J Geod (2014) 88:749–764

DOI 10.1007/s00190-014-0719-7

ORIGINAL ARTICLE

Robust estimation of deformation from observation differences

for free control networks
Krzysztof Nowel · Waldemar Kamiński

Received: 22 May 2013 / Accepted: 8 April 2014 / Published online: 9 May 2014
© The Author(s) 2014. This article is published with open access at Springerlink.com

Abstract Deformation measurements have a repeatable describes the algorithm of classical robust methods. Partic-
nature. This means that deformation measurements are per- ular attention was paid to the IWST method. In relation to
formed often with the same equipment, methods, geomet- classical robust methods, the optimization problem of the
ric conditions and in a similar environment in epochs 1 new REDOD method was formulated and the algorithm for
and 2 (e.g., a fully automated, continuous control measure- its solution was derived.
ments). It is, therefore, reasonable to assume that the results
of deformation measurements can be distorted by both ran- Keywords Robust estimation · Stability of reference
dom errors and by some non-random errors, which are con- points · Displacement · S-transformation · IWST · REDOD
stant in both epochs. In other words, there is a high prob-
ability that the difference in the accuracy and precision of
measurement of the same geometric element of the network 1 Introduction
in both epochs has a constant value and sign. The constant
errors are understood, but the manifestation of these errors is Absolute and relative control networks are distinguished in
difficult to determine in practice. For free control networks deformation measurements and analysis. Absolute control
(the group of potential reference points in absolute control networks consist of controlled points placed on the surveyed
networks or the group of potential stable points in relative object (e.g., on a dam, bridge) and potential reference points
networks), the results of deformation measurements are most usually placed outside the effect of factors causing deforma-
often processed using robust methods. Classical robust meth- tions. Relative control networks consist only of controlled
ods do not completely eliminate the effect of constant errors. points established on the surveyed object (e.g., on an area
This paper proposes a new robust alternative method called covered by mining damage). Both potential reference points
REDOD. The performed tests showed that if the results of of absolute networks and potential stable points of relative
deformation measurements were additionally distorted by networks (the points are on a potentially rigid network frag-
constant errors, the REDOD method completely eliminated ment) form FCN (free control networks). The problem of
their effect from deformation analysis results. If the results deformation analysis for FCN is, in fact, the problem of iden-
of deformation measurements are only distorted by random tifying the mutual stability of these network points. In this
errors, the REDOD method yields very similar deformation way, the datum for the displacement vector of all absolute or
analysis results as the classical IWST method. The numerical relative control network points is defined. It can be seen that
tests were preceded by a theoretical part. The theoretical part deformation analysis for FCN is, therefore, an essential task
in deformation analysis for control networks (Chen 1983;
K. Nowel (B) · W. Kamiński Caspary 2000; Denli 2008).
Institute of Geodesy, University of Warmia and Mazury in Olsztyn,
Methods of deformation analysis for FCN described in lit-
1 Oczapowskiego Str., 10-719 Olsztyn, Poland
e-mail: [email protected] erature can be classified according to different criteria. In this
paper, they are divided into robust methods and non-robust
W. Kamiński methods (Chrzanowski and Chen 1990; Caspary 2000; Setan
e-mail: [email protected] and Singh 2001). Today, the robust methods are applied more

123
750 K. Nowel, W. Kamiński

often. It should be stressed that “robust” does not refer here bility and the processing of the results. A higher reliability
to outliers in observations; “robust” refers here to single- level of a measurement produces a lower risk of making gross
point movements. The observations must be clean of outliers. and systematic errors. A high measurement reliability level is
This is very important. Classical robust methods are based on achieved by, among others, periodical instrument checks, the
the results of separate adjustments by the least squares (LS) use of self-checking measurement procedures and the intro-
method of two measurement epochs, i.e., on adjusted coordi- duction of corrections due to the disturbing effects of external
nates. These methods consist in S-transformation (Helmert conditions. The higher the reliability level of the processing
similarity transformation) of differences in these coordinates of measurement results, the higher the probability will be of
with the optimization condition of robust estimation. The detecting possible gross and systematic errors (internal relia-
obtained displacement vectors of single points are assessed bility) and the lower the effect of these undetected errors will
statistically for significance. The best-known methods from be on the quality of deformation analysis results (external
this group include the iterative weighted similarity transfor- reliability), e.g., Caspary (2000), Prószyński (2010).
mation (IWST) method, which fulfills the condition of the Deformation measurements are performed at different
minimum sum of absolute component values of displace- moments of time, but often with the same equipment, meth-
ment vectors (Chen 1983; Chen et al. 1990; Setan and Singh ods, geometric conditions and in a similar environment in
2001; Setan and Othman 2006; Gökalp and Tasçi 2009) and epochs 1 and 2. Moreover, the same instrument and target are
the least absolute sum (LAS) method, which fulfills the con- most often set on given points in both epochs. It is, therefore,
dition of the minimum sum of length values of displacement reasonable to assume that the results of deformation measure-
vectors (Caspary 1984; Caspary and Borutta 1987; Caspary ments can be distorted both by random errors and by some
et al. 1990; Setan and Singh 2001). Proposals for the use of non-random errors, which are constant in both epochs, i.e.,
other optimization conditions, e.g., the minimum objective errors with the same value and the same sign in both epochs.
function of the Huber method or the Danish method, can also This means that the error of measurement of the same geo-
be found in literature (Caspary and Borutta 1987). The IWST metric element of the network in both epochs may contain a
method serves as the prototype for classical robust methods. common factor, which is constant in both epochs. This fac-
The other robust methods are only modifications, differing in tor is, in fact, the difference in the measurement accuracy
the optimization condition used. The non-robust methods are and precision. The constant errors are understood, but the
also based on the results of separate adjustments by the LS manifestation of these errors is difficult to determine in prac-
method of two measurement epochs, i.e., on adjusted coor- tice. Classical deformation analysis of FCN is most often
dinates. Statistical tests are most popular in this group, e.g., performed by robust methods. In these methods, the dis-
the congruency test (Niemeier 1981; Setan and Singh 2001; placement vector is determined in robust S-transformation of
Denli and Deniz 2003). The congruency test method con- differences in adjusted coordinates. This approach does not
sists in finding the most numerous possible group of points completely eliminate the effect of constant errors and, there-
for which the value of the displacement test statistic does not fore, leads to reduced external reliability level of processing
exceed the critical value. Displacement vectors are computed and, consequently, to reduced quality of deformation analy-
on the basis of S-transformation of differences in adjusted sis results. Moreover, the displacement vector and its covari-
coordinates of these points with the optimization condition ance matrix are estimated separately in two different com-
of the LS method or on the basis of differences in coordinates putational processes. As a result, the displacement vector
obtained from free adjustments by the LS method. Although estimator and its covariance matrix estimator have different
many more other new methods for deformation analysis, e.g., stochastic properties, i.e., the actual and estimated accuracy
robust method based on Msplit estimation (Wiśniewski 2009; of the displacement vector estimator is different.
Duchnowski and Wiśniewski 2012) or R estimation (Duch- In the 1970s, the Polish geodesist Prof. Tadeusz Laz-
nowski 2010, 2013) or the non-robust Fredericton approach zarini dealt with this problem for control networks with a
(Gökalp and Tasçi 2009), can be found in literature, this paper defined datum. Prof. Lazzarini developed a method for joint
is limited to the presentation of only the most common solu- estimation of the displacement vector and its covariance
tions. matrix directly from differences in unadjusted observations
Geodetic methods are most often used to measure minor (Lazzarini et al. 1977; Chrzanowski and Chen 1990). This
deformations with values slightly exceeding errors in their approach completely solves, among others, the problem of
measurement. A special role in deformation analysis is, there- the effect of constant errors and the problem of discrepancy
fore, played by the quality of results, i.e., the accuracy of the between the actual and estimated accuracy of the displace-
determined displacement vectors and the correct assessment ment vector estimator. This method is still used successfully
of their statistical significance (effectiveness). The quality of in deformation analysis of control networks with a defined
deformation analysis results is conditioned by the measure- datum, among others in an automated system for continuous
ment precision (random errors) and the measurement relia- control measurements ALERT (Wilkins et al. 2003). Mod-

123
Robust estimation of deformation from observation differences 751

ern robust estimation and matrix g-inverse theory allow this adjusted coordinates and N− = (AT PA)− is any g-inverse,
approach to be expanded to FCN (control networks without fulfilling the property NN− N = N. The network is a free
a defined datum). Such expansion is the aim of this paper. In network. In this case, the network has a datum defect
fact, a new robust method for FCN deformation analysis will
de = u − rank(A), rank(A) = r < u (4)
be developed and examined in this paper, in which the dis-
placement vector will be determined in the process of robust the matrix A is the matrix of the columnarly incomplete rank
M estimation of differences in unadjusted observations. The and, therefore, the ordinary (classical) inverse N−1 will not be
proposed method was conventionally called robust estima- determined. The g-inverse determination methods are well-
tion of deformation from observation differences (REDOD). described in literature (e.g., Rao 1973; Caspary 2000). For
example, one of the g-inverses can be determined in a simple,
algebraic manner, by representation of the design matrix in
2 Robust transformation of deformation from the following form
coordinate differences  
A = A1 A2 , (5)
The robust FCN deformation analysis method proposed in where A1 ∈ n×r , r = rank(A1 ) = rank(A) and A2 ∈
this paper will be developed based on classical robust meth- n×de is, in fact, the matrix defining the datum for the geo-
ods and is specifically based on the IWST method. There- detic network. The final result of step 1 is the vectors x̂(e) (1)
fore, the algorithm of classical robust methods was derived and their cofactor matrices Qx̂(e) , where e is the number of
and described in this section. Particular attention was paid to the measurement epoch.
the IWST method. Of course, during LS estimation of the vector of coordi-
As previously mentioned, in classical methods the dis- nates, other important aspects that need to be considered are
placement vector is determined in the process of robust S- outlier detection (observations) (e.g., Baarda’s data snoop-
transformation of the differences in adjusted coordinates. ing, Pope’s test or some robust estimation) and tests of the
Generally, the algorithm always consists of the following compatibility between the two epochs (variance ratio test).
three steps. Moreover, in the case of heterogeneous networks, it is some-
times important to perform the estimation of the variance and
2.1 Step 1: LS estimation of the vector of coordinates covariance components.

LS estimation of the vector of coordinates is carried out sep- 2.2 Step 2: robust S-transformation of the displacement
arately for each measurement epoch. The computations can- vector
not distort measurement results or the relationships occurring
between them. The basic aim is to provide data for starting Displacements of individual points in an FCN cannot be
robust transformation. The result of this processing is the vec- determined on the basis of differences in the direct values of
tor of adjusted coordinates for all analyzed points in an FCN the vectors x̂(1) and x̂(2) . LS estimation of the vector of coor-
dinates can concern networks with different datum defect
x̂ = x0 + δ̂x , (1)
types in both epochs (because of different types of obser-
where x0 ∈ u×1 is the vector of the approximate coordi- vations). Moreover, constraints defining the datum for the
nates (the same in both epochs) and δ̂x ∈ u×1 is the LS geodetic network in epochs 1 and 2 can concern different
estimator of the vector of the corrections to the approximate points (e.g., because of damage to some points in epoch 2)
coordinates. or can concern the same unstable points. For this reason,
The functional model, which is the linearized form of the the datum for the geodetic network in epoch 2 can be freely
initial nonlinear relationships, has the form shifted, rotated and re-scaled in relation to the datum in epoch
1 (Chen 1983; Chen et al. 1990). The vector x̂(2) can, there-
lobs − l0 + v = Aδx , lobs ∼ N (Ax, C) , (2)
fore, be adjusted in a different datum than the vector x̂(1) and
where A ∈ n×u is the design matrix, l0 ∈ n×1 is the vec- this is why
tor of approximate observations, lobs ∈ n×1 is the vector of
d = x̂(2) − x̂(1) , (6)
actual observations, v ∈ n×1 is the vector of residuals of
observations and C ∈ n×n is the covariance matrix of obser- where d ∈ u×1 is the vector of displacement components
vations. The sought parameter vector estimator has the form for all analyzed points in an FCN. This difference x̂(2) − x̂(1)
is the so-called apparent (or raw) displacement vector.
δ̂x = Qx̂ ω, (3)
To reduce the vectors x̂(1) and x̂(2) to a common datum,
where ω = AT P(lobs −l0 ), P ∈ n×n is the weight matrix of the vector x̂(2) is transformed to the datum of the vector x̂(1) .
observations, Qx̂ = N− is the cofactor matrix of the vector of This task consists in fitting the vector x̂(2) into the vector

123
752 K. Nowel, W. Kamiński

x̂(1) reduced to the center of gravity of the network. The and yield greater variances. This decrease in efficiency is the
 has the form
transformed vector x̂(2) price to be paid for the robustness of the method.
In the IWST method, the objective function is the L1 -

x̂(2) = x̂(2) − Ht (7) norm of vector d, i.e., (Chen 1983; Chen et al. 1990; Setan
and Singh 2001; Setan and Othman 2006; Gökalp and Tasçi
 + Ht, where t ∈ h×1 is the vector of
because x̂(2) = x̂(2) 2009). The optimization problem of the IWST method has
S-transformation parameters, i.e., translation, rotation, scale the following form
distortion, h designates the number of all datum parame-
ters of a given network and H is the design matrix of S-
transformation to a common datum (epoch 1). For example, ,
for a 2-D network, the matrix H has the form
⎡ ⎤
.. .. .. ..
⎢ . . . . ⎥
⎢ 1 0 − ȳi(1) x̄i(1) ⎥
H∈ 2m×h
=⎢⎢ ⎥, (8) (9)
⎥
⎣ 0 1 x̄i(1) ȳi(1) ⎦
.. .. .. .. where d = [d1 , . . . , du ]T , di is the given component of
. . . . vector d for point i (d xi or dyi or dz i ), u is the number of
where m designates the number of points in a given network the displacement components for all m analyzed points in an
and x̄i(1) , ȳi(1) designate adjusted coordinates of point i in FCN. The final solution to the optimization problem of the
epoch 1 reduced to the center of gravity of the network. The IWST method has the following iterative form

 ⎫
first two columns refer to translation along the axes X, Y , d̂(k) =S(k) x̂(2) − x̂(1) ⎪
respectively, the third column refers to rotation around the (k) (k)

(k)
T ⎪
⎬
Q =S Qx̂(1) + Qx̂(2) S (10)
axis Z and the last column refers to scale distortion. More d̂  ⎪ ⎪
theory on that issue (among others, the matrix H for a 1-D W(k+1) = diag . . . , w (k+1)
i , ... ⎭
k=1,2,...
and 3-D network) can be found in papers by Chen (1983),
Caspary (2000) and Even-Tzur (2011). where S(k) = I − H(HT W(k) H)−1 HT W(k) is the matrix of
Since the displacement vector now has the form d = S-transformation of the displacement vector to a common
 − x̂ , consequently d = (x̂
x̂(2) (1) (2) − x̂(1) ) − Ht. Of course, datum in epoch 1, W(k=1) = I, I ∈ u×u is the identity
this system of equations has infinitely many solutions for the matrix and
 
displacement vector. To obtain a unique, sought value of this (k+1)  (k) 
wi = 1/ d̂i  (11)
vector, we also need to formulate proper objective functions.
Let us recall that FCN may consist of potential reference
It is possible that during the iterations (10) some d̂i(k)
points of absolute networks or potential stable points of rel-
may approach
 zero, causing numerical instabilities, because
ative networks. Hence, we can assume that most of these  (k) 
points (more than half) are actually stable. Unfortunately, 1/ d̂i  becomes very large. For this reason in practice, Eq.
there is a risk that single unstable points may occur among (11) is replaced with the equation
  
this group. It can be assumed that measurement errors at 
all epochs are a sample of random variables with a normal wi(k+1) = 1 d̂i(k)  + c , (12)
distribution with a mean of zero and accepted standard devi- where c is the precision for computations, e.g., 0.0001 m.
ation. In this case, the estimated displacements of actually Proposals for uses of other solutions in this task can be
stable points should also be a sample of random variables found in other papers, e.g., Marx (2013). The iterative
with such a distribution and the estimated displacements of process (10) isfinished when all components of the difference

actually unstable points should be random variables which  (k+1) 
d̂ − d̂(k)  are lower than the adopted precision for com-
are the outliers of this distribution. In other words, the vec-
tor of displacement components of FCN points may have putations c (convergence). The final robust estimator of the
a contaminated normal distribution (analogously as, e.g., a vector of displacement components for all analyzed points
vector of observations with suspected gross errors). This is in an FCN, i.e., vector d̂ R and the final cofactor matrix of
why the robust objective functions are very useful in defor- this vector, i.e., the matrix Qd̂ R , is obtained from the last
mation analysis for FCN. Thus, “outliers” do not refer here iterative step. Vector d̂ R consists of vectors d̂iR for single
to the observations. “Outliers” refer here to the single-point points in an FCN, i.e., (d̂ R )T = [. . . , (d̂iR )T , . . .], and the
movements. It should be stressed that if all of the FCN points matrix Qd̂ R , neglecting the correlation between the points
are certainly actually stable, then the LS and robust methods and contains cofactor matrices of these vectors, i.e., Qd̂ R =
lead to unbiased results, but robust methods are less efficient diag(. . . , Qd̂ R , . . .). Proposals for uses of other objective
i

123
Robust estimation of deformation from observation differences 753

functions ψ(t) in this task (among others, the objective func- network) (e.g., Cederholm 2003). In this case, the signifi-
tion of the LAS method, the Huber method or the Danish cance assessment consists in checking graphically whether
method) can be found in papers by Caspary and Borutta the vector d̂iR does not exceed the confidence interval for
(1987), Caspary et al. (1990), Setan and Singh (2001). determination of this vector (1-D network) or the confidence
ellipse for determination of this vector (2-D network) or the
2.3 Step 3: assessment of the significance of the confidence ellipsoid for determination of this vector (3-D
displacement vector network). Of course, both solutions (15), (16) give the same
results.
All d̂iR vectors are assessed statistically for significance. The It should be stressed that the matrix Qd̂ R is approximate
aim of this assessment is to check if these vectors are actu- since the relation between x̂(2) − x̂(1) and d̂ R is nonlinear
ally displacements or only the result of random measurement according to (10). Thus, the law of variance propagation
errors. The first step consists in checking the global signifi- does not apply. In this case, the proper cofactor matrix can,
cance test of course, be determined only by the Monte Carlo approach
 T  (e.g., Du Mond and Lenth 1987). Moreover, the assumption
d̂ R Q+R d̂ R that the Ti statistic has Fisher’s distribution is also approxi-
d̂
T = < F (α, u, f ) , (13) mate based on the assumption d̂ R ∼ N . The relation between
u σ̂02
(x̂(2) − x̂(1) ) ∼ N and d̂ R is nonlinear and, therefore, the esti-
where mator d̂ R does not have a normal distribution. The distribu-
 
σ̂02 = v̂(1)
T
P(1) v̂(1) + v̂(2)
T
P(2) v̂(2) f (14) tion of this estimator is not known and the initial assumption
d̂ R ∼ N is only approximate. Consequently, the assumptions
is the pooled variance factor estimator for both epochs and (d̂iR )T Q−1R (d̂iR ) ∼ χ2 and Ti ∼ F are also approximate,
d̂i
v̂(e) = Aδ̂x(e) − (l(e)obs − l0 ), f = f
(1) + f (2) , f (e) = n (e) − although the approximate F tests (13, 15, 16) are accepted.
u (e) + de(e) , α is the significance level, i.e., the probability
of making a Type I error in testing statistical, u = rank(Qd̂ R ) 3 Robust estimation of deformation from observation
and F is the critical value read from Fisher’s cumulative dis- differences
tribution function (Aydin 2013). If the global test passes, this
means that all d̂iR vectors are not statistically significant and, As previously mentioned, classical robust methods for defor-
thus, these points can be regarded as actually stable. In the mation analysis of FCN are based on the results of separate
opposite case, localization of single-point displacements is adjustments of observations for epochs 1 and 2, i.e.,
then performed. The single-point test, neglecting the corre-
obs
lation between the points, is l(1) LS estimation x̂(1)
−−−−−−−−−→
 T  robust S-transformation
d̂iR Q−1R d̂iR −−−−−−−−−−−−−−−−−→ d̂ R
d̂i
Ti = < F (α, u i , f ) , (15) obs
l(2) LS estimation x̂(2)
u i σ̂02 −−−−−−−−−→
The basis for such an approach is the assumption that vector
where u i = rank(Qd̂ R ) is the number of the dimension of the
i d is the difference in functions of independent observations
single-point displacement vector (1 or 2 or 3) (Chen 1983; obs ) − f (lobs )). If the observations are distorted in
(d = f (l(2) (1)
Setan and Singh 2001). Instead of the local test (15), neglect-
both epochs only by random errors, then such an approach is
ing the correlation between the points, one can also apply the
correct, but not if observations are additionally distorted by
local test based on the decorrelated matrix Qd̂ R . More details
constant errors.
are given in Caspary (2000). If the single-point test passes,
First (and most importantly), separate adjustments fulfill
this means that the displacement of point i is not statistically
the condition for the closing of geometric figures, which is not
significant and, thus, this point can be regarded as actually
significant in the deformation analysis process. Such a proce-
stable. In the opposite case, this point can be regarded as
dure, in the case of the occurrence of constant errors, leads to
unstable. The solution (15) can be used for 1-D, 2-D and 3-D
an unjustified increase in the values of the vectors v̂(1) , v̂(2)
networks. Moreover, another condition can be derived from
and, thus, the factor σ̂02 (14) and, consequently, may lead to
the condition (15) for better interpretation of results
 incorrect assessment of the statistical significance of the d̂iR
d̂iR ∈ E = f F, σ̂02 , Qd̂ R , (16) vectors (13), (15) or (16). Second, if the matrices P(1) , P(2)
i
are not proportional in both epochs (e.g., for GNSS vectors),
where E is the confidence interval (1-D network) or the set of vector d̂ R will be distorted by the effect of constant errors.
points formed by the confidence ellipse (2-D network), (e.g., These errors will not cancel out completely in the iterative
Kamiński and Nowel 2013) or the confidence ellipsoid (3-D process (10) because, in this case, their effect on the values

123
754 K. Nowel, W. Kamiński

of the vectors x̂(1) , x̂(2) (1) will be different. As a result, the tion, the solution to the system of Eq. (17) should be a robust
vector d̂ R will be partly distorted by their effect. Third, the estimator of vector d. In the REDOD method, as in the IWST
vector d̂ R and the factor σ̂02 are estimated independently, in method, the L1 -norm estimator of vector d will be used for
two different computational processes. As a result, the vec- this purpose. Moreover, because errors of geodetic observa-
tor d̂ R and the factor σ̂02 have different stochastic properties. tions are of random nature (according to normal distribution),
This discrepancy may lead to incorrect assessment of the the robust estimator of vector d should simultaneously fulfill
statistical significance of the vectors d̂iR . the optimization condition of the LS method for the vector
As mentioned in the introduction, when the displacement of residuals of differences in observations. Consequently, the
vector is determined directly from differences in unadjusted following optimization problem for the system of Eq. (17) is
observations, the above problems do not occur. Therefore, obtained
this paper proposes a new robust alternative method for defor-
mation analysis of FCN (REDOD). The basis of this method
is the assumption that vector d is a function of differences in
independent observations (d = f (l(2) obs − lobs )), i.e., ,
(1)

robust estimation
obs
l(2) − l(1)
obs − −−−−−−−−−−−−−−→ d̂ R

If observations in both epochs are additionally distorted by

constant errors, such an approach can eliminate the effect of
these errors from deformation analysis results. As a result (18)
of estimation, the vector of differences in observations is
where v ∈ n×1 is the vector of residuals of differences in
assigned only such a vector of residuals which enables deter-
observations, Q = (Q(1) + Q(2) ) ∈ n×n is the cofactor
mination of vector d̂ R , but “does not bend” observations to
matrix of differences in observations and P = Q−1  ∈
n×n
fulfill the conditions for the closing of geometric figures.
is the weight matrix of differences in observations.
An unjustified increase in the value of the factor σ̂02 (14)
To find the minimum of the first objective function φ (d),
and distortions of vector d̂ R are avoided this way. Moreover, the classical technique based on the differential calculus can
the vector d̂ R and the factor σ̂02 are estimated in the same be applied. Determining the gradient of the objective func-
computational process and, as a consequence, have the same tion φ (d) and then using the necessary condition for the
stochastic properties. existence of a minimum of this function (∂φ (d) /∂d)T = 0,
The algorithm of the proposed REDOD method consists the following system of normal equations is obtained
of the following two steps.
N d − ω = 0, (19)
3.1 Step 1: robust estimation of the displacement vector 
where N = AT P A, ω = AT P l(2) obs − lobs . The
(1)
It can be seen that if the same geometric network ele- objective function φ (d) is a convex function. Therefore, if
ments have been measured at both analyzed epochs, then the necessary condition is fulfilled, then the sufficient condi-
A(1) = A(2) = A, e.g., for continuous control measurements. tion for a minimum of this function is fulfilled as well. The
In this case, it is possible to construct a certain alternative and obtained system of normal Eq. (19) is, in fact, the set of per-
very natural deformation model, in which the vector of obser- missible solutions (constraints) for vector d, minimizing the
vation differences will be the vector of “observations” and objective function φ (d).
the sought displacement vector will be the vector of para- Continuing the solution of the optimization problem (18),
meters. If interepoch observations, which are described by we find the minimum of the second objective function ψ(d),
the model (2), are subtracted from eachother, we will obtain taking into account the constraints (19). The function ψ(d)
obs − l0 + v
(2) − l(1) − l
0 − v
(1) = is not repeatedly differentiable and its minimum cannot be
for two epochs l(2) obs

Aδx(2) − Aδx(1) and then found using the classical technique, based on differential cal-
  culus, as minimization of this objective function requires the
obs
l(2) − l(1)
obs
+v = Ad, l(2) obs
−l(1)
obs
∼ N (Ad, C ) (17) use of complex linear programming procedures (e.g., simplex
method). For this reason, in this paper, the function ψ(d) was
because δx(2) − δx(1) = d, v = v(2) − v(1) , C ∈ n×n replaced with an alternative, repeatedly differentiable func-
is the covariance matrix of differences in observations (Laz- tion with the form (Kadaj 1988)
zarini et al. 1977). Because the vector of displacement com- u 

ponents of FCN points may have a contaminated normal dis- ψA (d) = di2 + c2 , (20)
tribution with a mean of zero and accepted standard devia- i=1

123
Robust estimation of deformation from observation differences 755

where c is the precision for computations, e.g., 0.0001 m. incomplete rank (4), which is why the matrix M is singular
With the assumption that c is a numerically insignificant and there is no ordinary (classical) inverse M−1 . To determine
value, the component of the function ψA (d) is equivalent to the g-inverse M− , according to g-inverse theory, the matrix
the component of the function ψ(d) and then both functions M ∈ u×u can be reduced to block form
are equivalent in terms of optimization properties. Objective    −1 
M11 M12 M11 0
function ψA (d) is now repeatedly differentiable and its min- M= → M− = , (28)
M21 M22 0 0
imum can be found using the classical technique, based on
differential calculus. According to the optimization theory, where M11 ∈ r ×r is a non-singular matrix with rank r .
−1
the method of Lagrangian multipliers can be applied to find The g-inverse of the matrix M is then the submatrix M11 ,
the minimum of the objective function ψA (d) with the con- complemented with zeros to the dimension  u×u (Rao 1973;
straints (19). According to this method, the objective function Prószyński 1981). Of course, there are also other methods for
ψA (d) with the constraints (19) is replaced by a secondary determination of the g-inverse of this matrix. It can, therefore,
objective function without constraints (Lagrangian function) be seen that we have here the problem of a datum defect
u  for the geodetic network, which has not yet been solved.
L(d) = di2 + c2 − κT (N d − ω ) , (21) To reduce the matrix M to the form (28), it is sufficient to
i=1 represent the matrix A in block form (5). Therefore, in this
case, the problem of a datum defect for the geodetic network
where κ ∈ u×1 is the vector of Lagrangian multipliers. The
can be solved analogously as in the classical method, only in
necessary conditions for the existence of a minimum of the
this case it is done simultaneously for both epochs. Moreover,
objective function ψA (d) with the constraints (19) then have
let us recall that the same optimization conditions were also
the following form (Kuhn–Tucker necessary conditions)
proposed in both methods. It can, therefore, be seen that both
(∂ L (d) /∂d)T = 0 (22) methods differ only in the functional models, i.e., (9) and
κ (N d − ω ) = 0
T
(23) (18), while the other theoretical assumptions are common.
Returning to the computations, after substitution of the block
The objective function ψA (d) is a convex function. There- matrix A (5) into the matrix M, we obtain
fore, if the above necessary conditions are fulfilled, the suffi-  
(Q11 W1−1 Q11 + Q12 W2−1 QT12 )−1 0
cient conditions for a minimum of this function are fulfilled M− = , (29)
as well. Determining the gradient of the function L(d) and 0 0
then using the necessary condition (22), the following equa- where Q11 = AT1 P A1 , Q12 = AT1 P A2 and W =
tion is obtained diag(W1 , W2 ). After substitution of the matrix (29) into the
Wd̂ − N κ̂ = 0 (24) system of Eq. (27), we finally obtain
 −1 
where Wd̂ − DT DW−1 DT A1T P l(2)
obs
− l(1)
obs
=0 (30)
 

W = diag . . . , 1 d̂i2 + c2 , . . . (25) where D = [ AT1 P A1 AT1 P A2 ]. We compute on the basis
of the system of Eq. (30) the robust estimator of vector d
and N T = N , WT = W, because N and W are symmet-
  and its cofactor matrix. The computations are carried out in
rical matrices. The second necessary condition (23) can be an iterative cycle with the following formula (approximate
replaced by the condition (19), (for κ̂ = 0). Finally, after Newtonian process)
determination of vector d̂ from Eq. (24), the necessary con-  ⎫
dition (23) will assume the following form (system of normal d̂(k) = R(k) l(2)
obs − lobs
(1) ⎪
⎪
equations of correlates) (k) (k)


(k) T ⎬
Q =R Q(1) + Q(2) R (31)
d̂  ⎪
⎪
Mκ̂ − ω = 0, (26) W(k+1) = diag . . . , w
(k+1)
, ... ⎭
i k=1,2,...
where M = N W−1 N . If vector κ̂ solves the system of Eq.
−1 T 
(k) −1 T −1 T
(26), vector d̂ minimizes the objective function ψA (d) with where R(k) = W(k) D D W D A1 P is
the constraints (19). Therefore, after determination of vector the matrix of transformation of the vector of differences in
κ̂ from Eq. (26) and its substitution into Eq. (24), we finally observations into the displacement vector, W(k=1) = I and
obtain  
  (k) 2  
w
(k+1)
=1 + c2 =∼ 1 d̂ (k)  + c
Wd̂ − N M− ω = 0, (27) i d̂i i (32)

where M− is a g-inverse, such that MM− M = M. In the It should be stressed that there still remains problem of outlier
considered case of an FCN, A is the matrix of the columnarly detection (observations), which has not yet been solved. In

123
756 K. Nowel, W. Kamiński

IWST REDOD
this case, this problem can be solved, e.g., on the basis of
Baarda’s data snooping or Pope’s test, only in this case it is Deformation measurements l obs obs
(1) , l (2)

done on the basis of the vector of residuals of differences in

Estimation (1) Estimation (31)
observations from the first iterative step (31). Unfortunately, xˆ (1) , Q xˆ (1) , vˆ (1) dˆ ( k =1) , Qd(ˆk =1) , vˆ (Δk =1)
in this case, robust estimation “does not work” for detecting xˆ (2) , Q xˆ ( 2 ) , vˆ (2)
outliers (observations).
The iterative process
 (31) is finished
 when all components Outlier test Outlier test
 
of the difference d̂(k+1) − d̂(k)  are lower than the adopted
no Observations Observations no
precision for computations c (convergence). The final robust clean? clean?
estimator of the vector of displacement components for all yes yes
analyzed points in an FCN, i.e., vector d̂ R and the final cofac- Transformation (10)
tor matrix of this vector, i.e., matrix Qd̂ R , is obtained from dˆ ( k =1) , Qd(ˆk =1)

Robust S-transformation
the last iterative step.
Reweighting (12) Reweighting (32)

Robust estimation
wi( k +1) = 1/ ( dˆi( k ) + c) wi( k +1) = 1/ ( dˆi( k ) + c)
3.2 Step 2: assessment of the significance of the
displacement vector Transformation (10) Estimation (31)
dˆ ( k = k +1) , Qd(ˆk = k +1) dˆ ( k = k +1) , Qd(ˆk = k +1)

Assessment of the significance of the vectors d̂iR can of course

no no
be performed analogously as in the classical method, i.e., on Convergence? Convergence?
the basis of the approximate global (13) and local (15) or (16) yes yes
F tests. For the stochastic model of the REDOD method (18), Final results (10), (14) Final results (31), (33)
the variance factor estimator has the following form dˆ R , Qdˆ R , σˆ 02 dˆ R , Qdˆ R , σˆ 02

σ̂02 = v̂
T
P v̂ /(n − u + de) (33) Global significance test (13)
T<F ?
yes no
All points can be regarded
as actually stable
Local significance test (15) or (16)
4 Comparison of the classical IWST method Ti < F ? or dˆ iR ∈ E ?

with the proposed REDOD method yes no

Point i can be regarded Point i can be regarded
as actually stable as actually displaced
As previously mentioned, classical deformation analysis of
FCN is most often performed by robust methods. The best Fig. 1 Block diagram of the algorithms of the IWST and REDOD
methods
known methods from this group include the IWST method.
This method was used, among others, for deformation analy- L1 -norm (9, 18), or the probabilistic model for the obser-
sis of the Tevatron atomic particle accelerator complex at the vations, i.e., normal distribution (2, 17).
Fermilab laboratory in the US and for deformation analysis The theoretical principles of both methods differ only in
of the world’s largest strip copper mine in Chile. The IWST the functional deformation models, i.e., (9) and (18). Gener-
method has also been implemented in the automated ALERT ally, the functional models (9) and (18) assume that the dis-
monitoring system developed by the Canadian Centre for placement vector is a function of differences in adjusted coor-
Geodetic Engineering and in the universal GeoLab geodetic dinates and differences in unadjusted observations, respec-
computation software. Section 2 describes the algorithm of tively. It is worth noting once more that the functional model
classical robust methods. Particular attention was paid to the (18) has some advantages which the functional model (9)
IWST method. In Sect. 3, a new robust deformation analy- does not have, namely, the functional model (18) accepts
sis method was proposed based on classical robust methods. measurement errors constant in both epochs, which can be
This method was conventionally called REDOD. of any magnitude and may be different for individual geo-
The theoretical principles of the proposed REDOD met- metric elements of the network. Moreover, unlike model (9),
hod are very similar to the theoretical principles of the clas- model (18) ensures stochastic consistency between the dis-
sical IWST method. Both methods have in common, among placement vector estimator and the variance factor estimator.
others, the definition of the datum for the geodetic network, This results from the fact that both estimators are estimated
i.e., the zero-variance computational base (5), the optimiza- here in the same computational process.
tion condition for the displacement vector (the definition of As a consequence of using different functional models, the
the datum for the displacement vector), i.e., the minimal aithms of the IWST and REDOD methods differ significantly.

123
Robust estimation of deformation from observation differences 757

In Sects. 2 and 3, these algorithms were derived and described Table 1 Simulated observations in epoch 1 and epoch 2
in detail. Only their basic aspects are shown below, in the Observation designation Variant (m)
form of a block diagram (Fig. 1).
b e 1 2
obs
l(1) obs
l(2) obs
l(1) obs
l(2)
5 Numerical tests
A B −0.003 0.014 −0.001 0.016
Numerical tests were performed on simulated FCN. The tests B C −0.004 −0.011 −0.002 −0.009
compared deformation analysis results obtained from the C A −0.001 0.006 0.001 0.008
IWST and REDOD methods. A D 0.002 −0.005 0.004 −0.003
D C −0.001 0.004 0.001 0.006
5.1 Example 1: deformation analysis of the leveling FCN B D 0.006 −0.009 0.008 −0.007
b beginning, e end
This test was performed on the example of a leveling network
of four points A, B, C, D (Fig. 2). Computations were carried
method], i.e., the submatrix A2 consisted of the de = 1 last
out for two variants of simulated observations.
column in the matrix A, as was proposed in the theoretical
First, the vector of theoretical heights of controlled points
part of this paper (5). Standard deviations of simulated errors
in epoch 1 x(1) = [0.000, 0.000, 0.000, 0.000]T m, the
of height difference measurement were assumed as a priori
vector of theoretical displacements of controlled points d =
mean errors of height difference measurement. The values of
[0, 10, 0, 0]T mm and the vector of theoretical heights of
a priori mean errors were the basis for determining the matri-
controlled points in epoch 2 x(2) = [0.000, 0.010, 0.000,
ces P(1) , P(2) (IWST) and P (REDOD). Outlier detection
0.000]T m were assumed. Theoretical height differences in
was carried out using Baarda’s method. The critical values
both epochs were computed on the basis of the vectors x(1) 2 (3)/3 = 2.61 and t
were χ0.05 α0 /2 = 2.51 (β0 = 0.20) for the
and x(2) .
global and local test, respectively. In variant 1, in the IWST
In variant 1, the theoretical height differences were dis-
method, the global test passed. In variant 2, in the IWST
torted with random errors according to the normal distribu-
method, the global test failed in epoch 2 (σ̂0(2) 2 = 2.89 >
tion and the significance level α = 0.05. The same standard
2.61), but the local test passed for all standardized residu-
deviations of simulated errors of height difference measure-
als. In the REDOD method, the global test passed in both
ment σh = 4 mm were assumed in both epochs.
variants. In the IWST method, the test on the variance ratio
In variant 2, the theoretical height differences were dis-
passed in both variants [critical value F(0.05, 3, 3) = 9.28].
torted with random errors from variant 1 and additionally
The results of the deformation analysis are presented in
with constant errors εh = +2 mm.
Table 2.
The variants of simulated observations are presented in
In variant 1, both methods give the same, correct results.
Table 1. Carrying out in succession the computations (10),
In both methods, point B is considered unstable. In variant 2,
(14), (13) and (15), complete results were obtained using
in the IWST method, the value of vector d̂ R is the same as in
IWST method. Carrying out in succession the computations
variant 1. Therefore, constant errors have no effect on vector
(31), (33), (13) and (15) complete results were obtained
d̂ R . However, the results of the assessment of significance
using REDOD method. In both methods, constant heights
are different than in variant 1. This is caused by the effect
z D were assumed to eliminate the network’s datum defect
of constant errors. This effect was not eliminated from the
[Eq. (3) in the classical method and Eq. (27) in the proposed
values of the vectors v̂(1) , v̂(2) . This led to an increase in
the value of σ̂02 (14), underestimation of the values of T
(13), Ti (15) and, consequently, erroneous assessment of the
significance of vector d̂ BR . In the REDOD method, the effect
of constant errors is eliminated. Both the value of vector d̂ R
and the results of the assessment of significance are the same
as in variant 1.

5.2 Example 2: deformation analysis of the horizontal FCN

This test was performed on the example of a horizontal net-

work of six points A, B, C, D, E, F (Fig. 3). Computations
Fig. 2 Leveling FCN were carried out for four variants of simulated observations.

123
758 K. Nowel, W. Kamiński

Table 2 The results of the deformation analysis of the leveling FCN

Variant Point IWST REDOD

   
d̂iR = d̂z iR Qd̂ R = Q d̂z R d̂iR = d̂z iR Qd̂ R = Q d̂z R
σ̂02 i i T Ti σ̂02 i i T Ti
(mm) (m) (mm) (m)

1 A 0.5 1.01 0.0014 6.03a 0.12 0.5 1.05 0.0014 5.79a 0.11
B 11.0 0.0033 10.61a 11.0 0.0033 10.19a
C −1.0 0.0026 0.15 −1.0 0.0026 0.14
D −5.5 0.0033 2.79 −5.5 0.0033 2.68
2 A 0.5 1.88 0.0014 3.23 0.06 0.5 1.05 0.0014 5.79a 0.11
B 11.0 0.0033 5.68 11.0 0.0033 10.19a
C −1.0 0.0026 0.08 −1.0 0.0026 0.14
D −5.5 0.0033 1.50 −5.5 0.0033 2.68
a The test statistic exceeds the critical value for the local F(0.05, 1, 6) = 5.99 or global F(0.05, 3, 6) = 4.76 test

of the vectors x(1) , x(2) . The theoretical observations were

then distorted with random errors according to the normal
distribution and the significance level α = 0.05. However,
different standard deviations of simulated errors of angle
measurement were assumed in both epochs in this case. In
epoch 1, the standard deviation of angle measurement errors
at central points A, B, C σβ = 15cc , at central points D,
E, F σβ = 5cc and the standard deviation of distance mea-
surement errors σs = 3 mm were assumed. In epoch 2, the
standard deviation of angle measurement errors at central
points A, B, C σβ = 5cc , at central points D, E, F σβ = 15cc
and the standard deviation of distance measurement errors
σs = 3 mm were assumed.
In variant 3, observations were distorted with random
errors and constant errors. The vector εx = [3, 3, 3, 3, 3, 3, 3,
Fig. 3 Horizontal FCN 3, 3, 3, 3, 3]T mm was added to the vectors x(1) , x(2) and the
vectors of eccentric coordinates of the target x(1) ε , xε were
(2)
First, the vector of theoretical coordinates of controlled obtained. The eccentric observations were then computed
points in epoch 1 x(1) = [350.000, 200.000, 300.000, 300.000, on the basis of the theoretical coordinates of the instrument
ε , xε .
x(1) , x(2) and the eccentric coordinates of the target x(1)
200.000, 300.000, 150.000, 200.000, 200.000, 100.000, (2)
300.000, 100.000]T m, the vector of theoretical displace- In other words, observations distorted by constant errors in
ments of controlled points d = [5, 5, 0, 0, 0, 10, 0, 0, 0, 0, the form of eccentricity of the target in relation to the instru-
0, 0]T mm and the vector of theoretical coordinates of con- ment with the value of εx = +3 mm along the axes X and Y
trolled points in epoch 2 x(2) = [350.005, 200.005, 300.000, were obtained this way. These observations were then addi-
300.000, 200.000, 300.010, 150.000, 200.000, 200.000, tionally distorted with random errors from variant 1.
100.000, 300.000, 100.000]T m were assumed. In variant 4, observations were distorted with constant
In variant 1, observations were distorted with random errors from variant 3 and random errors from variant 2.
errors. Theoretical observations were computed on the basis The variants of simulated observations are presented in
of the vectors x(1) , x(2) . The theoretical observations were Table 3.
then distorted with random errors according to the normal Complete results were obtained in the IWST method by
distribution and the significance level α = 0.05. The same carrying out the computations (10), (14), (13) and (16) in
standard deviations of simulated errors of angle measure- succession. Complete results were obtained in the REDOD
ment σβ = 10cc and distance measurement σs = 3 mm method by carrying out the computations (31), (33), (13)
were assumed in both epochs. and (16) in succession. In both methods, constant coordi-
In variant 2, observations were also distorted with random nates yE , xF , yF were assumed to eliminate the network’s
errors. Theoretical observations were computed on the basis datum defect (Eq. (3) in the classical method and Eq. (27)

123
Robust estimation of deformation from observation differences 759

Table 3 Simulated observations in epoch 1 and epoch 2

Observation designation Variant (g), (m)
c l r 1 2 3 4
obs
l(1) obs
l(2) obs
l(1) obs
l(2) obs
l(1) obs
l(2) obs
l(1) obs
l(2)

A B C 33.0507 33.0457 33.0510 33.0455 33.0515 33.0465 33.0518 33.0463

A C D 37.4338 37.4364 37.4344 37.4360 37.4343 37.4369 37.4349 37.4364
A D E 37.4354 37.4320 37.4354 37.4329 37.4361 37.4327 37.4361 37.4336
A E F 33.0503 33.0473 33.0500 33.0477 33.0513 33.0484 33.0510 33.0488
B C D 37.4328 37.4402 37.4332 37.4404 37.4344 37.4418 37.4348 37.4420
B D E 33.0501 33.0488 33.0507 33.0495 33.0507 33.0494 33.0513 33.0501
B E A 59.0319 59.0370 59.0332 59.0368 59.0339 59.0389 59.0352 59.0387
C D F 59.0341 59.0289 59.0340 59.0292 59.0345 59.0293 59.0344 59.0296
C F A 33.0512 33.0513 33.0488 33.0496 33.0515 33.0516 33.0491 33.0499
C A B 37.4351 37.4277 37.4333 37.4282 37.4355 37.4282 37.4337 37.4287
D E F 33.0504 33.0522 33.0498 33.0492 33.0495 33.0513 33.0489 33.0483
D F A 37.4327 37.4338 37.4338 37.4356 37.4322 37.4333 37.4333 37.4351
D A B 37.4359 37.4326 37.4335 37.4303 37.4352 37.4320 37.4328 37.4297
D B C 33.0510 33.0527 33.0508 33.0509 33.0499 33.0517 33.0497 33.0499
E F A 37.4333 37.4326 37.4331 37.4332 37.4317 37.4310 37.4315 37.4316
E A B 33.0491 33.0508 33.0497 33.0512 33.0484 33.0501 33.0490 33.0505
E B D 59.0337 59.0331 59.0338 59.0325 59.0318 59.0312 59.0319 59.0306
F A C 59.0316 59.0332 59.0335 59.0310 59.0313 59.0329 59.0333 59.0307
F C D 33.0509 33.0512 33.0500 33.0528 33.0505 33.0509 33.0496 33.0525
F D E 37.4343 37.4342 37.4336 37.4339 37.4339 37.4338 37.4332 37.4335
– A B 111.804 111.807 111.801 111.801 111.806 111.809 111.803 111.803
– B C 100.007 99.993 99.993 99.997 100.004 99.990 99.990 99.994
– C D 111.801 111.811 111.800 111.807 111.797 111.807 111.796 111.803
– D E 111.798 111.802 111.802 111.804 111.797 111.801 111.801 111.803
– E F 100.001 100.001 99.999 99.997 100.004 100.004 100.002 100.000
– F A 111.804 111.812 111.805 111.809 111.808 111.816 111.809 111.813
c central, l left, r right

in the proposed method), i.e., the submatrix A2 consisted (σ̂0(1)

2 = 2.98 > 1.62, σ̂0(2)
2 = 3.42 > 1.62) and the local
of the de = 3 last columns in the matrix A, as was pro- test failed in both epochs. The largest standardized resid-
posed in the theoretical part of this paper (5). The theo- uals of epochs 1 and 2 were v̂DEF /σv̂DEF = 4.51 > 3.63
retical coordinates from epoch 1 were assumed as approxi- and v̂ABC /σv̂ABC = 5.36 > 3.63, respectively. These obser-
mate coordinates. Standard deviations of simulated errors of vations were deleted. The critical values at that point were
angle and distance measurement were assumed as a priori 2 (16)/16 = 1.64 and t
χ0.05 α0 /2 = 3.58 (β0 = 0.20) for
mean errors of angle and distance measurement. The val- the global and local test, respectively. In the IWST method,
ues of a priori mean errors were the basis for determining the the global test failed again in both epochs (σ̂0(1)
2 = 1.98 >
matrices P(1) , P(2) (IWST) and P (REDOD). Outlier detec- 1.64, σ̂0(2) = 1.92 > 1.64), but the local test passed now
2
tion was carried out by Baarda’s method. The critical values for all standardized residuals. In the REDOD method, the
2 (17)/17 = 1.62 and t
were χ0.05 α0 /2 = 3.63 (β0 = 0.20) global test passed in all variants. In the IWST method, the
for the global and local test, respectively. The global test test on the variance ratio passed in all variants [critical value
passed in variants 1 and 2 in the IWST method, but the global F(0.05, 17, 17) = 2.27].
test failed in both epochs in variant 3 in the IWST method The results of deformation analysis are presented in
(σ̂0(1)
2 = 2.51 > 1.62, σ̂0(2)
2 = 2.10 > 1.62), although the Table 4.
local test passed for all standardized residuals. In variant 4, In variant 1, both methods yield the same results (within
in the IWST method, the global test failed in both epochs the limits of accuracy of computations).

123
760 K. Nowel, W. Kamiński

Table 4 The results of deformation analysis of the horizontal FCN

Variant Point IWST REDOD
 
Qd̂ R = Qd̂ R =
d̂ R = ⎡ 
i ⎤ d̂ R = ⎡ 
i ⎤
 i R  i R
d̂ xi Q d̂ x R Q d̂ x R d̂ y R d̂ xi Q d̂ x R Q d̂ x R d̂ y R
σ̂02 ⎣ i  i i ⎦ E= σ̂02 ⎣ i  i i ⎦ E=
d̂ yiR Q d̂ y R d̂ x R Q d̂ y R d̂ yiR Q d̂ y R d̂ x R Q d̂ y R
(mm) i i i (mm) i i i
(m) (m)
ai ai
ϕi ϕi
bi bi
(g) (g)
(mm) (mm)

1 A 7.1 1.13 0.0025 0.0010 6.9 16 7.1 1.10 0.0025 0.0011 6.9 18
4.1 0.0010 0.0016 4.1 4.0 0.0011 0.0016 4.1
B −2.4 0.0021 0.0017 7.8 −44 −2.4 0.0021 0.0017 7.8 −41
−0.6 0.0017 0.0024 4.0 −0.7 0.0017 0.0025 4.1
C 1.0 0.0011 0.0001 7.3 0 1.0 0.0011 −0.0004 7.3 2
10.2 0.0001 0.0027 3.1 10.1 −0.0004 0.0027 3.0
D 0.4 0.0013 −0.0011 5.3 36 0.4 0.0014 −0.0011 5.2 36
−1.3 −0.0011 0.0017 2.8 −1.4 −0.0011 0.0017 2.8
E −0.8 0.0012 0.0007 3.9 48 −0.8 0.0013 0.0007 3.9 46
−0.7 0.0007 0.0012 2.7 −0.7 0.0007 0.0012 2.7
F −0.8 0.0016 −0.0011 4.9 −35 −0.7 0.0016 −0.0011 4.9 −35
0.2 −0.0011 0.0012 2.1 0.2 −0.0011 0.0011 2.0
2 A 4.8 0.73 0.0026 −0.0015 6.1 −28 5.1 0.74 0.0028 −0.0008 6.2 −10
2.9 −0.0015 0.0017 2.9 3.7 −0.0008 0.0019 4.1
B −0.3 0.0022 0.0013 5.2 23 −0.2 0.0018 0.0013 4.8 41
0.1 0.0013 0.0010 1.4 0.2 0.0013 0.0015 2.1
C 1.6 0.0013 0.0008 3.9 −30 2.1 0.0018 0.0006 4.3 −36
10.1 0.0008 0.0016 2.5 10.4 0.0006 0.0019 3.8
D 0.0 0.0010 −0.0003 2.3 −23 0.0 0.0012 −0.0004 3.6 7
−0.1 −0.0003 0.0009 1.9 −0.4 −0.0004 0.0016 2.6
E −0.1 0.0011 0.0011 5.1 −17 0.0 0.0008 0.0003 5.1 −1
−1.7 0.0011 0.0022 2.0 −1.1 0.0003 0.0023 1.8
F −0.2 0.0015 −0.0016 6.2 28 −0.9 0.0024 −0.0016 6.4 48
−1.7 −0.0016 0.0026 2.4 −1.4 −0.0016 0.0024 4.0
3 A 7.1 2.30 0.0025 0.0010 9.9 16 7.1 1.10 0.0025 0.0011 6.9 18
4.1 0.0010 0.0016 5.9 4.0 0.0011 0.0016 4.1
B −2.4 0.0021 0.0017 11.2 −44 −2.4 0.0021 0.0017 7.8 −41
−0.6 0.0017 0.0024 5.8 −0.7 0.0017 0.0025 4.1
C 1.0 0.0011 0.0001 10.5 0 1.0 0.0011 −0.0004 7.3 2
10.2 0.0001 0.0027 4.5 10.1 −0.0004 0.0027 3.0
D 0.4 0.0013 −0.0011 7.6 36 0.4 0.0014 −0.0011 5.2 36
−1.3 −0.0011 0.0017 4.0 −1.4 −0.0011 0.0017 2.8
E −0.8 0.0012 0.0007 5.6 48 −0.8 0.0013 0.0007 3.9 46
−0.7 0.0007 0.0012 3.9 −0.7 0.0007 0.0012 2.7
F −0.8 0.0016 −0.0011 7.1 −35 −0.7 0.0016 −0.0011 4.9 −35
0.2 −0.0011 0.0012 3.0 0.2 −0.0011 0.0011 2.0
4 A 3.5 1.95 0.0022 −0.0017 10.0 −46 5.1 0.74 0.0028 −0.0008 6.2 −10
1.4 −0.0017 0.0021 4.5 3.7 −0.0008 0.0019 4.1
B −6.2 0.0028 0.0010 10.0 9 −0.2 0.0018 0.0013 4.8 41
0.0 0.0010 0.0005 1.1 0.2 0.0013 0.0015 2.1

123
Robust estimation of deformation from observation differences 761

Table 4 continued
Variant Point IWST REDOD
 
Qd̂ R = Qd̂ R =
d̂ R = ⎡ 
i ⎤ d̂ R = ⎡ 
i ⎤
 i R  i R
d̂ xi Q d̂ x R Q d̂ x R d̂ y R d̂ xi Q d̂ x R Q d̂ x R d̂ y R
σ̂02 ⎣ i  i i ⎦ E= σ̂02 ⎣ i  i i ⎦ E=
d̂ yiR Q d̂ y R d̂ x R Q d̂ y R d̂ yiR Q d̂ y R d̂ x R Q d̂ y R
(mm) i i i (mm) i i i
(m) (m)
ai ai
ϕi ϕi
bi bi
(g) (g)
(mm) (mm)

C 1.3 0.0013 0.0009 6.4 −31 2.1 0.0018 0.0006 4.3 −36
11.8 0.0009 0.0016 4.0 10.4 0.0006 0.0019 3.8
D −1.0 0.0014 0.0000 4.9 0 0.0 0.0012 −0.0004 3.6 7
0.1 0.0000 0.0002 0.9 −0.4 −0.0004 0.0016 2.6
E 2.3 0.0021 0.0014 9.4 −43 0.0 0.0008 0.0003 5.1 −1
−2.8 0.0014 0.0023 6.2 −1.1 0.0003 0.0023 1.8
F −1.8 0.0015 −0.0010 10.7 10 −0.9 0.0024 −0.0016 6.4 48
−4.5 −0.0010 0.0029 5.2 −1.4 −0.0016 0.0024 4.0
In both methods, the global significance test of the displacement vector T (13) failed for all variants

In variant 2, the values of vector d̂ R and the parameters constant errors on the value of σ̂02 . In the REDOD method,
of confidence ellipses differ in both methods. This is caused the effect of constant errors is eliminated. Both the values
by disproportionateness of the matrices P(1) , P(2) (IWST) of vector d̂ R and the values of the parameters of confidence
to the matrix P (REDOD). However, these differences are ellipses are the same as in variant 2.
not significant. Moreover, the sum of absolute values of true Figure 4 presents graphical assessment of the significance
errors for the components
 of vector d̂ R is the same in both of vector d̂ R (16).
 R 
methods, d̂i − di  = 8.2 mm. In the case of the occurrence of only random errors (Fig. 4,
variants 1, 2), the results of assessment of significance are
In variant 3, in the IWST method, the value of vector d̂ R
correct in both methods. The displacements of points A and
is the same as in variant 1. Therefore, constant errors have no
C are considered significant in both variants.
effect on vector d̂ R . However, the values of the parameters
In the case of the occurrence of random errors and con-
of confidence ellipses differ significantly from the values in
stant errors (Fig. 4, variants 3, 4), the results of assessment
variant 1. This is caused by the effect of constant errors.
of significance in the IWST method are not correct. Only
This effect was not eliminated from the values of the vectors
the displacement of point C is considered significant in both
v̂(1) , v̂(2) . This led to an increase in the value of σ̂02 (14) and,
variants. This is caused by the effect of constant errors. In the
consequently, to an increase in the values of the parameters
REDOD method, the results of assessment of significance are
of confidence ellipses (16). In the REDOD method, the effect
correct. The displacements of points A and C are considered
of constant errors is eliminated. Both the values of vector d̂ R
significant in both variants.
and the values of the parameters of confidence ellipses are
the same as in variant 1.
In variant 4, in the IWST method, the value of vector d̂ R
6 Summary and conclusions
differs significantly from the value in variant 2. The sum of
absolute values of true errors for the values of vector d̂ R is sig-
  R 
 Deformation measurements have a repeatable nature. In other
nificantly higher than in variant 2, d̂i − di  = 26.9 mm. words, deformation measurements are performed at differ-
This is caused by the disproportion of matrix P(1) to matrix ent moments of time, but often with the same equipment,
P(2) . This in turn led to a disproportionate effect of constant methods, geometric conditions and in a similar environment
errors on the values of the vectors x̂(1) , x̂(2) (1) and, con- in epochs 1 and 2. Moreover, the same instrument and tar-
sequently, prevented complete reduction of these errors in get are most often set on given points in both epochs. It is,
the iterative process (10). The values of the parameters of therefore, reasonable to assume that the results of deforma-
confidence ellipses also differ significantly from the values tion measurements can be distorted both by random errors
in variant 2. As in variant 3, this is caused by the effect of and by some non-random errors, which are constant in both

123
762 K. Nowel, W. Kamiński

Fig. 4 The results of

assessment of the significance of
displacements in the case of the
occurrence of random errors
(variants 1, 2) and in the case of
the occurrence of random and
constant errors (variants 3, 4)

epochs, i.e., errors with the same value and the same sign 1. If the results of deformation measurements were addi-
in both epochs. There may be constant errors which can- tionally distorted by constant errors, the REDOD method
not be eliminated from measurement results, e.g., incom- completely eliminated their effect from deformation
pletely reduced atmospheric refraction, eccentricity of the analysis results,
target in relation to the instrument (caused by faulty exe- 2. If the results of deformation measurements were distorted
cution or assembly, non-vertical mounting of the centering only by random errors, the REDOD method yielded very
sleeve in the instrument pillar or inclination of the instrument similar deformation analysis results as the classical IWST
pillar with time). There may also be constant errors which can method.
be eliminated from measurement results although, for exam-
ple, economic and technical limitations of the measurement The REDOD method was developed for FCN, which has
project prevent their elimination (Caspary 2000). For exam- the same observation structure in both epochs. That is, the
ple, a pillar could have a GNSS mounted, as well as a target same geometric elements must be measured in the control
prism that RTS could observe (Fig. 5). In this case, the target network in both epochs. For example, this condition is always
prism and GNSS receiver cannot define the same physical met for automated, continuous control measurements. As
point. previously mentioned, there is a high risk that the error of
For FCN, the results of deformation measurements are measurement of the same geometric element of the network
most often processed using the robust methods. Classical in both epochs may contain a factor constant in both epochs
robust methods do not completely eliminate the effect of con- and, therefore, the REDOD method is strongly recommended
stant errors and, therefore, lead to reduced external reliability in this case.
level of processing and, consequently, to reduced quality of It should be noted, however, that sometimes, in practice,
deformation analysis results. control networks have a different observation structure in
This paper proposes a new robust alternative method, both epochs. For example, the classical technique in epoch 1
which was called REDOD. The performed numerical tests can be replaced with the GNSS technique in epoch 2. In this
showed that: case, the REDOD method can also be used. The input data for

123
Robust estimation of deformation from observation differences 763

Acknowledgments The authors wish to thank the editors and three

anonymous reviewers for their time, effort and help in improving the
manuscript.

Open Access This article is distributed under the terms of the Creative
Commons Attribution License which permits any use, distribution, and
reproduction in any medium, provided the original author(s) and the
source are credited.

References
Aydin C (2013) Power of global test in deformation analysis. J Surv
Eng 138(2):51–55
Caspary WF (1984) Deformation analysis using a special similarity
transformation. In: FIG Int Eng Surv Conf, Washington DC, pp 145–
151
Caspary WF, Borutta H (1987) Robust estimation in deformation mod-
els. Surv Rev 29(223):29–45
Caspary WF, Haen W, Borutta H (1990) Deformation analysis by sta-
tistical methods. Technometrics 39(1):49–57
Caspary WF (2000) Concepts of network and deformation analysis. The
University of New South Wales, Kensington
Cederholm P (2003) Deformation analysis using confidence ellipsoids.
Surv Rev 37(287):31–45
Fig. 5 Integrated deformation measurements (RTS, target prism and Chen YQ (1983) Analysis of deformation surveys- a generalized
GNSS receiver) for network of potential reference points (FCN) method. Technical Report No. 94 University of New Brunswick,
Fredericton, pp 54–72
Chen YG, Chrzanowski A, Secord JM (1990) A strategy for the analysis
of the stability of reference points in deformation surveys. CISM J
the REDOD method will then be pseudo-observations of the ACSGG 44(2):141–149
same geometric elements of the network in both epochs (e.g., Chrzanowski A, Chen YQ (1990) Deformation monitoring, analysis and
prediction-status report FIG XIX international congress. Helsinki
vectors). Such pseudo-observations can be computed based 6(604.1):83–97
on preadjusted coordinates. However, it should be noted that Denli HH, Deniz R (2003) Global congruency test methods for GPS
for networks which have different observation structure in networks. J Surv Eng 129(3):95–98
both epochs, there is a very low risk that error constant in both Denli HH (2008) Stable point research on deformation networks. Surv
Rev 40(307):74–82
epochs will occur and, therefore, in this case, the REDOD Du Mond CE, Lenth RV (1987) A robust confidence interval for loca-
method is not strongly recommended. As previously demon- tion. Technometrics 29(2):211–219
strated, if there are no error constant in both epochs, the Duchnowski R (2010) Median-based estimates and their application in
REDOD and IWST methods give much the same quality of controlling reference mark stability. J Surv Eng 136(2):47–52
Duchnowski R, Wiśniewski Z (2012) Estimation of the shift between
deformation analysis results. parameters of functional models of geodetic observations by apply-
he obtained results encourage further research into the lat- ing Msplit estimation. J Surv Eng 138(1):1–8
est robust solutions for deformation estimation from observa- Duchnowski R (2013) Hodges–Lehmann estimates in deformation
tion differences. The method of sign-constrained robust least analyses. J Geod 87(10–12):873–884
Even-Tzur G (2011) Deformation analysis by means of extended free
squares, which detects over 50 % of outliers (Xu 2005), and network adjustment constraints. J Surv Eng 137(2):47–52
robust estimation by expectation maximization algorithm, Gökalp E, Tasçi L (2009) Deformation monitoring by GPS at embank-
which is more sensitive to outliers than classical methods ment dams and deformation analysis. Surv Rev 41(311):86–102
(Koch 2013; Koch and Kargoll 2013), may be particularly Hekimoglu S, Koch KR (1999) How can reliability of robust methods
be measured? Third Turkish-German Joint Geodetic Days, Istanbul
useful. Moreover, new numerical tests should be performed 1:179–196
for new approaches, based on many thousands of simulated Kadaj R (1988) Eine Klasse von Schätzverfahren mit praktischen
data sets (e.g., Monte Carlo simulations). Not only measure- Anwendungen. Zeitschrift fur Vermessungswesen H8(117J):157–
ment errors should be randomized in individual sets, but also 165
Kamiński W, Nowel K (2013) Local variance factors in deforma-
the number and location of displaced points and the values tion analysis of non-homogenous monitoring networks. Surv Rev
and directions of these points’ displacements. The measure 45(328):44–50
of accuracy of the estimated displacements could be empir- Koch KR (2013) Robust estimation by expectation maximization algo-
ical standard deviation and the measure of significance test rithm. J Geod 87(2):107–116
Koch KR, Kargoll B (2013) Expectation maximization algorithm for
efficiency could be the mean success rate (MSR) (Hekimoglu the variance-inflation model by applying the t-distribution. J Appl
and Koch 1999). Geod 7:217–225

123
764 K. Nowel, W. Kamiński

Lazzarini T, Laudyn I, Chrzanowski A, Gaździcki J, Janusz W, Wiłun Setan H, Othman R (2006) Monitoring of offshore platform subsidence
Z, Mayzel B, Mikucki Z (1977) Geodetic measurements of displace- using permanent GPS stations. J Glob Position Syst 5(1–2):17–21
ments of structures and their surroundings. PPWK, Warsaw (in Pol- Setan H, Singh R (2001) Deformation analysis of a geodetic monitoring
ish) network. Geomatica 55(3):333–346
Marx C (2013) On resistant L p -norm estimation by means of iteratively Wilkins R, Bastin G, Chrzanowski A (2003) ALERT: a fully automated
reweighted least squares. J Appl Geod 7:1–10 real time monitoring system. In: FIG XI symposium on deformation
Niemeier W (1981) Statistical tests for detecting movements in repeat- measurements, Greece
edly measured geodetic networks. Tectonophysics 71(1981):335– Wiśniewski Z (2009) Estimation of parameters in a split func-
351 tional model of geodetic observations (Msplit estimation). J Geod
Prószyński W (1981) Introduction into applications of the generalized 83(2):105–120
Moore–Penrose inverse to chosen adjustment problems. Geod Car- Xu PL (2005) Sign-constrained robust least squares, subjective break-
togr 4(2):123–130 down point and the effect of weights of observations on robustness.
Prószyński W (2010) Another approach to reliability measures for sys- J Geod 79(1–3):146–159
tems with correlated observations. J Geod 84(9):547–556
Rao RC (1973) Linear statistical inference and its applications. Wiley,
New York

123