The magnitude of the global MSE, by Eq. (2.22), for all computed parameter sets
sn is presented in Table 2.2. Note that variations of the initial conditions, e.g. ψ0 or C(0),
may improve or worsen the global MSE. Recall, however, in the present study the
integration cap radius is strictly constrained to ψ0 = 2°.
It is informative to examine the components of Eq. (2.22). If the parameters of the
unbiased LS, WG and VK methods, all utilising bn = ( QnL + sn ), are applied then Eq.
(2.22) is specified as follows:

M ∞ ∞ 2
 2 
∑ (Q + sn ) dcn + c ∑ (Q ) cn + c ∑ 
2 L 2
m =c
2
N
2 L
n
2
n
2
− QnL − sn*  σ n2 . (2.30)
n =2 n = M +1 n =2  n − 1 

Apparently, the two first terms in the right side can be considered jointly. The first term
mirrors the contribution of the geopotential model errors (harmonics ≤ M) and the
second term is responsible for the truncation error (due to unavailability of the remote
zone gravity anomaly harmonics beyond degree M). Consequently, for minimising the
global MSE it is feasible to select such a value for M, at which the error of the
geopotential coefficients becomes equal to the power of the gravity signal. The third
term in Eq. (2.30) represents the terrestrial data errors. In practical computations the
infinite sum in Eq. (2.30) is truncated at the maximum degree nmax = 2000. Fig. 2.4
demonstrates that the adopted nmax is sufficient to cover all significant contributions.
Indeed, the error contributions have a progressively smaller effect on the geoid by
increasing degree. This is due to the fact that most of the power of the geoid is
contained in the low frequencies. Numerical values of the global MSE components
2
M nmax
 nmax
 2 
c 2  ∑ ( QnL + sn ) dcn + ∑ (QnL ) cn  and c 2 ∑ 
2 2
− QnL − sn*  σ n2 are cumulatively plotted in
 n = 2 n = M +1  n = 2  n − 1 
Figs. 2.4a-d. For simplicity the former counterpart is referred to as the “joint error”,
whereas the latter term is the terrestrial error, respectively.

Table 2.2. The expected global mean square error for six different modification methods of Stokes’
formula. Notations refer to Eq. (2.22). The modification limits are L, M and P. Initial conditions: (1) the
geopotential models GGM01s and EGM96, (2) terrestrial anomaly variance C(0) with the correlation
length 0.1°, (3) integration cap size ψ0 = 2°. Unit is centimetre.

Adopted Stochastic (least squares) Deterministic modifications

variance modifications
C(0) of Vaníček- Wong and
mN2 Unbiased Optimum Biased Simple
terrestrial Kleusberg Gore
gravity
L=M=67 L=M=67 L=M=67 L=M=P=67 L=31, M=67 L=0, M=360
0 mGal² Glob.err 3.281 3.280 4.307 6.018 6.017 12.093
1 mGal² Glob.err 4.886 4.886 5.552 6.913 6.938 13.499
4 mGal² Glob.err 7.915 7.914 8.214 9.084 9.161 17.036
9 mGal² Glob.err. 11.198 11.198 11.309 11.849 11.982 21.684
Geopotential model GRACE Gravity Model GGM01s EGM96

27
2.4.1 Deterministic methods (M = 67)

We begin with the inspection of the joint error for the deterministic methods; see Fig.
2.4a. Obviously, when the GGM01s is employed, the main contribution of the joint
error is due to M > 67 degrees cn (note an upward jump at M = 67), i.e. for the
truncation error. The joint error for the EGM96-based case (other conditions remained
the same) is much greater than the one employing GGM01s. This is due to greater
errors dcn of EGM96. For the VK and WG methods the joint error is practically the
same (see also results with C(0) = 0 mGal² in Table 2.2). The terrestrial errors for the
WG modification with C(0) > 0 mGal² are merely greater than the ones for the VK
method (see also the resulting global MSE in Table 2.2).

2.4.2 Simple modification method in conjunction with the full expansion of EGM96

The largest global MSE (see the last column in Table 2.2) is related to the simple
modification method by Eq. (2.5). This method utilises the full expansion of EGM96
(i.e. M = 360) in conjunction with the original Stokes function (i.e. L = 0) in the integral
kernel. The joint and terrestrial error components are remarkably larger than the
respective values of any other reviewed method, see Fig. 2.4b. [Note that dcn was not
re-scaled, see also the discussion in Appendix B.] Particularly, the terrestrial error
becomes extraordinarily large.
Assume for a moment that the terrestrial data is without errors (i.e. σ n2 = 0).
Considering the modification parameters sn = 0, then the terms in the brackets ( ) of the
joint error components reduce to Molodenskii’s truncation coefficients Qn; see Eq.
(2.30). As is well known, the larger the integration cap the less the coefficients Qn (i.e.
Qn → 0 with ψ0 → π). This provides a rationale to employ as large cap as possible for
this modification method (also targeted for the NKG geoid models). Intuitively, a limit
M < 360 may be acceptable only when compensated with a large data area. According
to the study by Sjöberg and Ågren (2002), however, in order to reduce the truncation
error in the NKG approach to a cm level, the integration cap must exceed 10°. [Note
that the “full” gravity field was used in their study. Obviously, the truncation error
becomes somewhat smaller when a r-c-r scheme or/and the topographically corrected
gravity anomalies in mountainous regions are utilised.]
Considering the deficiencies of the past geopotential models, one could assume
that good quality terrestrial data may somewhat compensate these shortages (e.g. Sideris
and Schwartz 1987). This was another reason to exploit vast data areas, wherever
possible. In this respect it is interesting to review the results of Vaníček and
Featherstone (1998). They concluded that even the error-free gravity data, when used in
a limited cap, can never completely correct the errors present in the geopotential model.
With access to up-to-date geopotential models, the accuracy of the low and intermediate
frequency bands can be considered almost errorless compared to the terrestrial data
errors. Subsequently, this problem is not relevant any more.
Indeed, the terrestrial data is never without errors (i.e. σ n2 > 0 mGal²) and these
errors will inevitably be present in the geoid estimation process. Returning to Eq. (2.30)
we see that the brackets of the last term reduce to [2/(n-1) - Qn]. Recall that Qn→ 0 with
ψ0→π.

28
 0.18
0. 1
0.16
EGM96
0.14
0.08

GLOBAL MSE [m]

GLOBAL MSE [m]

0.12

0.06 0. 1
GG M01s
0.08
0.04
0.06

0.04
0.02
0.02

0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
DEG REE [nm ax ] DEG REE [nm ax ]
Fig. 2.4a. Fig. 2.4b.

0.1 0.1

0.08 0.08
GLOBAL MSE [m]
GLOBAL MSE [m]

0.06 0.06
WG

0.04 LS (9 mGal2 ) 0.04 VK

2
LS (0 mGal )
0.02 0.02
LS

0
0 0 200 400 600 800 1000
0 200 400 600 800 1000
DEG REE [nm ax ]
DEG REE [nm ax ]
Fig. 2.4c Fig. 2.4.d

Fig. 2.4. Cumulative sum of the global MSE components. Unit is metre.

LEGEND
Notations refer to Eq. (2.30),
values nmax > 1000 are not visualised. Integration cap is 2 °, terrestrial data errors:
C(0) = 1 mGal² (dashed line), C(0)= 4 mGal² (dotted line) and C(0)= 9 mGal² (dash-dot line).

Fig. 2.4a. Vaníček and Kleusberg (1987) modification, (L = M = P = 67):

Joint contribution of the GGM01s errors and the truncation error (solid bold line),
joint contribution of the EGM96 errors and the truncation error (thin solid line).

Fig. 2.4b. Simple modification method (EGM96, L = 0, M = 360):

Joint contribution of the EGM96 errors and the truncation errors (solid bold line)
Note different vertical scale!

Fig. 2.4c. Unbiased (Sjöberg 1991) least squares modification (GGM01s, L = M = 67):
Joint contribution of the GGM01s errors and the truncation error, LS parameters consider C(0) as
follows:
C(0) = 0 &1 mGal² (solid bold line),
C(0) = 4 mGal² (lowermost thin line), and,
C(0) = 9 mGal² ( uppermost thin line), respectively.

Fig. 2.4d. Joint error when the GGM01s and M = 95 are utilised:
Unbiased LS modification, L = 95, (bold line),
Vaníček and Kleusberg, L = 67, (lowermost thin line),
Wong and Gore, L = 31, (uppermost thin line).

29
Hence, the greater the radius of integration the larger the magnitude of [ ], yielding a
large terrestrial data error. This necessitates, in contrast to the earlier conclusion, that
the integration cap size should be in reasonable balance when applying the simple
modification method. Note that the largest error is due to the first degrees of the
terrestrial data. Vaníček and Featherstone (1998) found that that the unmodified kernel
S(ψ) allows low-frequency terrestrial gravity data errors to pass, almost undiminished,
into the geoid. The WG and VK kernels attenuate these errors to a greater extent, but
not completely, however. Consequently, even when the residual anomalies are
employed in the Stokes integral it is impossible to eliminate (at least in that way), the
long wavelength errors in the geoid solution. Also a study by Ågren and Sjöberg (in
print) concludes that the simple r-c-r scheme with unmodified kernel S(ψ) is sensitive to
long-wavelength errors in gravity anomalies.
Vaníček and Featherstone (1998) show, also supported by this study, that the
methods with SL(ψ) are better when a small truncation cap is utilised. The results of the
following experiments reveal that LS modification methods allow reducing the expected
global MSE even further.

2.4.3 LS modification methods (GGM01s, L = M = 67)

The solid lines in Fig. 2.4c visualise the contribution of the joint errors for the unbiased
LS modification {note that different C(0) are utilised}. There are only insignificant
numerical differences among different LS modifications, but formally the global m N2
was the smallest for the optimum LS method. This observation is in agreement with
Sjöberg (2003d). When errorless terrestrial data is assumed, the expected m N2 of the LS
approach is almost two times less than those for the WG and VK methods; see Table
2.2. Conversely, the terrestrial data errors are similar to those of the deterministic
methods, cf. Figs. 2.4a and 2.4c. Consequently, the improved m N2 of the LS methods is
mainly due to reduced joint error. Further improvements can be achieved by reducing
the terrestrial gravity errors and/or by proper choice of the modification degree M
{recall the discussion after Eq. (2.30)}. Curiously, also the choice of C(0) affects, but
little, the joint error of the LS methods.
Obviously, with access to the GRACE geopotential models, the first part of the
joint error has lost some of its past significance. Furthermore, the gradiometric satellite
GOCE (Gravity and Ocean Circulation Explorer, to be launched during 2006 by the
European Space Agency) will provide unprecedented accuracy for geopotential
coefficients up to degree 270 (corresponding to the spatial resolution of 65 km).
However, for many practical applications a more refined regional geoid model may still
be necessary. Thereby the modification of Stokes’ formula does not lose its importance.
The terrestrial data errors become crucial and should be considered with the greatest
care (see also Ågren and Sjöberg in print). To illustrate this, compare Figs. 2.4a and
2.4c. The joint error of the deterministic methods is of the same range as the terrestrial
data error for C(0) = 4 mGal². When the LS modification is applied in conjunction with
the GGM01s then the joint and terrestrial data errors are of the same magnitude already
at C(0) = 1 mGal²; see Fig. 2.4c. Consequently, even small terrestrial data errors may
ruin the advantages of existing and future geopotential models in regional geoid studies.

30
2.4.4 Deterministic and LS modification methods (GGM01s, M = 95)

As already noted, the modification degree M influences the magnitude of the joint error.
Aiming at minimizing the global MSE, we tested also higher (than M = 67) degree
harmonics of GGM01s. Fig. 2.4d represents the case with the upper limit M = 95 of
GGM01s is utilised. At M = 95 the GGM01s harmonic error dcn becomes almost equal
to the gravity signal cn. The joint error of the LS method in the cm range was achieved.
Note, however, that this astonishing result for regional geoid modelling is based on a
rather unrealistic assumption that the terrestrial data is completely errorless. The joint
error of the deterministic models is reduced as well (compare with Fig. 2.4a), though to
a lesser extent. If we also account for the terrestrial errors, their contribution will be
almost similar to the ones in Figs. 2.4a and 2.4c, thus confirming the previous
discussion.

2.4.5 The expected global MSE, discussion

A variety of results for mN2 was obtained for different modification methods. Generally,
LS modifications give always superior results over the deterministic methods. With
errorless terrestrial data the LS results are almost twice better than the ones for the
deterministic methods. However, with the presence of terrestrial errors the differences
between the LS and deterministic methods degrade.
The results of this section, of course, are dependent on the reliability of the
applied stochastic models, see also discussion at the end of Section 2.3.5. Nevertheless,
we believe that the global MSE as evaluated here, may give a rather good idea about the
expected internal accuracy of different modification methods. It should emphasised that
one ought not overweight the results in Table 2.2, since the expected global MSE is
only a theoretical estimator, which need to be confirmed by some external datasets and
practical computations. Different results of geoid modelling can be expected in different
geographical regions. An external assessment of different modification methods can be
reached from the inter-comparison of the geoid models and some GPS-levelling data.
The results of such comparison are presented in Section 2.5.4.

2.5. Numerical investigations

Six methods {three deterministic and three stochastic (LS) modification methods} were
applied for the Baltic geoid modelling by Eq. (2.18). These methods are: simple (L = 0,
M = 360), WG (L = 31 and M = 67), VK (L = M = P = 67), biased, unbiased and
optimum LS modifications of Stokes’ formula. The expressions of the respective
modification coefficients (sn and bn) were presented in Table 2.1. Recall, that the LS
parameters sn depend on stochastic models and the adopted C(0). The LS geoid models
utilised four different (either considering C(0)= 0, 1, 4 or 9 mGal²) sets of parameters sn
for each LS modification method. Altogether 12 different LS geoid models are
computed. The final C(0) has not yet been specified, but C(0) = 1 mGal² is preferred
here. At the same time the parameters of the deterministic methods are invariant to these
quantities, thus only one geoid model for each deterministic method is computed.

31
 In order to map the geoid the target area is divided into 1.5´×3´ (similar to the
gravity anomaly grid) surface blocks. Totally, there are 250 rows in north-south and 170
columns in east-west directions, respectively. The free-air anomaly grid and ∆ĝn from
the GGM(s), along with the respective sets of the modification coefficients (sn and bn)
were inserted into Eq. (2.18), resulting in the geoidal heights. Inside each cell of
terrestrial data the continuous integrating function {SL(ψ)∆ĝ} is approximated to be
constant. The outcome of the unbiased LS modification is visualized in Fig. 2.5.

60
17 16

19 18
19
20

20

20
59

17
19
18
20
21
58 20
20
20

22
LATITUDE [°]

21

23 22 21
57
24 23
22

24 20
25
22
56

26 23
21
25
24 22
55
26 25
23
28 27
29 26
28
24

27
54 3 0

20 22 24 26 28
LONGITUDE [°]

Fig. 2.5. The Baltic gravimetric geoid model BALTgeoid-04. Computed from the unbiased LS
modification of Stokes’ formula. Upper limit of GGM01s and the modification degree of Stokes’ function
are both set to 67; terrestrial anomaly variance and correlation length are set to 1 mGal² and 0.1°,
respectively; integration cap size is 2°. Geoidal heights are given with respect to GRS-80. Unit is metre
and the contour interval is 0.5 m. The total area of the image corresponds to 300 000 km2.

32
The geoidal heights in the target area are decreasing towards the northeast, whereas the
extremes of 30 and 15.5 m are located in the southwest and northeast corners,
respectively (the length of this diagonal is ~800 km). The Baltic geoid model is mainly
smooth (with a standard deviation of the mean of 3 m), but it includes some local
irregularities in the NE part of the target area. Their occurrence is correlated with the
regional extremes of the gravity anomaly field (cf. Fig. 2.3).
From this point we focus on the five methods, which utilise SL(ψ) in the integral
and the GGM01s; see also PAPER A. The results of the simple modification method
will be discussed in Section 2.5.6.

2.5.1 Discrepancies between the geoid models

Even though the numerical tests involve five different modification methods the
discrepancies between any pair of the geoid models remain within ± 9 cm; see Fig. 2.6.
The maximum differences are discovered between the groups of stochastic and
deterministic methods. The stochastic geoid models are practically the same. The
outcomes of the unbiased and optimum LS modification methods match within 1 mm,
whereas their differences from the biased LS modification results range from +3.6 to
-5.2 cm (with a mean of -5 mm); see Fig. 2.7. At the same time the geoid models of the
WG and VK methods agree internally within ± 2 cm (with a mean of 0); see Fig. 2.8.
Subsequently, these two deterministic methods practically coincide.

60 2 0 60 0 0 2
8 -1 3
2 4
-1

6 2 0
-1

-1 1
-2 0
0
-1

-2
0
0

-2 -3
-4 0 -4 0 -1
-2

59 2 59 1
-4 -2
-6

-2 -1 2
0
2

-1

4 0 -2
-1
0

2 6 1 1
0 2
-2

58 -2
4

58 4
0

2
2

-1
-4

-4
0 1
-3
LATITUDE [°]

-3 -1
LATITUDE [°]

-2 0
0

-2

2 -1
-4

57 -4 -6 57
-2

-6 1
0 1
-1

0
-4

-2 -4 -1
-8

-1
-2

-4 -2 -4
-6

-6 0 -2
56 56
1

0
0

-
-4 -2 -1 2 -1 1
0

2 0 -3
-2
0

0 0 2
2 1 2
0 1
2

0
2

55 4 0 55 1
2
2

-2 -2 -1 -2
-1

0 2 2
4

-2

-4 -4 -3
54 -6 54
2

0
-4

20 22 24 26 28 20 22 24 26 28
LONGITUDE [°] LONGITUDE [°]

Fig. 2.6. Discrepancies between the unbiased LS Fig. 2.7. Discrepancies between the unbiased
and Vaníček-Kleusberg geoid models. Unit is cm and biased LS modifications. Unit is cm and
and the contour interval is 2 cm. The discrepancies the contour interval is 1 cm. Regional
range between ± 9 cm, regionally. maximum is +3.6 cm (brightest region) and the
minimum is –5.2 cm (darkest region),

33
2.5.2 Comparisons between regional geoid models utilising different GGM

One of the scopes of the present study is to assess the regional geoid improvement due
to the data from the GRACE mission. In contrast to the stochastic methods, the
modification parameters sn of the deterministic methods are invariant to the change of
underlying GGM. Consequently, if the terrestrial data in the area of interest remains
unchanged, then an enhancement of the regional geoid can only be deduced either by
the change of a GGM, or the limit M, or both {cf. Eq. (2.18)}. Both deterministic
methods were applied to compute the Baltic geoid model anew. Exactly the same
computational setup and local gravity data were employed, but differently, EGM96 is
used in the computations. The results are compared with the geoid models based on
GGM01s; see Fig. 2.9. As is expected, the deviations among the respective regional
geoid models are of long wavelength nature (recall that the modification degree M = 67
is utilized in all the cases), ranging from -6 to +17 cm, with a mean of +8 cm. Major
discrepancies of the geoid models occur in the western part of Latvia. However, slightly
different range of discrepancies may be expected when different upper modification
degree M is chosen; see e.g. a study in PAPER B. Nevertheless, in both cases the local
extreme of the discrepancies is related to the same location, i.e. West-Latvia. A
conclusion follows that the deviations among the recent geopotential models are more
crucial than the numerical discrepancies among the tested modification methods.

60 60 0 -4
1

6
8 4 -2
0

2
-2

-2
1

8 2 *
-2 0 10
59 59 10
-1

-1 1 12 6
4
12 8
14 14 12
0

58 -1 58 10
-1 16
6
0
LATITUDE [°]
LATITUDE [°]

0 0
16

57 57
8
14

16
1

1 12
14
14 10
56 56
1

12
12
10
1 1 10 8
8
6
8 6
55 55 4 4
0 0 6
0 0 4 2
2 2
0
-1 2 -2 0
-2
0

-1 0
54 54 0 -2 -4 -4
-1
20 22 24 26 28 20 22 24 26 28
LONGITUDE [°] LONGITUDE [°]

Fig. 2.8. Discrepancies between the Vaníček - Fig. 2.9. Discrepancies between the Wong-Gore
Kleusberg and Wong-Gore geoid models. Unit is geoid models utilising different geopotential models
cm, the contour interval is 1 cm. The (EGM96 is subtracted from GGM01s, both with
discrepancies range between ± 2 cm, regionally, L = 31 and M = 67). Unit is cm, contour interval is 2
cm. Black dots indicate the location of the GPS-
levelling points.

34
2.5.3 Verification with GPS-levelling data

As is well-known, a practical application of a geoid model is the transformation of

heights. A reasonable indication of the geoid model accuracy could be obtained from
the comparison with the GPS-derived and levelling heights.
Three sets of Estonian, Latvian and Lithuanian geodetic points (for their locations
see Fig. 2.9) are applied for an independent evaluation of the computed geoid models.
The average distance between 26 evenly distributed Estonian control points is 50 km,
whereas the combined error of GPS-derived and spirit levelled heights cannot be greater
than 2-3 cm. Note that most of these points are directly connected to the high-precision
levelling network and the geodetic heights are computed from the same GPS-campaign.
For more details see Ellmann (2001, Ch. 5.3.1). The Latvian and Lithuanian datasets
(53 and 110 points, respectively) have been provided by J. Kaminskis, State Land
Service of Latvia and E. Parseliunas, Vilnius Technical University, Lithuania. Those
control points are located more densely, whereas their accuracy seems to be rather
heterogeneous (e.g. a few obvious outliers were excluded from the final computations).
The geodetic coordinates of the control points are related to the respective national
realization of the new European Terrestrial Reference System ETRS-89. The spirit-
levelled normal heights of all points refer to the Baltic Height System 1977, which is
related to Kronstadt tide-gauge. Since Estonian territory is affected by the
Fennoscandian post-glacial rebound, this effect for Estonian points is reduced. For other
points zero land uplift is assumed. In the Baltic States normal heights (related to so-
called quasi-geoid) instead of orthometric heights (related to the geoid) are in use.
Rigorously, in such a comparison the separation between geoid and quasi-geoid has to
be considered. However, due the low elevation of the target area and relatively little
range of gravity anomalies, the maximum difference between the orthometric and
normal heights is only a few mm. Compared to the accuracy of the control points, this
correction is certainly negligible.
The GPS-levelling points form a surface, which here is called the “geometric
geoid model”. An i-th geometric geoid height, (Ngeom)i , is obtained at discrete points by
algebraically subtracting the levelling height Hi from a GPS derived height hi, i.e.
(Ngeom)i ≅ hi - Hi. The discrepancies between the geometric and gravimetric geoid
models are denoted by δN = Ngeom - N. For the evaluation of numerical outcomes we
utilise two well-known statistical terms: the standard deviation (STD) of the parameter
δNi and the root mean square error (RMS) of certain residuals (to be specified later).
The STD is calculated as

1 n´
σ STD = ∑
n´−1 i =1
(δ N i − δ N ) 2 , (2.31)

where n´ is the total number of points and δ N is the mean value. The RMS of residuals
v1, v2 … vn, is calculated as

1 n´ 2
σ RMS = ∑ vi
n´ i =1
. (2.32)

35
Table 2.3. Numerical statistics (STD and mean) of the comparison of the geoid models and national GPS-
levelling points (δN = Ngeom - N). The notation STD refers to Eq. (2.31). Unit is metre.

Country Statis- Unbiased LS Wong-Gore Simple

tics modification modification modification
L = M = 67 (L = 31 and M = 67) L = 0 and M =360
Baltic (joint) Mean -0.410 -0.411 -0.341 -0.429
189 points STD 0.058 0.058 0.090 0.077
Estonia Mean -0.387 -0.394 -0.308 -0.379
26 points STD 0.040 0.052 0.069 0.062
Latvia Mean -0.398 -0.414 -0.274 -0.393
53 points STD 0.060 0.070 0.070 0.067
Lithuania Mean -0.426 -0.413 -0.382 -0.458
110 points STD 0.057 0.053 0.080 0.071
Geopotential model GRACE Gravity Model GGM01s EGM96

The common Baltic geometric geoid is represented by the sum of the three
national datasets (altogether 189 points). Comparisons are also produced for each
country separately; see Table 2.3.
In all the cases a negative mean of differences δ N of the same range is detected.
This demonstrates thus that the Baltic height systems are in the average ~ 4 dm lower
than the gravimetric geoid models. Approximately the same results were obtained for
the geoid models utilizing EGM96. There could be many reasons for such an offset.
Further discussion on this subject is outside the scope of Chapter 2, but for more details
and other results, see PAPER B and Ellmann (2001).
Next we compare the STD of the deterministic (WG will be used in the
comparisons) methods with different underlying GGM. The STD of δN-sets are 5.8, 5.2,
7.0 and 5.3 cm for the Baltic, Estonian, Latvian and Lithuanian geoid models (based on
GGM01s), respectively. At the same time the corresponding STD values for EGM96
are 9.0, 6.9, 7.0 and 8.0 cm, respectively. In the former cases STD are thus notably
smaller from the ones utilising EGM96. This may most likely indicate the presence of
some long-wavelength uncertainties in the EGM96 solution, e.g. possible tilt between
EGM96 and the height system. For instance, in the computations of an earlier geoid
model for Estonia (Ellmann 2001, also in PAPER C) a small SW-NE oriented tilt (~30
cm to 350 km) was detected between the GPS-levelling data and the EGM96 based
regional geoid model. In other words, the smaller STD is a clear indication that
GGM01s is more consistent to the local data than the EGM96 based values. The last
column of Table 2.3 reflects the results for the simple modification method with a full
expansion of EGM96. The STD values are better than the WG modification employing
M = 67 and the model EGM96. Consequently, the long wavelength uncertainties of the
EGM96 harmonics below M = 67 can be suspected. Apparently these discrepancies are
somewhat corrected by the higher degree contributions of EGM96, but not completely,
however. At the same time the GGM01s model provides more or less correct long
wavelength contribution (see also discussion in Section 2.5.7).

36
2.5.4 Statistics of a four-parameter fitting

In the sequel we attempt to minimize the offsets (i.e. the vertical offset and possible tilt)
between the gravimetric and geometrical geoid models by introducing a polynomial fit.
In practice, the four-parameter model is often used; see Appendix C. The same sets of
GPS-levelling points are used for defining the transformation parameters between the
pairs of geoid models (i.e geometric and gravimetric). Thereafter, these parameters were
applied for fitting the gravimetric geoid models to the GPS-levelling points. As is
reported elsewhere, the post-fitting residuals are commonly used for evaluation of the
gravimetric geoid model accuracy. The numerical statistics, RMS, minimum and
maximum of post-fit residuals are presented in Table 2.4.
The best RMS values of the post-fit residuals are as follows: Baltic 5.3 cm (LS
modifications), Estonia 2.8 cm (unbiased LS), Latvia 5.6 cm (unbiased LS) and
Lithuania 4.2 cm (deterministic), respectively. It could be concluded thus, that the
accuracy of the tested modification methods is of the same level as is the accuracy of
the control points. Not coincidently the best statistics is related to Estonia, where the
quality of the GPS-levelling data is probably better than in the other data-sets.
Generally, the RMS of the post-fitting residuals of the LS modifications are slightly
smaller (up to 1 cm) compared to the ones of the deterministic methods. The only
exception is in Lithuania, where the accuracy of two deterministic methods is
marginally superior to any LS modification. Note that the same pattern can be observed
in Table 2.3 (STD values in columns 3 and 4).

Table 2.4. The statistics of verification of the geoid models with the GPS-levelling data, after the four-
parameter fit is applied. Notation RMS of the post-fit residuals refer to Eq. (2.32). Unit is metre.

Stochastic least Deterministic modification

squares modification
Post-fit
Unbiased Biased Vaníček- Wong - Gore Simple modification
Country resi-
Optimum Kleusberg Eq. (2.5)
duals
L = M = 67, L = M = 67 L = 31, M = 67 L = 0 L=0
C(0)= 1 mGal² M = 200 M = 360
The Baltic RMS 0.053 0.053 0.059 0.057 0.059 0.064 0.059
countries Min -0.132 -0.123 -0.146 -0.147 -0.133 -0.121 -0.180
(189 points) Max 0.155 0.168 0.192 0.189 0.194 0.205 0.226
RMS 0.028 0.031 0.034 0.036 0.035 0.065 0.038
Estonia
Min -0.059 -0.058 -0.062 -0.070 -0.066 -0.105 -0.076
(26 points)
Max 0.045 0.054 0.064 0.067 0.064 0.120 0.082
RMS 0.056 0.058 0.066 0.066 0.068 0.071 0.066
Latvia
Min -0.105 -0.103 -0.127 -0.126 -0.128 -0.171 -0.163
(53 points)
Max 0.145 0.156 0.180 0.178 0.182 0.149 0.155
RMS 0.047 0.044 0.043 0.042 0.043 0.052 0.046
Lithuania
(110 points) Min -0.104 -0.098 -0.079 -0.076 -0.076 -0.117 -0.121
Max 0.154 0.156 0.150 0.149 0.148 0.174 0.172
Geopotential model GRACE Gravity Model GGM01s EGM96 GGM01c EGM96

37
 Recall that each LS modification method applied four sets of parameters sn, which
considered different values of the terrestrial data variance, C(0) = 0, 1, 4 and 9 mGal².
The results with C(0) = 1 mGal² are presented in Table 2.4. The other LS geoid models
gave almost the same statistics of the post-fitting residuals. As a matter of fact,
variations of C(0) have only a minor influence to the quantities SL(ψ) and bn (see also
Chapter 3).
It is interesting to compare the results in Table 2.2 and Table 2.4. Apparently, the
global MSE for the cases C(0) > 1 mGal² of the former table are too pessimistic
compared to RMS in the latter. The global MSE and post-fit RMS are consistent for the
LS parameters with C(0) = 1 mGal². Therefore the LS parameters considering C(0) = 1
mGal² were preferred for the final solution.
As discussed in Section 2.3.5, there are some indications of systematic
discrepancies of the terrestrial data. These are naturally comprised in the low frequency
spectra, manifesting thus as an offset and/or a tilt. As is expected, the GPS-levelling fit
eliminates these long wavelength effects in a great deal. Remember, however, that
various GPS-levelling errors cannot be separated from the fitting results. For this reason
the global MSE is probably better suited for assessing the comparative accuracy
between different modification methods. The GPS-levelling data provides an
independent check and allows also to evaluate the suitability of the regional geoid for
some practical applications, such as the levelling by GPS. Since the LS modification
methods perform the best (see Table 2.4), thus the efforts put into minimization of the
global MSE resulted in notable rewards.
As a matter of fact, some additional geoid models adopting the GGM01s
harmonics M > 67 were tested as well. For the LS modification L = M was taken and for
the deterministic methods the predefined (see Section 2.3.2) L were retained, i.e. for the
VK and WG methods L = 67 and L = 31, respectively. The results were thereafter
compared with the GPS-levelling data. None of the experimentations with other upper
limits (incl. M = 95) produced any clearly better than our original choice, i.e. M = 67.
We therefore settled the solution with this choice of M.

2.5.5 Preferred modification method

The geoid model providing the “best” post-fit statistics could be preferred as a final
representation of the joint Baltic geoid. Considering also the study results in Table 2.2
we give our preference to the unbiased LS modification method of Sjöberg (1991),
which attempt to reduce all relevant errors in geoid modelling. The modification
parameters sn and the geoid model of this geoid are calculated from the following initial
conditions: (1) upper limit of GGM01s and the modification degree of Stokes’ function
are M = L = 67; (2) terrestrial anomaly variance and correlation length are set to 1
mGal² and 0.1°, respectively; (3) integration cap size is 2°.
The correction terms accounting for the Earth’s ellipticity and the contribution of
topographic and atmospheric masses are computed separately by means of the additive
corrections (see Chapter 4). These corrections are added to the geoid modelling
outcome. The new geoid model for the Baltic countries is named to as BALTgeoid-04.
As a matter of fact, BALTgeoid-04 (including all additive corrections) has already been
presented in Fig 2.5.

38
2.5.6 Comparison with the NKG results

The NKG96 (Forsberg et al. 1997) geoid model covers entirely the target area. The
standard deviation after a four-parameter fit for Lithuania (36 GPS-levelling points) and
Latvia (26 points) were 0.078 and 0.077 m. Including the most recent terrestrial data
and using exactly the same computational setup as for the NKG96 model, the new geoid
model (Forsberg 2001) was computed. The standard deviation of the post–fitting
residuals for Estonia (31 points) were 0.033 m. A recent study by Jürgenson (2003) is
devoted to the computations of the Estonian geoid model. He utilized the NKG geoid
computation software and the full expansion of EGM96, but differently from the
original NKG geoid models, significantly improved and very dense gravity coverage
from Russia. The number of GPS-levelling datapoints used for assessing the resulting
geoid was almost 50 (ibid. p. 114). The standard deviation of post-fitting residuals was
2.1 cm, which was achieved by a careful selection of the GPS-levelling data. Obviously,
very scrupulous handling of the new data has been advantageous as well. The accuracy
estimates of the other (listed in Section 2.3.4) geoid models for the same region range
between the aforementioned statistics.
In the present study the computations applied also the simple modification {see
Eq. (2.5), equivalent to Vincent and Marsh (1974) modification} and three geopotential
models - GGM01s, GGM01c and EGM96. These results are validated by the same sets
of GPS-levelling points (see Section 2.5.3). Our attempts to exploit M = 67 and M = 120
(the latter is the upper limit of GGM01s) and M = 200 (upper limit of GGM01c) did not
provide satisfactory results, i.e. the statistics of the models utilising Eq. (2.5) were
worse (generally, the less M the greater RMS) than the ones for the five methods with
SL(ψ); see also Table 2.4. As is discussed in Section 2.4 this is suspected mainly due to
large truncation error of the simple modification method, whereas in the methods
utilising SL(ψ) this error is efficiently mitigated. [Perhaps this problem for the methods
with S(ψ) is less critical when data from more extended areas would be used in the
integration.] Only the complete expansion of EGM96 (i.e. M = 360) provides
comparable post-fit accuracies with other results in Table 2.4. A direct conclusion
follows, that the present GRACE models do not qualify for the precise modelling by
application of Eq. (2.5). Possibly, even future geopotential models from GOCE data
alone (resolution up to degree of 270 expected) also would not be satisfactory in Eq.
(2.5). This yields that simple modification method could benefit from the new satellite
missions only if a high-degree (at least up to M = 360) geopotential model will be
elaborated. Recall that this necessitates the involvement of worldwide terrestrial data,
which brings along a lot of undesired problems.
Even though the results of the present study are generally superior to the NKG
joint models, there are several circumstances preventing direct comparison of the
results. These are:
(1) The fast Fourier transform (FFT, see e.g. Forsberg and Sideris 1993, Haagmans et
al. 1993) method is utilised in the computations of the NKG geoid models. Most often
all the data-points from rather large (rectangular-shaped) region are involved for
computing a single geoidal height. In contrast, in this study the data is restricted to a
spherical cap around each computation point.
(2) The number of Latvian and Lithuanian GPS-levelling points has remarkably
increased during the years after the NKG96 geoid model was computed.

39
(3) The NKG solution of Forsberg (2001) includes a Russian dataset, which was not
available for this study.
(4) The NKG96 computations yield height anomalies, whereas the outcomes of the
present study are geoidal heights.

2.5.7 Assessment of the geopotential models

The deterministic models utilizing EGM96 have almost similar RMS (a few mm larger,
though) as the models employing GGM01s. Since the fit to the GPS-levelling data
removes the long wavelength discrepancies, thus the quality of an underlying GGM
cannot be judged from the statistics of the post-fit residuals alone. Another comparison
is more useful. Note that the RMS of the post-fit residuals is reduced, of course,
compared to the corresponding STD value of (pre-fit) differences δN (Table 2.3).
Intuitively, the less the difference between pre- and post-fit statistics the better the
underlying GGM. An insignificant difference of the STD and post-fit residuals indicate
most likely an one-dimensional offset between the GPS-levelling data and the geoid
model. Revealed parallelism between the practical realisation of the national vertical
datum and the geoid models adopting GGM01s, gives the impression of greater
reliability, in contrast to somewhat tilted, EGM96 based geoid models. This comparison
indicates that the GGM01s is superior to EGM96 in the target area. Slightly different
numerical results can be expected when using different modification limit M, although
the same conclusion is reached; see PAPER B.

2.5.8 Practical considerations

It should be emphasised that the four-parameter polynomial fit of the GPS-levelling

data served only to assess of the geoid modelling outcome. For practical consideration
(e.g. levelling by GPS) the fitting of geometric and gravimetric geoid models could be
developed further, e.g. by constraining GPS-levelling points and adjusting the surface of
the gravimetric geoid between the control points. This suggests the combined
adjustment of the GPS-levelling and geoid models, i.e. by the method of Kotsakis and
Sideris (1999), or Jiang and Duquenne (1996), or least squares collocation (Tscherning
et al. 1992), etc. Strictly speaking, the resulting new surface is not a geoid model any
more, since it has no physical meaning after such a procedure. In the literature this type
of surface are sometimes called “height reference surface”.
In the Baltic countries the normal heights, thus related to the quasi-geoid, are used
in surveying practice. The quasi-geoid can be easily obtained applying very little (a few
mm in the target area) corrections to the BALTgeoid-04. The geoid to quasi-geoid
separation could be determined by a relation (as a function of Bouguer anomalies and
orthometric height) in Heiskanen and Moritz (1967, p. 327), more rigorously this
problem is treated in Sjöberg (1994 and 2000).

40
2.6 Conclusions
The modified Stokes’ formula employs the long wavelength geoid component from a
global geopotential model in conjunction with the regional terrestrial gravity data in a
truncated form of Stokes’ integral. In order to compute a high-resolution geoid model
for the Baltic countries six different (three deterministic and three stochastic)
modification methods were tested. The deterministic methods are Wong and Gore
(1969), Vincent and Marsh (1974) and Vaníček and Kleusberg (1987). The stochastic
methods are biased, unbiased and optimum LS modifications by Sjöberg (1984b, 1991,
2003d). The deterministic methods utilise originally the r-c-r scheme. For the sake of
the comparison, those deterministic methods are expressed such that the original surface
gravity anomaly, instead of the residual anomaly, is exploited in the Stokes integral.
Five methods utilise the modified Stokes function SL(ψ) as an integral kernel, whereas
the original (non-modified) Stokes function is used by one modification method. The
deterministic modification parameters for the WG and VK methods are selected such
that SL(ψ) becomes zero at the edge of predefined truncation cap.
The definition of the geoid modelling parameters is a rather delicate task and
requires a thorough consideration of the characteristics of the target area and available
datasets. Depending on the local gravity data quality, the chosen radius of integration
and the characteristics of an underlying GGM, the LS modification parameters sn vary.
A tentative assessment of different modification methods (and suitability of
computed modification parameters) is done by means of the expected global MSE. The
LS modification methods reduce the global MSE the most.
The accuracy of the low and intermediate frequency bands of GGM01s can be
considered almost errorless compared to the range of the terrestrial gravity data errors.
A conclusion follows that with access to up-to-date geopotential models the terrestrial
errors become more important in the modified Stokes formula. Consequently, even
small terrestrial data errors may ruin the advantages of existing and future geopotential
models in regional geoid modelling. This implies that the terrestrial data limits are the
next barrier, which has to be overcome for further improvement of regional geoid
models.
The modification methods with SL(ψ) were applied for the purpose of computing
geoid models for the Baltic countries. Interestingly, although the modification
parameters are defined differently, the numerical discrepancies between any pair of the
resulting geoid models remain within ± 9 cm in the target area.
The GPS-levelling data was applied for an independent evaluation of the accuracy
of the geoid models. The fitting of the geoid models to the GPS-levelling points reveals
that the accuracy of the tested modification methods is of the same level as is the
accuracy of the control points. Generally, the post-fit residuals from the LS
modification are slightly smaller (up to 1 cm) than the respective values of deterministic
methods. Theoretical superiority of the LS methods in the test area is thus proved by the
practical computations. However, in one country (Lithuania) the accuracy of both
deterministic modification methods is superior to the LS modification.
To secure similar accuracy for the geoid models by the simple modification
method, a high degree expansion of a GGM is required. This yields that the ”satellite-
only” geopotential models do not qualify as a basis for high-resolution regional geoid
models. All modification methods combining the modified Stokes function SL(ψ) and an

41
intermediate expansion of the geopotential models provide a reasonable accuracy even
for relatively small cap sizes.
An appropriate geopotential model is essential to determine the regional
gravimetric geoid model accurately. The new GRACE Gravity Model GGM01s is
exploited in the computations. There are some notable improvements of numerical
statistics (assessed by the GPS-levelling data) in the target area when utilizing the new
GGM-01s, instead of using EGM96. Unlikely to the regional models employing
EGM96, the more realistic one-dimensional offset between the GGM01s-related models
and the height system is identified in the target area. A conclusion follows that the
deviations between the two geopotential models may be more significant than the
numerical discrepancies between the modification methods.
Obtained results demonstrate, that the LS modification of Stokes’ formula is well
suited for computing high-resolution and accurate geoid models. The outcome of the
unbiased LS modification is completed with additive corrections, which have to be
introduced in the geoid determination process (see Chapter 4). This gravimetric geoid
model is named BALTgeod-04. The model is fitted to three sets of GPS-levelling
points, yielding the RMS values from 2.8 cm to 5.6 cm for the post-fitting residuals.
This order of discrepancies is sufficient for many practical applications.

42
Appendices
Appendix A. Equality from Sjöberg and Hunegnaw (2000, Eq. 3).

The generalised Stokes scheme by Vaníček and Sjöberg (1991) utilises the modified
Stokes function and residual gravity anomaly in the integral. The geoid estimator Ñ1 is
provided by {cf. Eq. (2.8)}

c  M
 M
2
N 1 =
2π ∫∫σ S L
(ψ ) 

∆ ˆ
g − ∑
n = 2
∆ ˆ
g n

d σ + c ∑
n = 2 n − 1
∆gˆ n . (A.1)
0

Sjöberg and Hunegnaw (2000) presented an equivalent to Eq. (A.1) method,

which uses the full gravity anomaly instead of the residual anomaly. Hereby we present
the derivation given in the original source (ibid.). Eq. (A.1) could be rewritten as

c  M
 M
2
N 1 =
2π ∫∫ S L
(ψ ) 

∆ ˆ
g − ∑
n=2
∆ ˆ
g n

d σ + c ∑
n =2 n − 1
∆gˆ n =
σ0
M M
. (A.2)
c c 2
= ∫∫σ S (ψ )∆gd
ˆ σ− ∫∫σ S (ψ )∑ ∆gˆ n dσ + c ∑ ∆gˆ n
L L

2π 0
2π 0
n = 2 n = 2 n − 1

The following relations are used

2π

∫∫σ = ∫∫σ −σ∫∫σ −

; ∫ d λ = 2π
λ =0
, (A.3)
0 0

and

M M
c  2 
2π ∫∫σ S L
(ψ ) ∑
n =2
∆gˆ n d σ = c ∑ 
n=2  n − 1
− sn  ∆gˆ n

M M
. (A.4)
c
∫∫ S (ψ )∑ ∆gˆ n dσ = c ∑ Q ∆gˆ n
L L

2π
n
σ σ
− 0
n =2 n =2

After applying the orthogonality principle of surface spherical harmonics, one arrives at

M
R
N 1 = N 2 = ˆ σ + c ∑ (QnL + sn ) ∆gˆ n ,
∫∫ S (ψ )∆gd
L
(A.5)
4πγ σ0 n =2

which holds for any set of parameters sn.

43
Appendix B. Models for gravity anomaly and data error degree variances

This appendix reviews some models for gravity anomaly and data error degree
variances, which can be used for estimating the global MSE by Eq. (2.22) and the LS
modification parameters (Chapter 3). The gravity anomaly degree variances are denoted
by cn, whereas the terrestrially measured and GGM-derived anomaly error degree
variances are denoted by σ n2 and dcn, respectively. This study applies some well-known
models for those quantities. Generally, the data errors could consist the random and
systematic error contribution. As is mentioned in Section 2.2, possible systematic errors
are neglected for the stochastic models of this study. This is supported by common
practice, that all known systematic errors are removed from computations, rather than
accounted for separately. Concerning unknown systematic discrepancies, these remain
unknown indeed and estimation of their actual range could be unsurmountable
challenge (and/or very often just a speculation).

Gravity anomaly degree variances.

The degree variance cn can be obtained by using coefficients ∆Cnm , Snm of disturbing
potential and fundamental constants (gravity mass constant GM and equatorial radius a)
of a GGM as follows:

(GM ) 2 n
cn =
a 4
( n − 1) 2
∑
m =0
( ∆Cnm
2
+ Snm
2
) . (B.1)

In practice the infinite sum in Eq. (2.22) must be truncated at some upper limit of
expansion. In this study nmax = 2000 is adopted. This is far beyond available GGM
harmonics. Higher degree cn could be generated synthetically to conform to the
predicted spectral characteristics of the Earths gravity field (e.g. following Kaula’s rule,
or some other model). In order to calculate the values cn for degrees n > M, we use the
formula and coefficients A’ = 425.28 mGal² and t = 0.999617 given by Tscherning and
Rapp (1974):

(n − 1)
cn = A '⋅ t n + 2 . (B.2)
(n − 2)(n + 24)

It should be noted that cn in Eq. (B.1) and (B.2) are for the gravity field not corrected for
any topographic effects (recall insignificant topography in the target area).

Geopotential harmonic error degree variances.

The degree variance dcn can be estimated from using standard error of the potential
coefficients δ Cnm , δ Snm (cf. Rapp and Pavlis, 1990):

(GM ) 2 n
dcn =
a 4
( n − 1) 2
∑
m =0
(δ C2nm + δ S2nm ) . (B.3)

44
Note that the coefficients δ Cnm , δ Snm are a natural part of many GGMs. Past geopotential
models, such as the EGM96, were utilising rather heterogeneous data-sets. As a matter
of fact, the variance by Eq. (B.3) is global and so not necessarily representative for the
area of interest. For instance, in the regions of a dense coverage (such as Fennoscandia)
of terrestrial data, that also goes into the solution of a GGM, the resulting dcn may be
too pessimistic. In order to obtain more realistic estimates for such a regions, the
variance dcn could be re-scaled, e.g. “tailored” to the area of interest. This study,
however, employs the new geopotential model GGM01s for the computation of the
regional geoid. This model is compiled from the GRACE satellite data and is assessed
to be highly accurate and homogeneous, with no distinction between offshore or
terrestrial areas. Thus, such a scaling procedure is not necessary in this study.

Terrestrial data error degree variances.

As it was emphasized in Section 2.4, the errors of the terrestrial data are very crucial in
the geoid modelling process. The terrestrial gravity anomaly ∆ĝ error degree variance
σ n2 can be estimated from an error degree covariance function C(ψ). The following
function could be used (cf. Heiskanen and Moritz 1967, p. 256)

∞
C (ψ ) = ∑σ n2 Pn (cosψ ) , (B.4)
n =2

where Legendre polynomials Pn(cosψ) are applied. The error degree variance of the
terrestrial gravity anomaly σ n2 is thus assumed to be of reciprocal distance type, which
can be estimated from the simple relationship (cf. Sjöberg 1986, Ch. 7)

σ n2 = cT (1 − µ ) µ n , 0 < µ < 1 , (B.5)

where the constants cT and µ can be estimated from a knowledge of the function C(ψ).
For estimation of σ n2 some assumptions for the error variance C(0) (the value of the
covariance function at ψ = 0) and the correlation length ψ° (to the point where C(ψ) is
degraded to half of C(0), thus C(ψ°) = 0.5C(0), see Moritz 1980, p. 174) need to be
adopted. In this study the correlation length 0.1° and the closed expression for the
covariance function C(ψ) in Sjöberg (1986, Eq. 7.2)

 1− µ 
C (ψ ) = cT  − (1 − µ ) − (1 − µ ) µ cosψ  , (B.6)
 1 − 2 µ cosψ + µ
2


are applied. If ψ = 0 then the variance by Eq. (B.6) becomes

C (0) = cT µ 2 , (B.7)

and thus it follows that

45
 1
C (ψ 0 ) = cT µ 2 . (B.8)
2

The solution for µ can be found iteratively. Inserting µ into Eq. (B.7) yields that
function C(ψ) is completely determined. Thereafter the degree variances σ n2 are
computed from Eq. (B.5).
Another estimators of σ n2 , e.g. by Wenzel (1981), Weber and Wenzel (1983) are
worth of mentioning. It should be noted that the latter approach is not isotropic, since
the error degree variances become also dependent on the azimuth. The advantages and
drawbacks of different estimators of the error degree variance, σ n2 , have been discussed
in the geodetic literature (e.g. Rummel 1997). This study, similarly to some earlier
studies at the KTH (e.g. Nahavandchi 1998, Ellmann 2001), utilises the Sjöberg’s
approach.
Note that degree variances σ n2 are to be used for estimating the global MSE. On
the other hand, regional geoid computations involve only local data, which may
drastically differ from the global average. For this reason it is recommended to define
such values for C(0), which correspond to the data actually involved in computations;
for more details see Section 2.3.5.
Stochastic models are essential not only for estimating the global MSE, but also
for defining the LS parameters and consequently also for the geoid modelling. In this
respect a study by Ågren and Sjöberg (in print) is credited. They studied some issues
when terrestrial data is utilized by different modification methods. Among other
essential findings they concluded, that the definition of a priori error model for the
terrestrial gravity anomaly is not a major problem for a successful application of the LS
methods.

Appendix C. Fitting of geometric and gravimetric geoid models

Comparisons of geoid models with GPS-levelling data usually apply the following basic
model (cf. Kotsakis and Sideris 1999):

hi − H i − N i = aTi x + ε i , (C.1)

where x is a vector of unknown parameters, aiT is a vector of known coefficients and εi

denotes an i-th residual random noise term, h is GPS-derived geodetic height, H
orthometric height and N gravimetric geoid height, respectively
The main problem under this approach is that εi terms will contain a combined
amount of GPS, levelling and geoid errors. Therefore it is very difficult to separate of
the various random and systematic effects (e.g. possible datum inconsistencies) between
aiTx and εi. In practice, the usual four-parameter polynomial fit is often used, i.e.

aTi x = x0 R + x1 cos ϕ i cos λi + x2 cos ϕ i sin λi + x3 sin ϕ i , (C.2)

46
where ϕ and λ are geodetic latitude and longitude, respectively. That corresponds to the
datum transformation model, which is described in Heiskanen and Moritz (1967, Ch. 5-
9). In our case, the two different datums correspond to the GPS-levelling datum and the
datum of the gravimetric geoid model. Three parameters x1, x2, x3 correspond to the
coordinate shifts, whereas the parameter x0 multiplied by the Earth’s radius R could be
interpreted as the scale factor. This kind of transformation, of course, is not a rigorous
coordinate transformation, since the parameters will absorb various errors as well.
In order to determine the “best fitting” solution the unknown regression
parameters in Eq. (C.2) need to be defined. Due to the errors in both datasets (i.e.
“geometric” and gravimetric geoid models) there is an infinite number of solutions for
regression parameters. For the minimization of the residuals εi, the least squares method
is commonly applied. The aim is to express each observation bi as a function of m
independent unknown parameters (corresponding to m necessary observations) and
estimate these parameters under the least squares principles. The solution for the vector
of unknown parameters x is obtained under the least squares conditions

ε T Pε = min . (C.3)

When defining regression parameters we take the vector of observations b as vector of

detected n differences between geometric and gravimetric geoid models, the vector of
residuals is ε, A is the design matrix and x is the vector of m unknown regression
parameters, such as

b - ε = Ax . (C.4)

Although the accuracy of GPS-levelling points can be various we assume here that all
elements of vector b are having same weight, i.e. matrix of weights P is an unit matrix
I. Since in the present study this information is used only for verification of the geoid
model accuracy, thus the possible influence of this simplification is equal to all
modification methods. If one aims to use the fitting procedure for practical applications
(e.g. levelling by GPS), the weighting issues need to be solved in more rigorous way.
The estimator of unknown parameters x is found, e.g. for the four-parameter
fitting by Eq. (C.2), as follows

xˆ = [ x0 x1 x2 x3 ] = ( A T A ) A T b .
T -1
(C.5)

After x has been estimated, one can easily obtain the least squares estimate of vector of
residuals ε as

εˆ = b - xˆ =  I - A(A T A) -1 A T  b . (C.6)

The residuals ε̂ are traditionally taken as the external indication of the geoid model
accuracy.
The accuracy of estimated regression parameters xi can be calculated by means of
a posteriori variance factor σ̂ 0 and variance- covariance matrix of estimators. The a

47
posteriori variance factor can be estimated from the least squares residuals ε̂ , the
number of GPS-levelling points n and number of parameters m as

εˆ Tεˆ
σˆ 0 = , (C.7)
n−m

and the variance-covariance matrices of the estimators are expressed as

ˆ ˆ = σˆ 0 ( A A )
2 T −1
Cxx
. (C.8)
Cεεˆ ˆ = σˆ 02 I - A(A T A)-1 AT 

The standard error of the each estimated parameters is calculated as

σˆ x = σˆ 0 Cx x .
i i i
(C.9)

Estimation of a posteriori variances can be produced for each set of the GPS-levelling
points. Such an analysis, accomplished with typical numerical results for one of the
countries (Estonia) under study can be found in Ellmann (2001, Ch. 5.6).

48
Chapter 3
_____________________________________________________________________________________
Determination of the least squares modification parameters

3.1 Expressions of the least squares modification methods

As already discussed in Chapter 2, the geoid models are dependent on the properties of
the modification procedure, i.e. the choice of the integration cap radius, upper
modification limits (L and M) and the coefficients sn and bn. The present chapter is
devoted for the practical determination of the LS modification parameters. First of all,
the LS parameters sn intend via minimising the expected global MSE to reduce all
relevant errors in geoid modelling. Reasoning this, the expression for the global MSE,
Eq. (2.22), is differentiated with respect to sn, i.e. mN2 / ∂sn . The obtained formula is then
equated to zero and the modification parameters sn are solved in the least squares sense
from the linear system of equations:

∑a
r =2
kr ⋅ sr = hk , k = 2,3,….L , (3.1)

where akr and hk are modification coefficients, which can be expressed via certain
coefficients Qn , Enk , cn , dcn and σ n2 ; see Table 3.1. The models for the gravity anomaly
degree variance cn, error degree variance dcn of the geopotential harmonics and
terrestrial data error degree variance σ n2 can be found in Appendix B of Chapter 2.
Chapter 3 focuses on the biased, unbiased and optimum methods of the LS modification
(Sjöberg 1984b, 1991 and 2003d), which basically deviate from each other in the choice
of parameters bn. Recall also the discussion at the end of Section 2.2. After the LS
parameters sn are defined, the expected global MSE, mN2 , can be estimated by an
equivalent expression {cf. Eq. (2.22)}:

mN2 = f − c 2sˆT h , (3.2)

where ŝ and h are the vectors of coefficients and

 2 ∞ 2
 2 
f = c ∑  2
− Qn  σ n + Qn2 qn  . (3.3)
 n − 1
n =2   

The last term qn is presented for each LS method in Table 3.1.

49
Table 3.1. Expressions of the least squares modifications of Stokes’ formula. The notations refer to Eqs.
(3.1)-(3.3), Table 2.1, and Appendix B (Chapter 2).

Para- Biased Unbiased Optimum

meters Sjöberg (1984b) Sjöberg (1991) Sjöberg (2003d)

b n= sn sn + QnL
(s n + QnL ) cn
cn + dcn
∞ ∞
2c 2 ∑  ( dcn + σ n2 ) sn* − ( cn + σ n2 ) QnL Enk
∞

∂mN2 2c 2 ∑  d n* sn* − d n*QnL Enk + Ω n Enk − 2c 2 ∑ (δ nk − Enk ) ×

= n=2 n =2 n =2
∂sn + Ω n Enk − Ω n + (QnL − sn* Enk ) σ n2  − Ω n + (QnL − sn* Enk ) d n  × ( QnL + sn* ) Cn − Ω n 

(σ r2 + dcr )δ kr − Ekrσ r2 − Erkσ k2 + ∞ ∞

akr = ∞ d k*δ kr − Erk d r* − Ekr d k* + ∑ Enk Enr d n* δ kr Cr − Ekr Ck − Erk Cr + ∑ Enr Enk Cn
+ ∑ Enk Enr (σ + cn )
2
n n =2 n =2
n=2
∞ ∞
Ωk − Qk d k* + ∑{Qn d n* − Ωn }Enk
∞
hk = Ωk − Qkσ k2 + ∑ Qn (σ k2 + cn ) − Ω n Enk Ωk − Qk Ck + ∑(QnCn −Ωn )Enk
n=2 n =2 n =2

L
sr = ∑a
r =2
kr ⋅ sr = hk , k = 2,3,….L

∞  2

 2 
mN2 = c 2 ∑  − Qn  σ n2 + Qn2 qn  − c 2sˆ T h
 n − 1
n=2   
 cn dcn
dcn ,if 2 ≤ n ≤ M  , if 2 ≤ n ≤ M
q n= cn   cn + dcn
cn otherwise c otherwise
 n
 s , if 2 ≤ n ≤ L 1 , if k = r ; *  dck , if 2 ≤ k ≤ M ;
sn* =  n ;δ = d k = σ k2 + 
 ck dck
, if 2 ≤ k ≤ M
 
Ck = σ k2 +  ( ck + dck )
0 otherwise
kr
 0 , otherwise c
 k , otherwise
c , if k>M
 k
π π
2k +1 2σ 2
Enk = enk ; Qn = ∫ S (ψ ) Pn (cosψ ) sinψ dψ ; enk = ∫ Pn (cosψ ) Pk (cosψ ) sinψ dψ ; Ω n = ; c =R/2γ n

2 ψ0 ψ0 n −1

3.2 Solving the LS parameters

As is emphasised in Sjöberg (1991 and 2003d, also in PAPER D) some difficulties may
be encountered when practically computing the LS modification parameters. To
illustrate this the system of Eq. (3.1) can be re-written in the matrix form as

Ax = b , (3.4)

where vector b consists of the modification coefficients hk, the symmetrical matrix A
comprises the coefficients akr = ark and x is the vector of unknown modification
parameters sr. For certain LS methods the matrix A may become numerically ill-
conditioned. In particular, the unbiased and optimum LS methods tend to become
extremely unstable. This is due to the occurrence of large clusters of very small values
in A, turning thus the problem into an ill-conditioned one. This also pertains that
rounding errors may be greatly amplified during the inversion of x = A-1b, hence
leading to a numerically unstable solution of x.
Some alternative methods, avoiding the matrix inversion, are acknowledged as
well. Since the matrix A is symmetrical, the Cholesky method for the solution of normal

50
equations (see Bjerhammar, 1973, p. 328) may be applied for defining LS parameters.
However, when solving the LS parameters the matrix is still ill-conditioned and thus the
focus of this study is given to another method - singular value decomposition (SVD).
A problem is improperly posed if it does not meet at least one of the following
three requirements: existence, uniqueness, or stability of the solution. In our case Eq.
(3.4) may become numerically unstable when a small integration cap is adopted for
computations. A way out of this problem is to incorporate some kind of regularization
into the solution procedure. Regularization usually aims at minimising the norm of
difference between two vectors, i.e. min||Ax-b||2. Simultaneously, the elements of x are
often desired to be stabile and thus the solution’s norm ||x|| should have reasonable size.
In all regularization problems the reduction of the solutions norm ||x|| is achieved by
allowing a larger residual norm ||Ax-b||.
A review of the numerical methods, which are feasible for solving the LS
parameters sn from an ill-conditioned system is presented in Section 3.3. Note that some
symbols of the mathematical inversion tools may denote a completely different quantity
in the LS modification theory. Therefore the preference hereafter to the end of Chapter
3 is given to the standard notation of the numerical treatment of ill-conditioned
problems. Most importantly the symbol n in the sequel corresponds to L (i.e. n = L-1,
since the parameters in this study start with s2, see Section 2.1) in Eq. (3.1). For
simplicity, the weight matrix P is adopted to be equal to the unit matrix, i.e. P = In
throughout this study.

3.3 Regularisation methods

3.3.1 Singular value decomposition

The SVD of the coefficient matrix A is a very important and useful numerical tool for
solving the ill-conditioned problems. Singular values are also known as eigenvalues of
the matrix A. The coefficient matrix A can be computed by means of the SVD

n
A = UΣV T = ∑ uiσ i v Ti , (3.5)
i =1

where U = (u1, …,un) and V = (v1…vn) are matrices with orthonormal columns (ui and
vi, respectively) satisfying the condition UTU = VTV = In. The diagonal matrix Σ has
non-negative singular values of A appearing in non-increasing order such that
σ1 ≥ σ2 ≥ … ≥ σn. For more details see Bjerhammar (1973, Ch. 27.9; also Hansen 1998).
Importantly, the sub-routines for computing the SVD are available in many
mathematical software packages (e.g. Matlab) today. With application of the SVD the
ordinary solution x = A-1b of Eq. (3.4), can be re-written

n
uTi b
A -1b = VΣ-1U T b ⇒ x LS = ∑ vi . (3.6)
i =1 σi

The use of the SVD in the analysis of ill-conditioned problems goes back to the
beginning of 1970-ies. For a more extended review see Hansen (1998). In the

51
mathematical literature (e.g. Hansen 1998) the ill-conditioned problems may be divided
into two classes: rank-deficient and discrete ill-posed problems, respectively. For
instance, the rank deficient problems are characterized that there are one or more small
(but non-zero) singular values well separated from the large ones. Conversely, the SVD
components of discrete ill-posed problems decay gradually to zero. However, in the
sequel we do not distinguish between the two type of the problems, because: (1) the
rank-deficient problems occurring in this study are most likely only numerically rank-
deficient and, (2) the methods under study provide quite similar results for the LS
parameters.
A useful indicator of non-stability, or ill-conditioning, is the condition number of
the matrix A, which is defined as the ratio between the largest and the smallest singular
values: cond(A) = σmax,/σmin. Obviously, matrices with 1/cond(A) less than the
computational precision of the computers are very ill-conditioned. Intuitively, the
solution norm ║xLS║ becomes very large due to near-infinity terms σi in denominator of
Eq. (3.6). Without any stabilization the solution will be therefore dominated by the last
SVD elements.
Two approaches suitable for solving the LS parameters will be presented in
Sections 3.3.2-3. These methods are truncated SVD and Tikhonov regularization. Both
can be expressed by means of SVD. For the facts not referenced in the text of Section
3.3, the source is Hansen (1998).

3.3.2 Truncated SVD

The truncated SVD (T-SVD) approach has been traditionally applied for the ill-
conditioned problems when the SVD values of the coefficient matrix A decay abruptly.
The most demanding task in connection with this problem is to compute the truncation
limit of the SVD values. For instance, the truncation parameter k is often defined as the
number of strictly positive SVD values, (i.e. k equals to the amount of the largest SVD
components before the “gap” only). Consequently, the idea of the T-SVD is to simply
treat the small singular values of A as zeros. For example, a T-SVD solution xk is
achieved by truncating Eq. (3.6) at k < n, i.e.

k
uTi b
xk = ∑ vi . (3.7)
i =1 σi

3.3.3 Tikhonov regularization

Tikhonov (1963) regularization can be applied for solving such ill-conditioned

problems, which are charaterised by gradual descent of the SVD values and with a large
condition number of matrix A. According to Hansen (1998, Ch. 5) this method can also
be applied where there is a gap in the SVD values. The Tikhonov approach is probably
the best investigated regularization method. For a review of other useful regularization
methods, see e.g. Hansen (1998).
Even though the matrix A for defining the LS modification parameters is always
symmetrical, we start with a more general A, with dimensions m × n, whereas m > n.

52
The expressions for symmetrical A can be easily developed from the general formulas.
More specifically, the system for the LS parameters is posed as in Eq. (3.12), yielding a
slightly different expression for a filter factor.
The Tikhonov regularization is more complicated than merely filtering out a
cluster of small singular values. In practical treatments, it is therefore necessary to
incorporate some kind of side constraints into Eq. (3.4). The Tikhonov regularization
scheme can be expressed as (cf. Hansen 1998, p. 72)

{
min Ax - b
2
2
+ λ 2 Lnx
2
2 }, (3.8)

where λ is the regularization parameter and Ln is a diagonal weighting matrix

(throughout this study the standard Ln is assumed, i.e. the identity matrix Ln = In). The
Tikhonov regularized solution, xreg, can be formulated by the normal equation matrix
(ATA) stabilized by λ2In :

x reg = ( A T A + λ 2I n ) A T b .
-1
(3.9)

Recall that an ordinary LS solution equals x = (ATA)-1ATb. The solution of Eq. (3.9)
can also be computed by means of the SVD of A, {cf. Eq. (3.5)}

n
uTi b
x reg = ∑ f i vi . (3.10)
i =1 σi

where the numbers fi are filter factors of the regularization method. The filter factors
must have the important property that as σi decreases, the corresponding fi tends to zero
in such a way that the contributions from the smaller σi are effectively filtered out. The
filter factors for Tikhonov regularization are given by

σ i2
fi = , (3.11)
(σ i
2
+ λ2 )

which demonstrates that the filtering effectively sets in for σi < λ.

The system for LS parameters yield a symmetrical matrix A. In this case (recall
that akr = ark in LS modification) the Tikhonovs problem Eq. (3.8) can be replaced with
the problem

( A + λIn ) x = b , (3.12)

and the filter factor fi for this system could be defined as (cf. Hansen 1998, p. 107)

σi
fi = . (3.13)
(σ i + λ )

53
 The filter factors play an important role in connection with regularization theory,
since the difference between the various regularization methods lies essentially in the
way that these fi are defined. No regularization method is complete without a method for
choosing the regularization parameter λ. Perhaps the most convenient graphical tool for
defining the Tikhonov regularization parameter is the so-called L-curve, which is a plot
for all valid parameters λ of the regularized solution norm ║xreg║ versus the residual
norm ║Axreg-b║. When the L-curve is plotted in log-log scale (for an ill-conditioned
matrix singular values span many orders of magnitude), it has very often a characteristic
L-shaped appearance (hence its name) with a distinct corner separating the vertical and
the horizontal parts of this curve. The exact definition of the L-curve corner point is not
a trivial problem and different methods may be applied (e.g. see Hansen 1998) for this.
For instance the L-curve’s corner could be defined as a point (with coordinates ζ and η)
on the curve

{ς ( λ ) ,η ( λ )} = {log Ax reg - b ,log x reg

2
} , (3.14)

which has the maximum curvature. The maximum curvature k(λ) can be defined from
classical differentiation formulas as

ς 'η ''− ς ''η '

k (λ ) = , (3.15)
{(ς ') 2
+ (η ' ) }
2 32

where differentiation is with respect to λ.

The L-curve displays clearly the compromise between minimization of the
residual and solution norms. It is impossible to construct any solution that corresponds
to a point below the L-curve; any regularized solution must lie on or above this curve.

5
10

4 8.8689e-016
10
SOLUTION NORM || x ||2

3
10

3.023e-014
2
10

1
10
5.7107e-010

0
10 1.6177e-008
1.2981e-0050.00036772
0.010416
-1
10 -10 -5 0
10 10 10
RESIDUAL NO RM || A x - b 2||

Fig. 3.1. Typical behaviour of the L-curve when solving the LS parameters.
Numbers on the L-curve indicate corresponding regularization parameter.

54
The optimal regularization parameter λ should correspond to a solution in which the
regularization (horizontal part in Fig. 3.1) and perturbation (vertical part in Fig. 3.1)
errors are balanced. In other words, by locating the corner of the L-curve one can
compute an approximation to the optimal regularization parameter. Obviously, a large
regularization parameter λ leads to over-smoothing, a small λ to under-smoothing of the
solution.

3.3.4 Discussion

The LS coefficients sn can be estimated either from the T-SVD or Tikhonov

regularization methods. For instance, in the latter method one has to impose some side
constraints (e.g. filter factors, or as it will be shown later - a priori knowledge of the
unknowns). The regularisation of the solution must therefore be seen in the framework
of biased estimator. This is the penalty we pay for getting a stable solution from the
regularization procedure. [Apparently, when employing the T-SVD method the
subsequent solutions are affected by this phenomenon as well. Namely, truncation of
the small SVD elements may be another source of the bias.] It can be shown that the
expectation of the estimator by Eq. (3.9) becomes

E {xˆ reg } = ( A T A + λ 2 I n ) A T E {b} = ( A T A + λ 2I n ) A T Ax ,

-1 -1
(3.16)

yielding that this estimator is biased towards zero (i.e. smoothed). This implies that the
signal variance of this biased estimator is also reduced. Nevertheless, when defining the
LS parameters, the regularization parameter λ is frequently in the order of 10-15. Hence,
due to the smallness of λ this bias remains insignificant.
According to Hansen (1998) the bias can be reduced by using a priori knowledge
of the parameters x. If an a priori estimate x* of the desired regularized solution is
available, then this information can be taken into account by including x* into the
regularization problem, which then can be rewritten as {cf. Eq. (3.8)}

{
min Ax - b 2 + λ 2 I n ( x - x *)
2 2
2 } . (3.17)

In this way, the regularized solution will be “biased” towards the a priori estimate x*.
As a matter of fact, the simple solution by Eq. (3.9) corresponds to the choice x*=0.
However, if an a priori x* is available then the regularized solution x can also be
elaborated by means of the SVD (cf. Hansen 1998, p. 74)

n
 uTb 
x reg = ∑  f i i + (1 − f i ) v Ti x*  v i . (3.18)
i =1  σi 

Here again, the numbers fi are filter factors for the particular regularization method and
they weight the contribution of b and parameters x*. For a large singular value σi the
filter factor fi {cf. Eq. (3.13)} is near one and thus the contribution from b dominates in
Eq. (3.18). In contrast, when fi decreases towards zero (i.e. in the case of small SVD
values) then the contribution from an a priori estimate x* becomes more important. As

55
the regularization parameter λ varies, the filter factors also vary and the balance
between the two contributions changes accordingly.
Some cautions should be emphasised. The regularized solution depends fully on
the quality of the a priori values, especially on the validity of the estimate x*. It is
desirable to use such x* which corresponds to E{x*} = x. However, it is almost
impossible to define, since the unknown x is unknown indeed. It should be mentioned,
however, when determining the LS modification parameters it is sufficient to limit the
computations to x* = 0. Recall that our ultimate goal is to compute the modified Stokes
function SL(ψ) and parameters bn, which appear to be rather insensitive to x*.

3.4 Numerical investigations

In recent computations of some regional geoid models at the Royal Institute of
Technology (e.g. Ellmann, 2001; Hunegnaw, 2001) the biased LS method, which gives
us a well-conditioned solution for the modification parameters sn, was favoured. First of
all, this choice could be explained by the occurrence of ill-conditioning for the unbiased
LS method (note that the optimum LS method was published in 2003). Nevertheless,
Sjöberg and Hunegnaw (2000) reported on solving the unbiased LS parameters for two
special cases when L = M = 20 and L = M = 180 and the terrestrial data were assumed
to be errorless.
The methods presented in this study, either the T-SVD or Tikhonov
regularization, are sufficient for solving the LS parameters from any combination of
initial conditions. The determination of the LS parameters starts from the SVD of the
matrix A consisting coefficients akr. Generally and independent on the chosen
combination of input values, the condition number for the biased LS method is
approximately from 102 to 103, when for the two other LS methods it ranges from 1015
to 1018. In other words the SVD elements of the unbiased and optimum methods range
over 18 orders of magnitude, which yield an extremely ill-conditioned system of linear
equations. In contrast, the system for the biased LS method is relatively well-
conditioned, which allows us to find the unknowns sn as the product of outright inverse
of the matrix A and the vector of hk coefficients. For the ill-conditioned systems the
straightforward inverse of A leads to parameters sn, which oscillate within the range of
± 105 and may thus be meaningless.
In order to find a solution from such a system one needs to detect the nature of the
ill-conditioned problem. This can be judged by inspecting the behaviour of the SVD
values. When solving the unbiased and optimum LS modification parameters with up-
to-date estimators of the degree variances cn, dcn and σ n2 , the most likely case to occur
is the well-determined gap in the singular values of the matrix A. Conversely, a gradual
decay of the SVD components appears when solving the biased LS modification
parameters. For the next step, one is encouraged to solve the unknown parameters sn by
a non-truncated SVD, i.e. by applying Eq. (3.6). Due to reasons explained earlier the
norm of parameters ||x|| may be rather large. In many cases, however, the corresponding
residual norm ||Ax-b|| does not exceed the magnitude 10-14…10-13. Thereafter the
parameters bn (see Table 3.1) are computed. An interesting feature (and important
criteria of the succeeded solution of sn) is, that for the same initial conditions (i.e. C(0),
M, L and ψ0) the parameters bn of the three LS methods (biased, unbiased, optimum)

56
are numerically very similar. Recall that the parameters sn of the biased LS are obtained
from a well-conditioned solution. If the outcome of the regularization deviates too much
(say more than 10-2) from those of the biased solution, sn = bn, this may be an indication
of regularization-related troubles. When such a case occurs, one need to continue
experiment with different truncation limits as explained in Section 3.3. At the same
time, the residual norm ||Ax-b|| of the final parameters to be used in geoid modelling is
recommended to be less than 10-12.
The LS parameters employed for the computations of Chapter 2 were obtained
from application of the T-SVD. Practically the same results were obtained by the
application of the Tikhonov regularization, Eq. (3.10). Therefore the user is
recommended to experiment with different regularization strategies, compare them and
take own decisions from these experiments.
The behaviour of resulting LS parameters bn is portrayed in Fig. 3.2. The
variations between different LS methods are insignificant {for all parameters
bn ≈ 2/(n - 1)}, which allowed us to denote the LS solutions with a single bold line in
Fig. 3.2. As a matter of fact, even though the variations of C(0) may have a dramatic
effect on the parameters sn, but at the same time the final parameters bn {and SL(ψ)}
remain similar (see also discussion in Section 2.4). Generally, C(0) = 0 - 9 mGal2
induced deviations of bn are of the same range (i.e. ~10-3) as are the deviations between
different modification methods.
Furthermore, the typical behaviour of SL(ψ) across the integration cap was
investigated. Importantly, any LS parameter set of our experiments “forces” the
modified Stokes’ function to zero (approximately) at the edge of the integration cap; see
Fig. 3.3. Even though the kernels of the unbiased and optimum LS estimators may take
unrealistic values outside the integration cap (Ågren and Sjöberg, in print 2004), this is
irrelevant to the geoid computation problem. Recall, that there is no integration beyond
the cap radius ψ0.

1200
1.8

1.6
1000

1.4
STOKES´ FUNCTION

800
1.2
n
PARAMETERS b

1
600
0.8

0.6 400

0.4
200
0.2

0
0
10 20 30 40 50 60 0 0.5 1 1.5 2
MODIFICATION DEGREE SPHERICAL DIST ANCE [°]

Fig. 3.2. Least squares modification parameters bn Fig. 3.3. Behaviour of the LS modified Stokes
(solid line; GGM01s model, modification degree function SL(ψ), {solid line; GGM01s, modification
L = M = 67, variance of terrestrial data C(0) = 1 degree L = M = 67, cap size ψ0 = 2°, variance of
mGal2, integration cap size ψ0 = 2°) versus the terrestrial data C(0) = 1 mGal2} versus the Stokes
Molodenskii’s truncation coefficients Qn (dashed function S(ψ) (dashed line). SL(ψ) values for
line). ψ < 0.1° are not shown. Note that the parameters sn
force the function SL(ψ) to almost zero at ψ0.

57
 Since the computations of Chapter 2 concerned also some deterministic methods,
it is appropriate to compare those with the LS methods. The parameters bn of the WG
and VK methods are numerically very similar to the LS parameters. Note, however, that
the selection of the modification parameters of the two deterministic methods were
constrained by some conditions {e.g. zero crossing of SL(ψ) at the integration cap
radius}. If one changes any of the initial modification criteria, more significant
discrepancies between different modification methods can be expected. The behaviour
of the parameters with some other modification criteria can be found in PAPER D, and
Ellmann (2001). The appearance of Molodenskii’s coefficients Qn {applied for the
simple modification method, Eq. (2.5)} is also visualized in Fig. 3.2. The simple
modification method utilizes the original (unmodified) Stokes function S(ψ), which is
marked in Fig. 3.3 by a dashed line.
The application of different approaches and regularization parameters may result
in a number of meaningful sets of the LS modification parameters sn. This prompts for
selecting the most useful sn set to be utilized for the geoid computations. It is reasonable
to give the preference to such a solution, which minimizes the expected global MSE mN2
furthermost. Thus, all calculated sn parameters should be tested for this criteria by
inserting them into Eq. (3.2).

3.5 Conclusions
This chapter of the Dissertation contributed to the important computational aspects of
the LS parameters sn. Aiming at minimising the error sources in geoid modelling, the
LS parameters sn are solved from a linear system of equations. Some numerical
difficulties may be encountered when solving this system. In particular, the parameters
of the unbiased and optimum LS modifications need to be defined from an extremely
ill-conditioned system. This calls for the application of some mathematical
regularization procedure. Two regularization methods (the T-SVD and Tikhonov
regularization), which are sufficient for obtaining a meaningful solution for any
combination of initial conditions, were reviewed. Both regularization methods provide
almost equivalent results. One is recommended to experiment with different
regularization strategies, compare them and draw conclusions from the test results.
Numerical procedures for assessing the quality of obtained parameters were
recommended. A typical outcome was presented and discussed. An important
characteristic of the successfully solved parameters is the expected global MSE, which
also gives a tentative idea about the expected (internal) accuracy of the geoid model.

58
Chapter 4
_____________________________________________________________________________________
Additive corrections to the geoid models

4.1 General
Geoid determination by Stokes’ formula requires that the disturbing potential be
harmonic above and on this boundary surface. In the classical approaches (i.e. the r-c-r
technique) the topographic and atmospheric masses above sea level are reduced inside
the geoid or completely removed. This implies a change of gravity corresponding to the
direct topographic (and atmospheric) effect. Furthermore, the application of Stokes’
formula requires that the gravity anomalies under the integral be given at sea level. In
modern context, the surface gravity anomaly is referred to ground level (Molodenskii
1945). Consequently, the gravity anomaly corrected for the direct effect must be
analytically continued to the sea level. This corresponds to the downward continuation
effect. After Stokes’ formula has been applied, the effect of restoring the topography
(and atmospheric masses) must be accounted for, yielding thus the indirect effect on the
geoid. Stokes’ formula for the gravimetric geoid calculations holds only on a spherical
reference, i.e. the input data must be given also on the sphere. The actual shape of the
geoid deviates from the global reference ellipsoid by at most 100 m. The systematic
error caused by neglecting the flattening of the ellipsoid is about 0.003 of the geoidal
height, reaching thus several decimetres.
The above effects become more apparent if Stokes’ original formula is corrected
as follows (cf. Sjöberg 2003c, Eqs. 1 and 3):

R
4πγ ∫∫
N= S (ψ )( ∆g + δ∆g *
dir + δ∆g dwc + δ∆g a )dσ + δ N I + δ N Ia + δ N e , (4.1)
σ

where ∆g is the surface gravity anomaly; δ∆g dir *

is the direct topographic effect on
gravity anomaly (referred to sea level); δ∆gdwc is the downward continuation effect;
δ∆ga is the direct atmospheric effect on gravity anomaly, δNI is the indirect topographic
effect on the geoid; δ N Ia is the indirect atmospheric effect on the geoid, and δNe is the
geoid correction for the spherical approximation of the Earth used in Stokes’ formula to
the surface of a reference ellipsoid. The other terms have already been determined in
Chapter 2 {see e.g. Eq. (2.4)}.
There are many ways how to account for these effects. In this study so-called
additive corrections are applied. Numerous publications (the references will be given
when appropriate) by Prof. L.E. Sjöberg demonstrate that all the elements in Eq. (4.1)

59
could be accounted on the geoid directly. The additive corrections are computed
separately and thereafter applied to the approximate geoidal height, which is obtained
from the Stokesian integration with the original surface anomaly. A comprehensive
theoretical summary is given in Sjöberg (2003c).
The expressions of the additive corrections can also be developed for the modified
Stokes formula. The emphasis of Chapter 4 is on the modification methods that employ
parameters bn = (sn+QnL). These methods are Wong and Gore (1969), Vaníček and
Kleusberg (1987) and the unbiased LS modification (Sjöberg 1991); see Table 2.1. The
geoid model for any of these methods can be expressed {cf. Sjöberg 2003c, Eq. 4a}

M
R
N = L
ˆ σ + c ∑ (QnL + sn ) ∆gˆ n + δ N comb
∫∫ S (ψ )∆gd
Mod
+ δ N atm
Mod
+ δ N dwc
Mod
+ δ N eMod , (4.2)
4πγ σ0 n =2

where δ N comb
Mod Mod
is the combined topographic effect, N atm is the total atmospheric effect,
δ N dwc
Mod
is the downward continuation effect and δ N eMod is the ellipsoidal correction.
Note that throughout the Dissertation it is assumed that the zero and first degree effects
of the corrections can be accounted for separately; see e.g., Sjöberg (2001a) and the
numerical studies in PAPERS E and F. It should be noted that for modification methods
employing bn = sn the expressions of additive corrections can also be developed. See
e.g. studies by Ellmann (2001, Ch. 3) and PAPER E.
Chapter 4 summarizes the numerical investigations of the additive corrections in
Eq. (4.2). There is no intention to duplicate the derivations and therefore only the final
expressions will be presented. Importantly, all the corrections can be expressed by
means of spherical harmonics, which is of computational advantage. In the target area
(see Section 2.3.3) the additive corrections are insignificant, as accounted on the geoidal
height they reach only a few mm. In order to demonstrate typical features of the additive
corrections we extend our computations to the region, where the magnitude of some
effects becomes more prominent due to significant topographic elevations. The new
study area extends from 52° to 71.5° northern latitudes and from 4.5° to 37° eastern
longitudes (see Fig. 4.1), including the whole of Fennoscandia and the Baltic Sea
region. The results of the additive corrections for different geographical regions, or even
worldwide, can be found in PAPERS E-G. The topographic and atmospheric effects are
considered in Sections 4.3 and 4.4, respectively. Section 4.5 deals with the downward
continuation effect, the ellipsoidal correction is considered in detail in Section 4.6. A
short conclusion follows.

4.1.1 Topographical elevations

The topographic, atmospheric and downward continuation effects can be expressed in

spherical harmonics of topographic heights. Referring to a surface point at latitude (ϕ)
and longitude (λ) the elevation H of the arbitrary power v, could be represented as

∞ n
H v (ϕ , λ ) = ∑ ∑H v
Y (ϕ , λ )
nm nm , (4.3)
n = 0 m =− n

60
 v
where H nm denotes the normalised spherical harmonic coefficient of degree n, order m,
and it can be determined by

1
= ∫∫σ H (ϕ , λ )Ynm (ϕ , λ )dσ ,
v v
H nm (4.4)
4π

where Ynm stands for the fully normalized spherical harmonic defined e.g. in Heiskanen
and Moritz (1967, p. 31). In the numerical computations the spherical harmonics with
resolution 30´ (ensuing the degree of expansion as of 360) were applied. The present
v v
study exploits the harmonics H nm , provided by Dr. H. Fan. For more details of H nm see
Fan (1998). The distribution of topographical masses, calculated by means of the first
power of elevation H {Eq. (4.3)} is visualised in Fig. 4.1.
The elevation extremes are 0 m at the shoreline and the highest peaks of the
mountain range along the Norwegian coast are above 2000 m. Due to the resolution 30´
the topography is certainly smoothed, for instance, the highest peak in Fennoscandia is
reduced to 1400 m. Nevertheless, in lower elevations the spherical harmonic
representation is nearer to the actual elevations.

30
10

1 00
0

70 200
5 00 2 00
30

68
0 200
0 0 750 50 2 00 1 00
10 7 5 100
66
7 50
0
20

64 30 20
0
0
50
LATITUDE [°]

0
750 10 20
10

0
1 00

0 0
0
500

62 1
00 0

30
0
12

30
10

0
200

2 00 30 20
60
30
0
75
5 00

100

30
30

1 00

58
2 00 1 00
0
20
10

30
30

1 00

56

0 10 0
10
200

54
20 0
30
100 10 0
52
5 10 15 20 25 30 35
LONGITU DE [°]

Fig. 4.1 Topographical elevations in the spherical harmonic

representation. Resolution 30´ × 30´ (55 × 55 km2). Target area is
enclosed by the rectangle. Unit is metre.

61
4.2 Combined topographic effect
This section is a brief summary of PAPER E. Even though that study focused on the
biased LS modification, the theoretical considerations and numerical results are similar
to those of the unbiased LS modification.
The most frequently applied “r-c-r” techniques of topographic masses suffer from
significant errors in the estimation of the direct and indirect effects. This is due to the
far-zone contribution (often neglected in practical computations) to the direct and
indirect topographic effects (Sjöberg 2000; Novak et al., 2001). The combined
topographic effect, as the sum of direct and indirect effects, can replace the traditional
approaches. Moreover, as proposed by Sjöberg (see e.g. 1994), the correction counting
for this effect could be made on the geoid directly.
The combined topographic effect is also independent of the method for the
topographic reduction (Sjöberg 2003c). By the summation of direct (δNdir) and indirect
(δNI) effects the reduction-related effect mostly diminishes. Note that the direct
topographic effect on the surface gravity anomaly includes also the secondary indirect
effect (Sjöberg 1997 and 2000). The combined topographic effect δ N comb Mod
can be
evaluated by means of surface spherical harmonics of elevation H squared:

2π Gρ ∞
δ Ncomb
Mod
= δ Ndir + δ NI =−
γ
∑H
n=2
2
n , (4.5)

where G is Newton’s gravitational constant and ρ is topographic mass density. The

topographic effect of the geoidal height is simply proportional to the elevation of the
computation point squared. Interestingly, the same expression of the topographic effect
is also valid for the original Stokes formula.
To illustrate the distribution of the combined topographic correction by Eq. (4.5)
we present the results of numerical studies in Fennoscandia and the Baltic Sea region;
see Fig. 4.2. As is expected the isolines of elevation and combined topographic effect
(cf. Figs. 4.1 and 4.2, respectively) are similarly shaped. The combined effect is always
negative. It decreases very rapidly in higher elevations, caused by the presence of H2 in
Eq. (4.5). The regional minimum in Norwegian mountains is below -15 cm. Recall that
the target area comprises mostly sea and topography below 100 m yielding that the
topographic effects are missing at all or are insignificant.
The 30´ resolution and the second power terms of H are sufficient for the accurate
determination of the combined effect in areas with average elevation lower than 4 km,
i.e. for most of the regions worldwide. In principle, the corrections could be derived to
arbitrary powers of topographic height. The expressions for the topographic effects,
including also higher power terms of elevation, have been published, e.g. in Sjöberg
(1977), Martinec (1998), Nahavandchi and Sjöberg (1998). The H3 terms of topographic
effects do not give significant improvement and can be safely neglected in lowland.
The combined effect is obviously also dependent on the density of topographic
masses. As emphasised in Sjöberg (1994), if the density of topographic masses varies
within 5%, the propagated geoid error could be as large as a few decimetres, globally.
However, for all areas below 1300 m this error in combined effect is within 1 cm. Thus,
the limited knowledge of topographical density is not a problem in the target area.

62
 -2-1
-10
70 -5-2 0 -5
-2

-2
-1
-5

-2
68 0 0
-5 0 -1
-1 0 -5 -5 -2
-1

-5
-20
66

20

-2
-5-

-1
-1

-1 0
64 -1

-2
-5-1 -2 -2

.
LATITUDE [°] -2
-1 0
62 0 -1
- 20 -5 0 0
-1 0 -50
5
-1 -2
-5

0 -1
-1
5 0
60 0 -5 -1 -2
-2-1

-1

-2
0

-2

-2
58 -1
-1
-5 -2 -1
-2
-1 -2
56 -5

-2
-5

-1
-5
54
-1 -2
-2

-1 -2
52
5 10 15 20 25 30 35
LONG ITU DE [°]

Fig. 4.2. Combined topographic effect. Resolution 30´ × 30´

(55 × 55 km2). Unit is millimetre.

According to Sjöberg (in print b) the lateral density effect of the geoidal height can also
be computed as a simple correction, which is proportional to the density variations and
the elevation squared. If the lateral density of topographic masses is adequately known,
then more accurate results can be obtained by Sjöberg (2000, Eq. 113; also in PAPER E
Eqs. 10 and 11). Due to the difficulty of obtaining reliable density information,
however, a constant density value is used in many traditional approaches and in the
present study. It should also be mentioned that the formulas of the topographic effect
are valid with slopes of topography less than 45°. However, the occurrence of this
undesired event with application of 30´ resolution spherical harmonics is believably
very rare, and definitely does not occur in our area of interest. An alternative way is to
correct the gravity anomalies for the attraction of topographical masses, and thereafter
utilise them in the modified Stokes formula.

4.3 Total atmospheric effect

This section is related to a study of Ellmann (2001, Ch. 3.2). Even though the focus of
that study was on the biased LS modification, the results are alike to those of the
unbiased LS modification.
The atmospheric masses above the geoid surface, like the topographic masses
discussed previously, must be removed, because of demands on the geodetic boundary
value problem. According to the recommendation of the International Association of
Geodesy (IAG), the effect of the atmospheric masses may be directly reduced on gravity
anomalies at the computation point (Moritz 1992). This method considers the direct

63
gravity effect (maximum at the sea level, +0.87 mGal) as the only atmosphere related
correction in Stokes’ formula. The indirect effect due to its small value (below 1 cm) is
usually neglected in this approach. However, this holds only if Stokes’ integral is
applied for the entire Earth. Conversely, if the area of integration is limited to a
spherical cap, the total mass of the atmosphere causes a significant bias (error) in the
zero degree term (Sjöberg 1998a). Subsequently, the application of the IAG approach in
a modified Stokes’ formula may easily cause a large bias. As a matter of fact many
datasets ignore the IAG recommendation. For instance, the IAG atmospheric correction
for the Estonian gravity survey is not considered at all (pers. comm. T. Oja, 2001). For
other datasets this information was not available. As emphasised in Sjöberg (1999),
instead of adding the IAG correction to gravity anomaly, the influence of the
atmospheric masses could be directly converted into a geoid model. Similarly to the
Mod
treatment of the topographic effects, the total atmospheric effect N atm can be evaluated
a
as the sum of the direct ( N dir ) and indirect ( N Ia ) atmospheric effects, and expressed by
means of surface spherical harmonics of elevation H (cf. Sjöberg and Nahavandchi,
1999)

2πρ0 R  ∞ 2 M ∞
n+2 
δ N atm
Mod
= N dir
a
+ N Ia = −  ∑
γ  n =2 n − 1
H n − ∑ ( sn + Qn ) n
L
H − ∑ QnL Hn  ,
2n + 1 
(4.6)
n =2 n = M +1

where ρ0 is the density of the atmosphere at the sea level. Note that Eq. (4.6) contains
parameters sn and QnL . Exactly the same modification coefficient set sn as applied in the
geoid determination has also to be utilized for the evaluation of the atmospheric effects.

4
4
2
70 0
-2 2

68
2
-2 0
4

66
2

2
-2

2
0

64
LATITUDE [°]

-2

62
-2

-8 -6 4
0
2

0
4

-4
0

60 2
-2

58
4
4

56
2

4
54
4

52
5 10 15 20 25 30 35
LONGITU DE [°]

Fig. 4.3. Atmospheric effect for the unbiased LS modification

method. Resolution 30´ × 30´ (55 × 55 km2). Unit is mm.

64
Thus, the total atmospheric effect is proportional to the elevation, but also dependent on
modification coefficients sn. The total atmospheric effect depends also on the upper
limit M. Since most of the power of the harmonics Hn is contained in low frequencies,
the atmospheric effect is strongly reduced when a reasonably large M and suitable
parameters {(sn + QnL) ≈ 2/(n-1) for the unbiased LS modification} are utilised. For
instance, even though M = 360 was adopted in conjunction with the modification
parameters bn = Qn, the resulting total atmospheric effect varied within 2.5 cm (i.e.
twice as much as for the unbiased LS modification) in the whole region. Hence,
topographic and atmospheric effects must be carefully considered if the simple
{equivalent to Vincent and Marsh (1974)} modification method is utilised. The total
atmospheric effect for the unbiased LS modification is portrayed in Fig. 4.3. There are
only insignificant numerical differences when the WG and VK parameters are used.

4.4 Correction for downward continuation of gravity anomaly

Due to a well-known boundary condition (see Section 2.1) the gravity anomaly
observed at the terrain surface must be continued downward to sea level. This leads to a
correction for downward continuation of the gravity anomaly.
Traditionally, the downward continuation and direct effects are treated together.
This relies on the rather rough approximation, that the surface gravity anomaly changes
linearly with elevation (see e.g. Moritz, 1980, p. 415). A more strict approach is to
reduce the surface gravity anomaly by the direct topographic effect and then continue
the reduced gravity anomaly downward to sea level by the inversion of Poisson’s
integral formula (see e.g., Bjerhammar, 1962, Heiskanen and Moritz, 1967, p. 317).
This approach has lately been studied by Martinec and Vaníček (1994a), Martinec
(1998), Hunegnaw (2001). If the direct effect accounts for all topographic masses,
Poisson’s integral should be solvable. In practice, however, the estimation of the direct
effect is based on the approximate knowledge of topographic mass. This leaves some
residual attraction, which makes the downward continuation problem ill-conditioned.
Sjöberg (1998b, 2001b) presented approaches in which the downward
continuation problem is solved without inversion of Poisson’s integral and separately
from other topographic effects. A new estimator of the downward continuation effect in
the modified Stokes formula can be expressed (cf. Sjöberg 2003a)

∞
δ N dwc
Mod
= δ N dwc − c ∑{QnL + sn* } × {∆g n − ( ∆g s ) n } (4.7)
n =2

where δNdwc is the downward continuation effect on the geoid corresponding to the
original Stokes formula {the complete expressions in spherical harmonics of the
topographic height H can be found in Sjöberg (2003a, Eqs. 30a-d)}, ∆gn and (∆gs)n are
the Laplace harmonics at sea level and for the of the surface gravity anomaly, given by
the formulas Eq. (2.7) and (ibid. Eqs. 29a-b), respectively. The downward continuation
effect is presented here only for the sake of the completeness. In the target area the
downward continuation effect is less than 1 mm and therefore it was not considered for
the numerical analysis.

65
4.5 Ellipsoidal corrections
PAPERS F and G are related to the numerical studies of the ellipsoidal correction in the
original and modified Stokes’ formula. We begin with the presentation of the ellipsoidal
correction for the original Stokes formula. Thereafter these expressions will be adapted
to the modified Stokes formula (Sjöberg 2003b; and, in print a).
Rigorously, geoid determination by Stokes’ formula holds only on a spherical
boundary, yielding that the “ideal” gravity anomaly ∆g0 is consistent with the well-
known boundary condition (Heiskanen and Moritz 1967, p. 88)

∂T 2
∆g 0 = − − T , (4.8)
∂r r

where r is geocentric radius and T is the disturbing potential. Eq. (4.8) is often assumed
to be referred to the mean Earth sphere (MES, with radius R). In practise, however, ∆g
is observed on the Earth’s mean sea level (MSL), which can be approximated by an
oblate ellipsoid. Note that the Stokesian integration utilises commonly those “practical”
gravity anomalies. The boundary condition for ∆g could be expressed as (Heiskanen
and Moritz 1967):

∂T ∂γ T ∂T 2
∆g = − + =− − T + δ ge , (4.9)
∂h ∂h γ ∂r r

where h is the geodetic height and δge is the ellipsoidal correction for the gravity
anomaly. Gravity is measured along the direction of the plumb-line, which differs from
the radial direction by order e2 (the squared eccentricity of the Earth’s ellipsoid). Also,
the discrepancy between MSL (where gravity is observed) and MES (where gravity is
needed in Stokes’ formula) is of order e2. Therefore the last term of Eq. (4.9) is often
expressed to order e2 {see also Eq. (4.13)}. These problems could be expressed as
follows (cf. Cruz 1986)

 ∂∆g 0  2 ∂ T
∆g 0 − ∆g = ( R − re )   + e  2 − 3cos θ − sin θ cosθ
2
 (4.10)
 ∂r  r = R  ∂θ  a

where re and θ are geocentric radius and co-latitude of the computation point, and a is
equatorial radius of the Earth. The first term of the right side reflects the
upward/downward continuation of ∆g from MSL to MES, whereas the last term
accounts for the discrepancies between the radial direction and the plumb-line. The
application of these corrections is often neglected in practical computations. This yields
an error in geoid determination of the order of the Earth’s geometrical flattening
(approx. 1/300) times the geoidal height, reaching thus several decimetres, globally.
Hence, for an accurate geoid model it is important to estimate the ellipsoidal correction.
Different authors have studied the ellipsoidal correction for the original Stokes
formula, see e.g. Molodenskii et al. (1962), Moritz (1980), Martinec and Grafarend
(1997), Fei and Sideris (2000), Heck and Seitz (2003), and references therein. A new
integral solution for the ellipsoidal correction was published by Sjöberg (2003b). For

66
the derivation of the ellipsoidal correction different boundary conditions are posed,
where the disturbing potential is commonly expressed as a series expansion of
eccentricity e2. The solutions differ in various respects, preventing their direct analytical
comparison. Accordingly, PAPER F aimed to contribute with the numerical comparison
of the three ellipsoidal corrections for the original Stokes formula - Martinec and
Grafarend (1997), Fei and Sideris (2000), and Sjöberg (2003b). The three methods
provide the correction as an integral over the full solid angle with various kernel
functions. The data sources ∆g and N0 (approximate geoidal height) are integral
arguments in the Martinec-Grafarend and Fei-Sideris methods, respectively, whereas
Sjöberg’s approach exploits both types of data. Furthermore, Sjöberg’s ellipsoidal
correction can be developed into a spherical harmonic series as

GM ∞ n Ynm
δ Ne = k´N 0 + e2 ∑ ∑ {[3 − ( n + 2) Fnm  Cnm − ( n + 1) GnmCn−2,m − ( n + 7) EnmCn+2,m} ,(4.11)
2aγ n=2 m=−n n − 1

where GM is the product of the Earth’s mass and the gravitational constant, n and m are
the degree and order of harmonic series, the coefficients Cnm are the fully normalised
harmonic coefficients of the disturbing potential, Ynm is the fully normalized spherical
harmonics (cf. Heiskanen and Moritz 1967, p. 31). The coefficients Enm, Fnm, Gnm can be
derived through relations presented in Martinec (1998; also in Sjöberg 2003b, eqs. A6-
8). The scale factor k´= (a-R)/R is needed for integration over the bounding sphere
instead of MES. Importantly, the solution by Eq. (4.11) is convenient for practical
computations, since the singularity problem is eliminated and tedious global integration
is avoided. For some other ellipsoidal corrections in the spherical harmonic
representation see Moritz (1980, Ch. 39), and Heck and Seitz (2003). It is notable that if
the term k´N0 in Eq. (4.11) is excluded, the expressions for degrees n ≥ 2 of the Sjöberg
(2003b) and Heck-Seitz (2003) methods are practically equivalent.
A remarkable finding is that the range and global distribution of the Fei-Sideris
(2000) and Sjöberg (2003b) ellipsoidal corrections are very similar; see PAPER F
(Table 1 and Fig. 2). This result is surprising in the view of that their derivational
approaches and resulting expressions are considerably different, yielding, however, the
numerical discrepancies of only about ±1 cm, globally. This suggests that the two
methods practically coincide. Generally, in the area where the gravity anomalies are
above MES (i.e. between latitudes ± 45°) the sign of the ellipsoidal correction coincides
with the sign of the geoidal height. In regions where the gravity anomalies are below
MES (i.e. in sub-polar latitudes) the ellipsoidal corrections generally have opposite sign
to the respective N. Intuitively, the absolute values of geoidal heights increase when ∆g
need to be continued downward to MES and decay when ∆g is continued upwards.
Interestingly, a strong correlation between the Martinec-Grafarend ellipsoidal
correction and the approximate geoid model is detected; see PAPER F (Fig. 1). The
local extremes of the geoid and ellipsoidal corrections have the same sign in the same
regions, thus the corrections practically mirror the geoid model. Roughly, the magnitude
of ellipsoidal correction comprises 0.5-1% of N0. As a consequence, the Martinec-
Grafarend (1997) method provides almost twice as large correction values, compared to
the other methods {see also the results in Huang et al. (2003) and Heck and Seitz
(2003)}. The reasons for large numerical discrepancies between Martinec-Grafarend

67
(1997) and other studied methods are yet unknown, thus calling for further analytical
studies of the derivations.
A detailed comparison between the approaches by Molodenskii et al. (1962),
Moritz (1980), Martinec and Grafarend (1997) and Fei and Sideris (2000) can be found
in Huang et al. (2003). They used EGM96 for generating the synthetic data. The
outcomes of our study (PAPER F: Table 1 and Figs. 1-2) agree well with the results in
Huang et al. (2003, see Table 1 and Figs. 2c-d). Huang et al. (2003) concluded that
Martinec-Grafarend’s solution demonstrates the best conformity to the synthetic geoid,
while the other three methods reveal deficiencies in the first-degree spherical
harmonics. Note, that Sjöberg’s approach was not included in their study.
In contrast to the original formula, the ellipsoidal correction for the modified
Stokes formula has received less attention in the geodetic literature and it is often
neglected in the practical computations. The ellipsoidal correction for the modified
Stokes formula is derived in Sjöberg (in print b) and it can be expressed as a series of
spherical harmonics:

R ∞
 2  a − R ˆ a 
δ N eMod =
2γ
∑  n − 1 − s
n =2
n − QnL  
 R
∆g n + δ g ne 
R 
, (4.12)

where δge is the ellipsoidal correction of the anomaly {cf. Eq. (4.9)-(4.10)}, which can
be presented by (ibid., Eq. 29e; also in PAPER F, Eq. 14)

∞ n
GM
δ g e = e2 ∑ ∑ Y {[3 − (n + 2) F ] C
nm nm nm −( n + 1)GnmCn −2,m − ( n + 7) EnmCn +2,m } . (4.13)
2a 2 n = 2 m =− n

Note that the component δge of the ellipsoidal correction is dependent on the maximum
harmonic degree of a GGM expansion. In practical computations δge and ∆gn were
estimated from the full expansion of the EGM96. For the magnitude and distribution of
δge ; see PAPER G (Fig. 5). In particular, the numerical verification with high-
resolution data reveals that the truncation of the Cnm series at n = 360 may cause a 10-
15% omission error, which absorbs into final ellipsoidal correction of the geoidal height
(PAPER G). This omission error is insignificant for the modified Stokes formula.
Generally, the largest corrections δ N eMod are either related to the equatorial areas
with large geoid anomalies, or high gravity anomaly regions in sub-polar latitudes. This
can be judged from two approximate formulas for estimation of the ellipsoidal
correction, which are derived in PAPER G (Eqs. 20 and 25). Even though the spherical
harmonic series by Eq. (4.12) (and corresponding integral formula, see PAPER G, Eq.
15) provides the most accurate results, the approximate formulas provide reasonable
approximation for the expected range of the ellipsoidal correction
As shown in PAPERS F and G the absolute range of the ellipsoidal correction in
the LS modification of Stokes’ formula does not exceed 1-2 cm, globally. This seems to
prove the common assumption that the ellipsoidal correction could be neglected in the
modified Stokes formula, if cm accuracy is considered as sufficient. This does not
mean, however, that the ellipsoidal correction can be abandoned for any modification
method. Note that spherical harmonic representation of δ N eMod {i.e. Eqs. (4.12)} are
dependent on the modification coefficients sn, and are also related to the integration cap

68
size (via term QnL ). As is expected, the larger the integration cap, the greater the
ellipsoidal correction. Various sets of modification coefficients sn were tested.
Apparently, a precondition for a little ellipsoidal correction is directly related to the
modification parameters bn = (sn + QnL), see Eq. (4.12). Obviously, when bn ≈ 2/(n-1),
then the resulting ellipsoidal correction is insignificant. For the WG, VK and LS
modification methods (Sjöberg 1984b, 1991 and 2003d) this condition is usually
satisfied. If the modification parameters bn differ notably from 2/(n-1), this effect should
be carefully assessed, especially in the regions with large gravity anomalies.
Recall that Chapter 2 considered also the simple modification case with sn = 0.
The first term of Eq. (4.12) reduces now to [2/(n-1) – Qn]. Recall that Molodenskii’s
truncation coefficients Qn depend solely on the integration cap size. The larger ψ0 the
larger are discrepancies between Qn and 2/(n - 1), and thus the greater the ellipsoidal
correction. The first term of Eq. (4.12) increases, yielding greater magnitude of the
ellipsoidal correction. This may result in a correction of several centimetres even for a
small truncation cap. For instance, the ellipsoidal effect in the study region varied
within 2.5 cm for the modification parameters bn = Qn (M = 360, integration cap
ψ0 = 2º). The conclusion follows that the ellipsoidal correction cannot be ignored for the
simple modification method in conjunction with a large integration domain. An
approximate expression for estimating the ellipsoidal correction for the cases with
bn = Qn can be found in Sjöberg (2003e, Eq. 39).

-4 2
0
-2

-4
70
2
-2 0
68 2

0 -2
2 4
0

0
66
2
6
2 0
64
LATITUDE [°]

0
7

62
-2

2
6
4

60
0

2
6

58 7
4

4
56
4
4 2
54 4

4 4
52
5 10 15 20 25 30 35
LONGITU DE [°]

Fig. 4.4. Distribution of the ellipsoidal correction for the unbiased

LS modification method. Resolution 30´ × 30´ (55 × 55 km2). Unit
is mm.

69
 We conclude thus that it is more secure to use the methods which apply the
modified Stokes function SL(ψ) than the methods utilising S(ψ) in the integral kernel. In
the former case the corrections are very small, and when ignored, no notable damage is
caused. The ellipsoidal corrections for the unbiased LS modification are portrayed in
Fig. 4.4. Note that the largest corrections are related to the sites with strongest
(negative) gravity anomalies (see Fig. 2.3). Note the opposite sign of the correction at
these sites.

4.6 Conclusions
The corrections which have to be applied in a geoid determination process were
reviewed in this part of the Dissertation. We propose that the classical approaches for
estimating various effects in Stokes’ formula are preferable replaced by the additive
corrections, which hold for the original Stokes formula as well as for the modified
Stokes formula.
The additive corrections are applied directly to the (approximate) geoidal height,
which is obtained from the Stokesian integration with surface gravity anomalies.
Compared to the classical approaches the new formulas are much simpler and easy to
implement in practical computations. Importantly, all additive corrections can be
expressed by means of spherical harmonics, which is of computational advantage. For
instance, the combined topographic effect of the geoidal height is simply proportional to
the elevation of the computation point squared. Another advantage of the combined
topographic effect is that it is completely independent of the choice of method of
reduction of the topography. More rugged terrain may require more sophisticated
approaches and high-resolution data. In such a regions probably it is reasonable to
account for topographic effects by using more rigorous formulas, or correct the
anomalies for the attraction of the topographic masses.
The expressions of the additive corrections must be consistent with the
modification method of Stokes’ formula. This yields that the additive corrections may
become dependent on modification coefficients sn.
Within the target area the total sum of all the additive corrections for the unbiased
LS modification of Stokes’ formula does not exceed 1 cm. The same magnitude of the
additive corrections can be expected for the Wong-Gore, Vaníček-Kleusberg, biased
and optimum LS modification methods. The geoid model BALTgeoid-04 (computed by
the unbiased LS modification, see Chapter 2) was supplemented with the additive
corrections. Subsequently, Chapter 4 completes the geoid modelling procedures.

70
Chapter 5
______________________________________________________________________
Conclusions, discussion and future investigations

One of the main objectives of physical geodesy is the determination of the geoid. The
geoid supports many professional, economic and scientific activities and functions,
ranging from navigation, mapping and surveying to the construction and exploitation of
nationwide infrastructure objects; from detecting the variations of the ocean currents to
the interpretation of seismic disturbances. A refined regional geoid model, in particular,
enables the user in many cases to replace the traditional height determination techniques
by faster and more cost-effective GPS measurements.
The Dissertation is devoted to the determination of a high-resolution geoid model
for the three Baltic countries. We impose the geoid determination problem by applying
the modified Stokes formula with additive corrections. This task is more explicitly
expressed by Eq. (4.2). The Dissertation tackles the aspects, both theoretical and
practical, of determining of all the elements of this formula. The results of the
investigations were presented in Chapters 2-4 as follows: (Ch. 2) Modified Stokes’
formula with application to the Baltic geoid, (Ch. 3) Determination of the least squares
modification parameters and, (Ch. 4) Additive corrections to the geoid models. More
detailed conclusions were drawn at the end of each chapter. The focus here is given only
to the most important results of the Dissertation.
In order to compute a high-resolution geoid model six different (three
deterministic and three stochastic) modification methods of Stokes’ formula were
applied. The deterministic methods are Wong and Gore (1969), Vincent and Marsh
(1974), and Vaníček and Kleusberg (1987). The stochastic methods are biased, unbiased
and optimum least squares modifications by Sjöberg (1984b, 1991, 2003d). The
deterministic methods utilise originally the residual anomaly in the integral. For the
sake of the comparison, those deterministic methods are expressed such that the original
surface gravity anomaly, instead of the residual anomaly, is exploited in Stokes’
integral. Five methods utilise the modified Stokes function SL(ψ) as an integral kernel,
whereas the original (non-modified) Stokes function is used by one modification
method (Vincent and Marsh 1974).
Note that all terms of modified Stokes’ formula and the additive corrections may
include the modification coefficients sn. The definition of the modification and geoid
modelling parameters is a rather delicate task and requires a thorough consideration of
the characteristics of the target area and available datasets. Depending on the local
gravity data quality, the chosen radius of integration and the characteristics of used
GGM, the LS modification parameters sn vary.

71
 The least squares modification methods attempt to minimise all relevant error
sources in geoid modelling. For this task a set of LS modification parameters is
determined from a system of linear equations, aiming at minimizing the expected global
mean square error of the geoid estimator. Regrettably, for certain LS methods some
numerical instability may be encountered when practically computing the modification
parameters. In order to obtain a meaningful, stable solution, the tools of mathematical
regularization need to be applied. Chapter 3 reviewed two methods, which are sufficient
to cover the different cases when defining the LS modification parameters sn. Both
methods utilise the singular value decomposition and provide rather similar numerical
results.
The numerical procedures for assessing the quality of obtained parameters were
discussed. A typical feature of the LS parameters sn is that the integration kernel
becomes almost zero at the pre-selected integration cap radius. This is also an important
feature for the deterministic methods, since the truncation error is minimum (de Witte
1966, Meissl 1971) when this condition is fulfilled. Accordingly, the parameters of the
VK and WG methods are selected such that that SL(ψ) is zero at the integration radius of
this study, i.e. at ψ0 = 2º.
The expected global MSE is a useful quantity to analyse the contribution of
different error sources in the geoid estimation process. This study employs the new
GRACE model GGM01s and EGM96 in the practical computations. Unlike the past
geopotential models, GGM01s is highly accurate and homogeneous. However, in order
to achieve a high resolution for the regional geoid models the terrestrial data cannot be
abandoned even in the GOCE era. The recent technological and theoretical advances
have created preconditions to achieve 1-cm accuracy for a high-resolution regional
geoid model. Disappointingly, the deficiencies in historical terrestrial gravity data
corrupt this objective remarkably. The shortages of the gravity data cannot be
completely removed. Nevertheless, similarly to some earlier studies (e.g. Vaníček and
Featherstone 1998; Ågren and Sjöberg, in print) we conclude that the methods utilising
SL(ψ) mitigate the terrestrial data errors more efficiently than the modification methods
utilising unmodified Stokes’ function S(ψ) in the integral. Follow-up studies should be
devoted for reducing the terrestrial data errors even further. Concerning the LS
modifications, this problem necessitates the development of such stochastic models,
which will consider more precisely the non-homogeneity of the terrestrial data, both in
accuracy and coverage. At the same time, the capacities of existing stochastic models
(such as listed in Appendix B of Chapter 2) are most likely not exhausted, yet.
Intuitively, the improved stochastic models could enhance the outcome of the LS
modification methods even further (also emphasised by Nahavandchi 1998). The results
of the present study reveal that the LS modification methods allow the most efficient
reduction of the expected global MSE. Even though the global MSE may provide an
idea about the expected internal accuracy of different modification methods, these
theoretical estimators need to be verified by practical computations and some
complementary datasets.
The six modification methods were applied for computing the geoid models for
the Baltic countries. Although the modification parameters are defined differently, the
numerical discrepancies between various geoid models exploiting SL(ψ) and GGM01s
remain within ± 9 cm. At the same time the differences between the GGM01s and
EGM96 based deterministic geoid models range in the target area from –6 to +17 cm.

72
This implies that to-date (2004) the deviations between the recent geopotential models
are more crucial than the numerical outcome of the tested modification methods.
As expected, the discrepancies between the upcoming geopotential models will be
reasonably small, which prompts for more careful selection of geoid modelling methods
and parameters. Preferably, several different modification methods should be
simultaneously put on the trial, following the methodology outlined in this dissertation.
Three sets (one from each country) of GPS-levelling points were used for an
independent evaluation of the computed geoid models. The main conclusions drawn
from this comparison are:
(1) The GGM01s based regional geoid models agree with the national height systems
better than the EGM96 based geoid models.
(2) The accuracies of the tested modification methods are at least of the same level as is
the accuracy of the control points.
(3) The RMS of the post-fitted residuals is smallest in Estonia, whereas more scattered
residuals are observed in Latvia.
(4) The LS modification methods perform best overall. However, the deterministic
methods are formally better in Lithuania.

The best RMS of the GPS-levelling post-fit residuals were as follows: 5.3 cm for
the joint Baltic geoid model and 2.8, 5.6 and 4.2 cm for Estonia, Latvia and Lithuania,
respectively. Undoubtedly, the comparison of the geoid modelling outcomes with the
GPS-levelling data is an important indicator of applicability of the regional geoid
models for many practical tasks. It cannot be excluded, however, that some systematic
biases between geodetic datasets (including all the three components - GPS, levelling
and gravity) remained in the comparison. Consequently, a careful and versatile revision
of all available geodetic data is desired. This can only be done within the frame of
international cooperation, desirably resulting in modernization of the gravity databases
and vertical networks of the Baltic countries.
Generally, the post-fit residuals from the LS modification are slightly smaller (up
to 1 cm) than the respective values of deterministic methods. This implies that the
efforts put into minimization of the global MSE have been successful. In general, all the
modification methods combining the modified Stokes function SL(ψ) and an
intermediate expansion of the geopotential models, provide a reasonable accuracy even
for rather small cap sizes. Conversely, to secure the cm accuracy in the geoid
modelling by the simple modification method, the employment of a high degree
expansion of a GGM and data from a large integration cap are necessary. This yields
that the “satellite-only” models in conjunction with unmodified S(ψ) do not qualify for
computations of high-resolution regional geoid models.
The geoid model computed by the unbiased LS modification provides the “best”
post-fit statistics and it is thus preferred as the final representation of the joint Baltic
geoid. The modification parameters of this model are calculated from the following
initial conditions: (1) geopotential model GGM01s, (2) upper limit of the GGM01s and
the modification degree of Stokes’ function are set to 67, (3) terrestrial anomaly
variance is assumed to be 1 mGal2 with a correlation length of 0.1°, and (4) the
integration cap size 2° is used. This approximate geoid model is supplemented by
separately computed additive corrections (the combined topographic and atmospheric
effects and the ellipsoidal corrections), which completes the geoid modelling
procedures. The new geoid model for the Baltic countries is named BALTgeoid-04.

73
 It should be noted that some of the additive corrections are dependent on the
modification coefficients sn and the selected upper limit of the GGM. We conclude that
the modification methods utilising the kernel SL(ψ) are rather insensitive to the
ellipsoidal and atmospheric effects. The only exception is the combined topographical
effect, which is either not or insignificantly dependent on the modification parameters.
Since the application of the combined topographic effect for mountainous regions was
not compared against other approaches, usability of the combined topographic effect in
such of areas should be additionally verified by future studies. The rigorous formulas in
Sjöberg (2000) or topographically corrected gravity anomalies are possible candidates
for such a study.
It is demonstrated that the application of the LS modification methods is
advantageous for the Baltic countries. However, the suitability of LS methods should be
tested also for other regions. To master rather complicated expressions, which are
attributed to the determination of the LS modification parameters, one need to elaborate
special computation software. Due this reason, obviously, the LS methods have been of
limited use. This problem could be overcome, however. One of the secondary goals of
this dissertation was to develop and explain the computing routines, which are useful
for solving the LS parameters. The computation programs will be made available to the
public. It is our humble hope that availability of such software increases also the
eagerness of geodesists to apply the LS modification methods for their own data in
different regions. Hopefully, the suggestions and restrictions outlined in this dissertation
are useful for this assignment.

74
______________________________________________________________________
References

This list comprises 105 bibliographical references. PAPERS A-G are listed separately,
see page 82.

Bilker M, Mäkinen J, Poutanen M, Forsberg R, Scherneck H-G, Vermeer M (2001) A proposal to use the
Baltic Sea area as a test field for the gravimetric satellites. IAG Scientific Assembly, Budapest, 2001.

Bjerhammar A (1962) Gravity reduction to a spherical surface. Report of the Royal Institute of
Technology, Geodesy Division, Stockholm.

Bjerhammar A (1973) Theory of errors and generalized matrix inverses. Elsevier Scientific Publishing
Company, Amsterdam, London, New York.

Bruns H (1878) Die Figur der Erde. Publikation des Kniglichen Preussischen Geodet. Institutes, Berlin.

Cruz JY (1986) Ellipsoidal corrections to potential coefficients obtained from gravity anomaly data on the
ellipsoid. Rep. 371, Dept. Of Geodetic Science and Surveying. The Ohio State University, Columbus.

de Witte L (1966) Truncation errors in the Stokes and Vening Meinesz’ formulae for different order
spherical harmonic gravity terms. Aerospace rep TR-669 (SG 230-35)-5.

Ellmann A (2001) Least squares modification of Stokes formula with applications to the Estonian geoid.
Royal Institute of Technology, Division of Geodesy Report No. 1056 (Licentiate Thesis), Stockholm.

Evans JD, Featherstone WE (2000) Improved convergence rates for the truncation error in geoid
determination, J Geod, 74: 239-248.

Fan H (1998) On an Earth ellipsoid best-fitted to the Earth’s surface. J Geod 72: 511-515.

Featherstone WE (2003) Software for computing five existing types of deterministically modified
integration kernel for gravimetric geoid determination, Computers and Geosciences, 29(2): 183-193.

Featherstone WE, Kirby JF, Kearsley AHW, Gilliland JR, Johnston GM, Steed J, Forsberg R, Sideris MG
(2001) The AUSGeoid98 model of Australia: data treatment, computations and comparisons with GPS-
levelling data. J Geod, 75: 313-330.

Featherstone WE, Evans JD, Olliver JG (1998) A Meissl-modified Vaníček and Kleusberg kernel to
reduce the truncation error in gravimetric geoid determination. J Geod, 72, 154-160.

Fei ZL, Sideris MG (2000) A new method for computing the ellipsoidal correction for Stokes’ formula. J
Geod, 74: 223-231 and 671.

Forsberg R (1993) Modelling the fine structure of the geoid: methods, data requirements and some
results. Surv Geophys, 14:403-418.

75
Forsberg R (1998) Geoid tailoring to GPS – with example of a 1-cm geoid of Denmark. In Vermeer M,
Ádám J (eds): Second continental Workshop on the geoid in Europe. Budapest, 10-14 March, 1998.
Reports of the Finnish Geodetic Institute 98:4, pp. 191-197.

Forsberg R (2001) Development of a Nordic cm-geoid with basics of geoid determination. In B. G.

Harsson (ed.): Nordic Geodesy towards the 21st century. Lecture notes for Autumn School, Nordic
Geodetic Commission, pp. 67-88, Statens kartverk, Geodetic Publication 2001:1.

Forsberg R, Sideris MG (1993) Geoid computation by the multi-band spherical FFT approach. Manusc
Geod, 18: 82-90.

Forsberg R, Kaminskis J, Solheim D (1997) Geoid for the Nordic and Baltic Region from Gravimetry and
Satellite Altimetry. In J. Segawa, H. Fujimoto and S. Okubo (ed.): Gravity, Geoid and Marine geodesy.
IAG Symposium Series volume 117, pp. 540-547, Springer Verlag.

Forsberg R, Strykowski G, Solheim D, Jürgenson H (in print) NKG-2003 geoid model of the Nordic and
Baltic area. IAG Symp. Series. IUGG 2003, Sapporo.

Haagmans R, de Min E, van Gelderen M (1993) Fast evaluation of convolution integrals on the sphere
using 1D FFT and a comparison with existing methods for Stokes’ integral. Manusc Geod, 18: 227-241.

Hagiwara Y (1976) A new formula for evaluating the truncation error coefficient. Bull Geod, 50:131-135.

Hansen PC (1998) Rank-deficient and Discrete Ill-Posed Problems: numerical aspects of linear inversion.
SIAM, Philadelphia.

Heck B, Grüninger W (1987) Modification of Stokes’s integral formula by combining two classical
approaches. Proceedings of the XIX General Assembly of th eInternational Union of Geodesy and
Geophysics, Vol. 2. Vancouver, Canada, pp. 309-337.

Heck B, Seitz K (2003) Solutions of the linearized geodetic boundary value problem for an ellipsoidal
boundary to order e3. J Geod, 77: 182-192.

Heiskanen WA, Moritz H (1967) Physical geodesy. W.H. Freeman and Company. San Francisco.

Huang J, Véronneau M, Pagiatakis SD (2003) On the ellipsoidal correction to the spherical Stokes
solution of the gravimetric geoid. J Geod, 77: 171-181.

Hunegnaw A (2001) Geoid determination over Ethiopia with Emphasis on Downward continuation of
gravity anomalies. Royal Institute of Technology, Division of Geodesy Report No. 1057 (Doctoral
Dissertation), Stockholm.

Jekeli C (1981) Modifying Stokes’s function to reduce the error of geoid undulation computations. J
Geophys Res 86(B8): 6985-6990.

Jiang Z, Duquenne H (1996) On the combined adjustment of gravimetrically determined geoid and GPS
levelling stations. J Geod, 70: 505-514.

Jürgenson H (2001) Gravity measurements and calculations of the geoid model for Estonia. In V.
Ingemann and F. Madsen (eds.): Proceedings from Seminar of the Danish & Baltic sector Programme,
Riga March 2-3, 1999, pp. 94-110, Copenhagen.

Jürgenson H (2003) Determination of Estonian Precision Geoid PhD Thesis, Estonian Agricultural
University, (in Estonian), Tartu.

Kaminskis J, Forsberg R (1997) A new detailed geoid for Latvia. In J. Segawa, H. Fujimoto and S. Okubo
(eds.): Gravity, Geoid and Marine geodesy. IAG Symposium Series volume 117, pp. 621-628, Springer.

76
Kearsely AHW (1988) Tests on the recovering of precise geoid height differences from gravimetry. J
Geophys Res, 93(B6): 6559-6570.

Kotsakis C, Sideris MG (1999) On the adjustment of combined GPS/levelling/geoid networks. J Geod,

73: 412-421.

Lemoine FG, Kenyon SC, Factor JK, Trimmer RG, Pavlis NK, Chinn DS, Cox CM, Klosko SM, Luthcke
SB, Torrence MH, Wang YM, Williamson RG, Pavlis EG, Rapp RH, Olson TR (1998) The development
of the joint NASA GSFC and NIMA Geopotential Model EGM96. NASA/TP-1998-206861.

Martinec Z (1998) Boundary-value problems for gravimetric determination of a precise geoid. Springer-
Verlag, Berlin, Heidelberg, New-York.

Martinec Z, Vaníček (1994) Direct topographical effect of Helmert’s condensation for a spherical geoid.
Manuscr Geod, 19: 257-268.

Martinec Z, Grafarend EW (1997) Solution to the Stokes’ boundary value problem on an ellipsoid of
revolution. Studia geophysica et geodaetica, 41:103-129.

Meissl P (1971) Preparation for the numerical evaluation of second order Molodenskii- type formulas.
Dept. of Geodetic Science, OSU Report No. 163, Columbus, Ohio.

Molodenskii MS (1945) The principal problems of geodetic gravimetry. Trudy TsNIIGAiK, 131,
Moscow (in Russian).

Molodenskii MS, Eremeev VF, Yurkina MI (1962) Methods for study of the external gravitational field
and figure of the Earth. (transl. From Russian 1960). Israel program for Scientific translations, Jerusalem.

Moritz H (1966) Linear solutions to the geodetic boundary-value problems. Rep. 79, Dept Geod Sci, Ohio
State Univ, Columbus.

Moritz H (1980) Advanced physical geodesy. Herbert Wichmann, Karlsruhe/Abacus Press, Tunbridge
Wells.

Moritz, H (1992) Geodetic Reference System 1980, Bull Geod, 66: 187-192.

Nahavandchi H (1998) Precise gravimetric-GPS geoid determination with improved topographic

corrections applied over Sweden. Doctoral Dissertation, Division of Geodesy Report No.1050. Royal
Institute of Technology, Department of Geodesy and Photogrammetry, Stockholm.

Nahavandchi H, Sjöberg LE (1998) Terrain correction to power H3 in gravimetric geoid determination. J

Geod, 72: 124-135.

Nahavandchi H, Sjöberg LE (2001) Precise geoid determination over Sweden using the Stokes-Helmert
method and improved topographic corrections. J Geod, 75: 74-88.

Novak P, Vaníček P, Martinec Z, Véronneau M (2001) Effects of the spherical terrain on gravity and the
geoid. J Geod, 75: 491-504.

Omang OCD, Forsberg R (2002) The northern European geoid: a case study on long-wavelenght geoid
errors. J Geod, 76: 369-380.

Parseliunas E, Forsberg R (1999) The improvement of the geoid model of Lithuania. IUGG 1999
Birmingham, Abstracts, vol. A, p. A.436.

77
Parseliunas E, Buga A (1995) GPS activities, coordinate systems, geoid determination in Lithuania. In
Vermeer M (ed): Coordinate Systems, GPS and the geoid. Reports of the Finnish Geodetic Institute 95:4,
pp. 17-30. Helsinki.

Paul NK (1973) A method of evaluating the truncation error coefficients for geoidal height. Bull Geod
110: 413-425.

Rapp R, Rummel R (1975) Methods for the computation of detailed geoids and their accuracy. Dept. of
Geodetic Science and Surveying Report, No. 208. The Ohio State University, Columbus, Ohio.

Rapp R, Pavlis NK (1990) The Development and analysis of geopotential coefficient models to spherical
harmonic degree 360. J. Geophys. R., 95(B13): 21885-21911.

Rummel R (1997) Spherical spectral properties of the Earth’s gravitational potential and its first and
second derivatives. In Sansó F and Rummel R (eds.): Geodetic Boundary Value Problems in View of the
One Centimeter Geoid, Lecture Notes in Earth Sciences 65, pp. 359–404, 1997.

Rummel R, Müller J, Oberndorfer H, Sneeuw N (2000) Satellite Gravity Gradiometry with GOCE. In:
Rummel R, Drewes H, Bosch W, Hornik H (eds.) Towards an Integrated Global Geodetic Observing
System. IAG Symp. Series 120, 66-72, Springer Verlag.

Sansó F (1997) (ed) Lecture notes, International School for the Determination and Use of the Geoid.
International Geoid Service, DIIAR-Politecnico di Milano, Milan.

Sideris MG, Schwartz K-P (1987) Improvement of medium and short wavelength features of geopotential
solutions by local gravity data. Boll Geod Sci Aff, 46(3): 207-221.

Sildvee H (1998) Gravity measurements of Estonia. Reports of the Finnish Geod Inst, 98:3, Masala.

Sjöberg LE (1977) On the errors of spherical harmonic developments of gravity at the surface of the
earth. Depart. of Geodetic Science, Report no. 257, The Ohio State University.

Sjöberg LE (1980) Least squares combination of satellite harmonics and integral formulas in physical
geodesy. Gerl Beitr Geophys 89(5): 371-377.

Sjöberg LE (1981) Least squares combination of satellite and terrestrial data in physical geodesy. Ann
Geophys, 37(1):25-30.

Sjöberg LE (1984a) Least squares modification of Stokes and Vening-Meinesz formulas by accounting
for errors of truncation and potential coefficient errors. Manuscr Geod, 9: 209-229.

Sjöberg LE (1984b) Least squares modification of Stokes and Vening-Meinesz formulas by accounting
for errors of truncation, potential coefficients and gravity data. The Dept. of Geodesy report No. 27,
University of Uppsala.

Sjöberg LE (1986) Comparison of some methods of modifying Stokes formula. Bollettino di Geodesia e
Scienze Affini, 46: 229-248.

Sjöberg LE (1991) Refined least-squares modification of Stokes formula. Manuscr Geod, 16: 367-375.

Sjöberg LE (1993) Terrain effects in the atmospheric gravity and geoid corrections. Bull Geod, 64: 178-
184.

Sjöberg LE (1994) On the terrain effects in Geoid and Quasigeoid determinations using Helmerts’s
second condensation method. Division of Geodesy, TRITA. GEOD Report 1036, 1994, Royal Institute of
Technology, Stockholm.

78
Sjöberg LE (1996) The terrain effect in geoid computation from satellite derived geopotential models.
Bollettino di Geodesia e Scienze Affini, 55: 385-392.

Sjöberg LE (1997) The total terrain effect in gravimetric geoid determinations. Bollettino di Geodesia e
Scienze Affini, 56: 209-222.

Sjöberg LE (1998a): The atmospheric geoid and gravity corrections. Vermeer M, Ádám J (eds): Second
continental Workshop on the geoid in Europe. Budapest, 10-14 March, 1998. Reports of the Finnish
Geodetic Institute No. 4, pp. 161-167.

Sjöberg LE (1998b) The exterior Airy/Heiskanen topographic-isostatic gravity potential, anomaly and the
effect of analytical continuation in Stokes’ formula. J Geod, 72: 654-662.

Sjöberg LE (1999) The IAG approach to the atmospheric geoid correction in Stokes formula and a new
strategy. J Geod, 73: 362-366.

Sjöberg LE (2000) Topographic effects by the Stokes-Helmert method of geoid and quasi-geoid
determinations. J Geod, 74: 255-268.

Sjöberg LE (2001a) Topographic and atmospheric corrections of gravimetric geoid determination with
special emphasis on the effects of harmonics of degrees zero and one. J Geod, 75: 283-290.

Sjöberg LE (2001b) The effect of downward continuation of gravity anomaly to sea level in Stokes’
formula. J Geod, 74: 790-804.

Sjöberg LE (2003a) A solution to the downward continuation effect on the geoid determined by Stokes’
formula. J Geod, 77: 94-100.

Sjöberg LE (2003b) Ellipsoidal corrections to order e2 of geopotential coefficients and Stokes’ formula. J
Geod, 77: 148-154.

Sjöberg LE (2003c) A computational scheme to model the geoid by the modified Stokes formula without
gravity reductions. J Geod, 77: 423-432.

Sjöberg LE (2003d) A general model of modifying Stokes’ formula and its least-squares solution. J Geod,
77: 459-464.

Sjöberg LE (2003e) The correction to the modified Stokes formula for an ellipsoidal Earth. In Santos M
(ed.): Honouring the academic life of Petr Vaníček, UNB Technical Report No 218, New Brunswick, pp
99-110.

Sjöberg LE (in print a) A spherical harmonic representation of the ellipsoidal correction to the modified
Stokes formula. In print J Geod.

Sjöberg (in print b) The effect on the geoid of lateral topographic density variations. In print J Geod.

Sjöberg LE (submitted) On the theory Used in Geoid Modelling by the Remove-Compute-Restore

Technique. Submitted to J Geod.

Sjöberg LE, Hunegnaw A (2000) Some modifications of Stokes formula that account for truncation and
potential coefficient errors. J Geod, 74: 232-238.

Sjöberg LE, Ågren J (2002) Some Problems in the Theory used for the NKG Geoid Model. In M.
Poutanen and H. Suurmäki (Eds.): Proceedings of the 14th General Meeting of the Nordic Geodetic
Commission, pp. 121-129, Kirkonummi/Helsinki.

Stokes GG (1849) On the variations of gravity on the surface of the Earth. Trans Camb Phil Soc 8:672-
695.

79
Tapley B, Bettadpur S, Chambers D, Cheng M, Choi K, Gunter B, Kang Z, Kim J, Nagel P, Ries J, Rim
H, Roesset P., Roundhill I (2003) Gravity Field Determination from CHAMP Using GPS tracking and
Accelerometer Data: Initial Results. In: C. Reigber, H. Lühr, P. Schwintzer (Eds.) First CHAMP Mission
Results for Gravity, Magnetic and Atmospheric Studies, 1 st CHAMP Science Meeting, 21-24 January,
2002, Potsdam, Germany. Berlin, Springer.

Tikhonov AN (1963) Solution of incorrectly formulated problems and the regularization method, Soviet
Math. Dokl., 4 (1963), pp. 1035-1038; English translation of Dokl. Akad. Nauk. SSSR, 151: 501-504.

Tscherning C, Rapp R(1974) Closed covariance expressions for gravity anomalies, geoid undulations and
deflections of the vertical implied by anomaly degree variance models. Department of Geodetic Science,
Report No. 208.

Tscherning CC, Forsberg R, Knudsen P (1992) The GRAVSOFT package for geoid determination. Proc.
1-st continental workshop on the geoid in Europe, Prague 1992, pp. 327-334.

Vaníček P, Krakiwsky E (1986) Geodesy: the concepts. Elsevier. North- Holland, Amsterdam, New
York, Oxford, Tokyo.

Vaníček P, Kleusberg A (1987) The Canadian geoid – Stokesian approach. Manuscr Geod, 12: 86-98.

Vaníček P, Sjöberg LE (1991) Reformulation of Stokes’s theory for higher than second-degree reference
field and modification of integration kernels. J Geophys Res, 96(B4): 6529-6339.

Vaníček P, Kleusberg A, Martinec Z, Sun W, Ong P, Najafi M, Vajda P, Harrie L, Tomasek P, ter Horst
B (1996) Compilation of the precise regional geoid. Department of Geodesy and Geomatics Engineering
Technical report No. 184, University of New Brunswick, Canada.

Vaníček P, Featherstone WE (1998) Performance of three types of Stokes’s kernel in the combined
solution for the geoid. J Geod, 72: 684-697.

Vermeer M (1994) A fast delivery GPS-gravimetric geoid for Estonia. Rep. of Finnish Geod. Inst.
Helsinki.

Vermeer M (1995) Two new geoids determined at the FGI. Report 95:5 Finnish Geodetic Institute.

Vincent S, Marsh J (1974) Gravimetric global geoid. In proc. of Int. Symp. On the use of artifical
satellites for geodesy and geodynamics, ed G. Veis, National Technical University, Athens, Greece

Weber G, Wenzel H-G (1983) Error covariance functions of sea gravity data and implication for geoid
determination, Presented to IAG General Assembly, Tokyo 1982. Marine Geodesy, 7: 199-226.

Wenzel H-G (1981) Zur Geoidbestimmung durch Kombination von Schwereanomalien und einem
Kugelfunktionsmodell mit Hilfe von Integralformeln. Z. Vermess 106:102-111.

Wenzel H-G (1983) Geoid computation by least squares spectral combination using integral kernels.
Proceedings IAG General Meeting, Tokyo, pp. 438-453, Springer Verlag.

Wenzel H-G (2000) Global Models of the gravity field of high and ultra-high resolution. In Sanso (ed):
Lecture notes, International School for the Determination and Use of the Geoid, hold at Johor, 2000.
International Geoid Service.

Wong L, Gore R, (1969) Accuracy of geoid heights from the modified Stokes kernels. Geophysical
Journal Royal Astronomy Society, 18: 81-91.

Ågren J, Sjöberg LE (in print) Comparison of some methods for modifying Stokes' formula in the GOCE
era. 2nd International GOCE user workshop, hold at ESA-ESRIN, Frascati, Italy March 08-10, 2004.

80
Part TWO
______________________________________________________________________

Publications

(not included in the electronical version of the Dissertation)

81
______________________________________________________________________
List of Publications

PAPER A:
Ellmann A (submitted) Testing of some deterministic and stochastic modifications of
Stokes’ formula: a case study for the Baltic countries. Submitted to Journal of Geodesy.

PAPER B:
Ellmann A (in print a) Effect of GRACE satellite mission to gravity field studies in
Fennoscandia and the Baltic Sea region. Proc. Estonian Acad. Sci. Geol., 53 (No. 2).

PAPER C:
Ellmann A (2002) An improved gravity anomaly grid and a geoid model for Estonia.
Proc. Estonian Acad. Sci. Geol., 51: 199-214.

PAPER D:
Ellmann A (in print b) On the numerical solution of parameters of the least squares
modification of Stokes’ formula. IAG Symp. Series. IUGG 2003. Springer Verlag.

PAPER E:
Ellmann A, Sjöberg LE (2002) Combined topographic effect applied to the biased type
of the modified Stokes formula. Boll. Geod. Sci. Aff., 61: 207-226.

PAPER F:
Ellmann A (in print c) A numerical comparison of different ellipsoidal corrections to
Stokes’ formula. IAG Symp. Series. IUGG 2003. Springer Verlag.

PAPER G:
Ellmann A, Sjöberg LE (in print) Ellipsoidal correction for the modified Stokes
formula. Boll. Geod. Sci. Aff., 63 (No. 3).

82
"No matter how much you know about a subject, there's always
someone else who knows more; including yourself, one week later"

Anonymous
Geodesy Report 1000 Series
Dissertations, Licentiate Theses & Research Reports

1001 Sjöberg, Lars E. Some Methods of Estimating Geoid Undulations by Least Squares
Combination of Stokes's Formula and Geopotential Coefficients. November 1985.
1002 Sjöberg, Lars E. A Comparison of Some Modification Methods to Stokes's Formula Using
GEM 9 Potential Coefficients. November 1985.
1003 Sjöberg, Lars E. and Huaan Fan. Studies on the Secular Land Uplift and Long Periodic
Variations of Sea Level around the Coasts of Sweden. August 1986.
1004 Sjöberg, Lars E. and Huaan Fan. A Comparison of the Modified Stokes's Formula and
Hotine's Formula in Physical Geodesy. November 1986.
1005 Fan, Huaan. A General Model for Least Squares Adjustment and Estimation. November 1986.
1006 Galvenius, Göran. Metoder och instrument för inmätning av underjordiska ledningar (Procedures
and Instruments for Positioning of Sub-surface Pipes and Cables). November 1986.
1007 Fan, Huaan and Lars E. Sjöberg. On the Estimation of Variance-Covariance Components with
and without Invariance in the Gauss-Helmert Model. December 1986.
1008 Cederholm, Torbjörn. System för rationell hantering av lägesbunden information. Delrapport 1:
Metoder för insamling av lägesdata - en översikt (Systems for Efficient Handling of Spatially
Related Data. Part One: Data Collection Methods - an Overview.). February 1987.
1009 Vanicek, Petr and Lars E. Sjöberg. A Note on Vertical Crustal Movement Determination
Techniques. February 1987.
1010 Sjöberg, Lars E. and Huaan Fan. Experiences with the UNB Land Uplift Modelling Program
Using Swedish Levelling and Tide Gauge Data. March 1987.
1011 Galvenius, Göran. Metoder och instrument för inmätning av underjordiska ledningar - Teoretiska
studier och fältförsök (Procedures and Instruments for Measuring Positions of Subsurface Cables
and Pipes - Theoretical Studies and Field Research). 1988.
1012 Cederholm, Torbjörn. System för rationell hantering av lägesbunden information. Delrapport 2:
System för geodetisk datafångst (Systems for Efficient Handling of Spatially Related Data. Part
Two: Field Survey Data Collection System). 1988.
1013 Holmberg, Stig C. Planning Support Facilities in Geoinformatic Systems: An Environment for
Handling, Visualizing and Communicating Geoinformation. September 1988. (Doctoral
Dissertation)
1014 Sjöberg, Lars E., Huaan Fan and Erick Asenjo. Studies on the Secular Change of Gravity in
Fennoscandia. September 1988.
1015 Fan, Huaan and Lars E. Sjöberg. First Order Terrain Corrections to Geopotential Coefficients - A
Numerical Investigation. January 1989.
1016 Galvenius, Göran. Metoder och instrument för inmätning av underjordiska ledningar - Fortsatta
studier. January 1989.
1017 Asenjo, Erick and Lars E. Sjöberg. Tröghetspositionering - en överblick med tonvikt på
utjämning av Ferranti-data (Inertial Surveying - A Review with Emphasis on the Adjustment of
Ferranti Data). March 1989.
1018 Grafarend, Erik W. The geoid and the Gravimetric Boundary Value Problem. August 1989.
1019 Fan, Huaan. Geoid Determination by Global Geopotential Models and Integral Formulas. August
1989. (Doctoral Dissertation)
1020 Egeltoft, Tomas. GPS-bestämning av stompunkter vid Husarviken (Determination of Control
Points at Husarviken with GPS). December 1989.
1021 Sjöberg, Lars E. Systematic Tropospheric Errors in Geodetic Positioning with the Global
Positioning System. January 1990.
1022 Holmberg, Stig C. User´s Requirements in Geoinformatic Systems. February 1990. (Doctoral
Dissertation)
1023 Cederholm, Torbjörn and Johan Larsén. GIS-terminologi, Terminologiprojektets insamlingsfas
(GIS-terminology, collection phase). August 1990.
1024 Sjöberg, Lars E. Contributions to the Problem of Modifying Stokes's Formula. September 1990.
1025 Pan, Ming. GPS Orbit Improvement. May 1991.
1026 Alemu, Abera. The Gravity Field and Crustal Structur of the Main Ethiopian Rift Valley. January
1992. (Doctoral Dissertation)
1027 Egeltoft, Tomas. Variance Component Estimation in Geodetic Networks. October 1992.
(Licentiate Thesis)
1028 Nord, Tomas. Numerical studies in Physical Geodesy. December 1992. (Doctoral Dissertation)
1029 Pan, Ming. The First Baltic Sea Level GPS Campaign: Data Processing. December 1992.
(Licentiate Thesis)
1030 Fan, Huaan. Some Studies on the Gravity Field in Antarctica. May 1993.
1031 Mcharo, Ibrahim. Underground Geodetic Measurements. May 1993. (Licentiate Thesis)
1032 Zhao, Shaorong. Design of Observational Precision of Monitoring Networks Taking into Account
of A Priori Deformation Information. November 1993.
1033 Zhao, Shaorong and Lars E Sjöberg. Theory for Inversion of Dynamic Geodetic Data:
Mathematical Rationale and Some Applications. December 1993.
1034 Zhao, Shaorong. Detection of Active Fault Segment by Joint Inversion of Observed Gravity and
GPS Baseline Changes. February 1994.
1035 Jansson, Patric. Underwater Positioning. November 1994. (Licentiate Thesis)
1036 Sjöberg, Lars E. On the Terrain Effects in Geoid and Quasigeoid Determinations Using Helmert's
Second Condensation Method. December 1994.
1037 Almgren, Kjell. Detail Surveying with GPS. March 1995. (Licentiate Thesis)
1038 Hagenäs, Per. Positioning with Supported GPS - A Simulation Study. April 1995. (Licentiate
Thesis)
1039 Pan, Ming. Application of the Global Positioning System to the Baltic Sea Level Project. August
1995. (Doctoral Dissertation)
1040 Egeltoft, Tomas. Data Analysis and Adjustment in Precise Levelling with Applications to the
Third Swedish Precise Levelling. January 1996. (Doctoral Dissertation)
1041 Nsombo, Peter. Geoid Determination Over Zambia. June 1996. (Doctoral Dissertation)
1042 Galvenius, Göran. Reflector Constants of EDM. April 1996.
1043 Sun, Wenke and Lars E. Sjöberg. Gravitational Potential Changes of a SNREI Earth Model to a
Point Surface Mass Load. November 1996.
1044 El-Rabbany, Ahmed. Temporal Characteristics of Multipath Errors. November 1996.
1045 Pan, Ming, Lars E. Sjöberg, Erick Asenjo, Abera Alemu and Laike M. Asfaw. An Analysis of
the Ethiopian Rift Valley GPS Campaign in 1989 and 1994. January 1997.
1046 Sun, Wenke, Lars E. Sjöberg. A Global Topographic-Isostatic Model. October 1997.
1047 Almgren, Kjell. A new method for GPS ambiguity resolution on-the-fly at short baselines.
February 1998. (Doctoral Dissertation)
1048 Jansson, Patric. Precise Kinematic GPS Positioning with Kalman Filtering and Smoothing- Theory
and Applications. April 1998. (Doctoral Dissertation)
1049 Sjöberg, Lars E. Linear Estimation of GPS Phase Ambiguities from Dual Frequency Code and
Phase Observables. December 1998.
1050 Nahavandchi, Hossein. Precise Gravimetric-GPS Geoid Determination with Improved
Topographic Corrections Applied over Sweden. Doctoral Dissertation. December 1998.
1051 Horemuz, Milan and Lars E Sjöberg. Quick GPS Ambiguity Resolution For Short And Long
Baselines. March 1999.
1052 Jan Danielsen. Determination of land uplift in areas not covered by repeated levellings. With
application to South-Norway. June 1999. (Licentiate thesis)
1053 Wenke Sun and Lars Sjöberg. Tidal Effects on Sea Bottom Positioning by Combining GPS
and Sonar Observations. March 2000.
1054 Stig-Göran Mårtensson. Height Determination by GPS – a Practical Experiment in Central
Sweden. May 2001. TRITA-GEOFOTO 2001:12. (Doctoral dissertation)
1055 Jonas Ågren. Processing of the 1992, 1994 and 1997 Campaigns on the Northern GPS
Deformation Traverse. May 2001. TRITA-GEOFOTO 2001:13. (Licentiate thesis).
1056 Artu Ellmann. Least squares modification of Stokes formula with application to the Estonian
geoid. November 2001. TRITA-GEOFOTO 2001:22. (Licentiate thesis)
1057 Addisu Hunegnaw. Geoid Determination over Ethiopia with Emphasis on Downward
Continuation of Gravity Anomalies. November 2001. (Doctoral dissertation)
1058 Huaan Fan. Development of the Mozambican Geoid Model MOZGEO 2002. TRITA-INFRA
2002:13, March 2002.
1059 Lars Sjöberg, Erick Asenjo, George Stoimenov and Peter Nsombo. Sustainable Land
Development in Zambia. September 2002.
1060 Matthias Adam. Establishment of a calibration field for a mobile mapping system. January
2003.
1061 Artu Ellmann. The geoid for the Baltic countries determined by the least squares modification of
Stokes’ formula. TRITA-INFRA 2004-013. March 2004. (Doctoral dissertation)