Hammer et al.

Earth, Planets and Space (2021) 73:54

https://doi.org/10.1186/s40623-021-01357-9

FULL PAPER Open Access

Geomagnetic Virtual Observatories:

monitoring geomagnetic secular variation
with the Swarm satellites
Magnus D. Hammer1* , Grace A. Cox2, William J. Brown2, Ciarán D. Beggan2 and Christopher C. Finlay1

Abstract
We present geomagnetic main field and secular variation time series, at 300 equal-area distributed locations and
at 490 km altitude, derived from magnetic field measurements collected by the three Swarm satellites. These Geo-
magnetic Virtual Observatory (GVO) series provide a convenient means to globally monitor and analyze long-term
variations of the geomagnetic field from low-Earth orbit. The series are obtained by robust fits of local Cartesian
potential field models to along-track and East–West sums and differences of Swarm satellite data collected within a
radius of 700 km of the GVO locations during either 1-monthly or 4-monthly time windows. We describe two GVO
data products: (1) ‘Observed Field’ GVO time series, where all observed sources contribute to the estimated values,
without any data selection or correction, and (2) ‘Core Field’ GVO time series, where additional data selection is carried
out, then de-noising schemes and epoch-by-epoch spherical harmonic analysis are applied to reduce contamination
by magnetospheric and ionospheric signals. Secular variation series are provided as annual differences of the Core
Field GVOs. We present examples of the resulting Swarm GVO series, assessing their quality through comparisons with
ground observatories and geomagnetic field models. In benchmark comparisons with six high-quality mid-to-low
latitude ground observatories we find the secular variation of the Core Field GVO field intensities, calculated using
annual differences, agrees to an rms of 1.8 nT/yr and 1.2 nT/yr for the 1-monthly and 4-monthly versions, respectively.
Regular sampling in space and time, and the availability of data error estimates, makes the GVO series well suited for
users wishing to perform data assimilation studies of core dynamics, or to study long-period magnetospheric and
ionospheric signals and their induced counterparts. The Swarm GVO time series will be regularly updated, approxi-
mately every four months, allowing ready access to the latest secular variation data from the Swarm satellites.
Keywords: Geomagnetism, Secular variation, Geodynamo, Earth’s core, Swarm satellites

Introduction situation we are forced to rely on carefully monitoring

The geomagnetic field undergoes gradual change, evolv- the geomagnetic field and its changes in order to provide
ing year by year in a process known as geomagnetic the information necessary for navigation and orientation
secular variation. These changes are thought to result applications, and for descriptions of near-Earth radiation
primarily from motions of liquid metal in the Earth’s belts and current systems.
outer core but this process is not yet well enough under- To make progress beyond field monitoring, detailed
stood to allow accurate predictions of future behavior, information on the geomagnetic variations as a function
even a few years ahead (e.g., Alken et al. 2020a). In this of space and time must be combined with knowledge of
the underlying physical processes. With the rapid devel-
opment of numerical geodynamo models over the past
*Correspondence: [email protected]
1
Division of Geomagnetism, DTU Space, Technical University of Denmark, decade (e.g., Aubert and Finlay 2019), there is now the
Centrifugevej 356, Kongens Lyngby, Denmark prospect of assimilating such information into realistic
Full list of author information is available at the end of the article

© The Author(s) 2021. This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing,
adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and
the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material
in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material
is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the
permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creat​iveco​
mmons​.org/licen​ses/by/4.0/.
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 2 of 22

models, such that the processes underlying secular vari- describing the Swarm GVO series and presenting exam-
ation can be better understood. ple validation comparisons with ground observatories.
For both monitoring long-term geomagnetic variations, In addition to taking account of the most important rec-
and for data assimilation applications, it is an advantage ommendations from earlier GVO studies, the series pre-
to have processed satellite magnetic field data available sented here also take advantage of principal component
on a well organized grid, with a regular sampling rate in analysis (PCA) (Cox et al. 2018) and spherical harmonic
space and time. The Geomagnetic Virtual Observatory analysis (SHA) in an effort to better isolate the core field
(GVO) method is one approach to obtain such a dataset. signal.
The Geomagnetic Virtual Observatory method was In the “Data” section we describe the input data from
first proposed by Mandea and Olsen (2006) as a tool the Swarm satellite mission, and the adopted data selec-
for making satellite magnetic field measurements easily tion strategies. In the Sect. “Methodology” we describe
accessible as time series of the vector geomagnetic field in detail how the GVO series are calculated. The Sect.
at pre-specified locations. The GVO method involves fit- “Results” presents examples of GVO time series, derived
ting a scalar magnetic potential to satellite magnetic field using Swarm measurements from December 2013 to
observations from a chosen time window and within a March 2020 and describes comparisons with ground
local region, defined by a cylinder centered on a GVO observatory magnetic field series and global field model
target point. The potential is then used to compute the predictions. In the Sect. “Discussion and conclusions” we
magnetic field at the GVO target point such that a mean reflect on what can be learned from these comparisons,
magnetic field over a chosen time window at satellite alti- describe possible applications for the GVO series and
tude is determined; see Fig. 1. The GVO time series thus mention ideas for extending and improving the present
mimics the time series produced by ground-based mag- GVO approach.
netic observatories on timescales of months and longer.
The main advantage of the GVO time series is that they Data
can be produced at any sites of interest that are covered This section describes the satellite magnetic field meas-
by satellite data, and in particular, can provide a global urements used to derive the Swarm GVO time series. The
grid of time series derived from measurements made by GVO products take as input vector magnetic field meas-
similar instruments onboard satellites such as the Swarm urements in the form of the Swarm Level 1b (L1b) prod-
trio. uct MAGX_LR_1B, which contains quality-screened,
Applications of the GVO time series include geomag- calibrated and corrected measurements given in physi-
netic jerk studies (Olsen and Mandea 2007), comparisons cal SI units (nT) in a North, East, Center, hereafter NEC,
with spherical harmonic (SH) based geomagnetic field reference frame. For the results presented here, we
models (Olsen et al. 2009, 2010), core flow studies (Kloss use Swarm data versions 0506 from the 1-Dec-2013 to
and Finlay 2019; Rogers et al. 2019) and data assimilation 30-Mar-2020.
studies (Barrois et al. 2018). The GVO method can also From the Swarm L1b 1Hz magnetic field data, two
be used to derive estimates of the magnetic field gradient separate data chains are produced. Data chain (a) simply
tensor (Hammer 2018). extracts all available measurements using a sub-sampling
Focusing on the core magnetic field, initial studies of 15s.
showed that the original GVO series were contaminated Data chain (b) extracts, again using a sub-sampling of
by ionospheric and magnetospheric sources (Beggan 15s, only those measurements that satisfy the following
et al. 2009; Domingos et al. 2019; Olsen and Mandea dark, geomagnetic quiet-time selection criteria:
2007; Shore 2013). Recommendations for improving the
original GVO concept and better removing such contam- • Gross measurement outliers for which the vector
ination have been proposed (Hammer 2018; Shore 2013). field components deviate more than 500 nT from the
Some of these improvements were implemented in more predictions of the latest CHAOS field model (here
recent GVO series that have been used for core flow version CHAOS-7.2 (Finlay et al. 2020)) are rejected
studies by Barrois et al. (2018); Kloss and Finlay (2019); • The Sun is at least 10◦ below horizon
Rogers et al. (2019); and Whaler (2017). • Geomagnetic activity index Kp < 30
Here, we present details of an updated processing GVO • Time rate of change of Ring Current (RC) index
scheme that has been developed during a Swarm DISC |dRC/dt| < 3nT/hr−1 (Olsen et al. 2014)
(Data, Innovation and Science Cluster) project and is • Merging electric field at the magnetopause
now being used to produce regularly updated Swarm Em < 0.8mVm−1, (Olsen et al. 2014)
GVO time series as an official ESA Level 2 product. • Constraints on IMF requiring Bz > 0nT and
The primary purpose of this paper serves as a reference |By | < 10 nT.
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 3 of 22

�dlEW = [BlSWA (r1 , t1 ) + BlSWC (r2 , t2 )]/2. For a particu-

lar orbit of SWA, the corresponding SWC data were
selected to be that closest in latitude with the condition
that |�t| = |t1 − t2 | < 50 s.

Methodology
This section describes in detail the algorithms used to
derive the following Swarm GVO products:

1) 1- and 4-monthly time series of the ‘Observed Field’

2) 1- and 4-monthly time series of the ‘Core Field’ and
its secular variation.

Each product involves time series of spherical polar com-

ponents of the vector magnetic field on an approximately
equal-area global grid of 300 locations at an altitude of
490 km above the mean Earth radius. An overview of the
algorithm is presented in Fig. 2.

GVO locations and timestamps

Fig. 1 Illustration of the Geomagnetic Virtual Observatory concept;
For a given GVO target point, and considering a speci-
satellite magnetic measurements from within a target cylinder are fied time window of either 1 or 4 months, input data that
used to infer field time series at the GVO location given by a red dot. fall within a cylinder of horizontal radius rcyl = 700 km
Note the cylinder radius is not to scale, the actual cylinder footprints around the target point, and which also satisfy the rel-
are shown in Fig. 3 evant selection criteria (see Sect. "Data"), are extracted.
The GVO locations are specified in spherical polar coor-
dinates rGVO = (r, θ , φ), at fixed radius r = ra + hGVO
2-hourly means are computed from 1-min values of the where hGVO is the height above the Earth’s mean spheri-
solar wind and IMF from the OMNI data-base (http:// cal radius ra = 6371.2 km. For the Swarm data described
omniw​eb.gsfc.nasa.gov). hGVO = 490 km, so the GVOs are located at approxi-
The Swarm GVO method described in the next section mately the mean orbital height of the Swarm satellites
makes use of sums and differences of the satellite mag- during 2013–2020, considering each of the lower pair to
netic field measurements. Denoting the input magnetic contribute with half weighting.
data at a given position r by dl (r) = l̂ · B(r), where B(r) The GVO time series are provided in a global approxi-
is the vector magnetic field and l̂ is a unit vector in the mately equal area grid based on the sphere partitioning
component direction, then dl and dl denote sums algorithm of Leopardi (2006). Selecting a number of
and differences of the vector magnetic field compo- GVO grid points, and an associated target cylinder
nents, respectively. Both along-track (AT) and East–West search radius rcyl that avoids overlap of the target cylin-
(EW) data sums and differences are considered such ders to ensure independent data, involves a trade-off;
that �dl = (�dlAT , �dlEW ) and �dl = (�dlAT , �dlEW ). decreasing the number of target points and increasing
Along-track data differences are calculated using the the search radius allows for more data within each GVO
15-s differences �dlAT = [Bl (r, t) − Bl (r + δr, t + 15s)], cylinder but at the same time lowers the spatial resolu-
where δr = (δr, δθ, δφ) is the change in position. A tion. Preliminary tests with Swarm data suggested that
15-s along-track difference with a satellite speed of 300 GVO grid locations provided a suitable balance
≈ 7.7 km/s corresponds to a distance of 115 km (Olsen (Hammer 2018). If higher spatial resolution is required,
et al. 2015). Along-track sums are similarly calcu- longer time windows than used here are necessary in
lated as �dlAT = [Bl (r, t) + Bl (r + δr, t + 15 s)]/2. order to obtain stable GVO estimates. The surface area
The East–West differences are calculated as dS covered by each GVO target cylinder is the total sur-
�dlEW = [BlSWA (r1 , t1 ) − BlSWC (r2 , t2 )] having an face area A divided by the number of GVOs, NGVO = 300,
East–West orbit separation between the Swarm Alpha i.e., dS ∼ A/NGVO = 4π(ra + hGVO )2 /NGVO , where ra is
(SWA) and Charlie (SWC) satellites of ≈ 1.4◦, cor- the Earth’s mean radius 6371.2 km and hGVO is altitude of
responding to 155 km at the equator (Olsen et al. GVOs, here 490 km. Equating this area with the area of a
2015). The East–West sums were calculated as
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 4 of 22

circle surrounding the GVO, 2 , gives a target cylinder

π rcyl secular variation series are computed from annual dif-

search radius of rcyl = 4(ra + hGVO )2 /NGVO ≈ 700km, ferences of the 1-monthly and 4-monthly series, so their
where we have rounded down to the nearest hundred kil- timestamps are shifted by 6 months compared with the
ometers to ensure no overlap. The distance between any field series. GVO epoch times are for formal reasons
two GVOs is thus ≈ 1400km. This corresponds roughly given in GVO product files as milliseconds since 01-Jan-
to SH degree n = 14 , since the SH degree n is associated 0000 00:00:00.000, following the convention for Swarm
with a horizontal wavelength at satellite altitude is data products.
√
n ∼ 2π(ra + hGVO )/ n(n + 1) (Backus et al. 1996).
With a target cylinder search radius of 700km, approxi- Data pre‑processing
mately 80% of the data are used; the combined area of the In order to derive the Observed Field GVO time series,
cylinder footprints thus does not span the entire area of we start from the geocentric spherical polar components
the spherical surface, but the independence of each GVO of the vector magnetic field, Bobs = (Br , Bθ , Bφ ), and sub-
estimate is ensured. The top panel in Fig. 3 illustrates the tract predictions, BMF , from the IGRF main field model
locations of the 300 globally distributed GVOs and the (Alken et al. 2020b) for spherical harmonic degrees
footprint of the data target cylinders for each GVO. The n ∈ [1, 13]. This results in the following Observed Field
grid also contains GVOs at the North and South Poles. At residuals
these positions the (r, θ, φ) frame is defined by letting θ be (1)
δBobs = Bobs − BMF .
aligned along the Greenwich meridian, r point upwards
and φ completes the right-handed coordinate system. These residuals are used to derive the GVO estimates as
When computing the main field at the North/South Pole described in Sect. "GVO model parameterization and
from field models, the average of the main field values estimation" below. Note that IGRF predictions at the
evaluated 0.1◦ in latitude from the North/South Pole at GVO target points and times are added back during this
longitudes 0◦ and 180◦ was used. procedure.
For the 1-monthly series each GVO estimate has a In order to derive 1-monthly Core Field GVO time
timestamp corresponding to the 15th day of the consid- series from data chain (a), predictions from a lithospheric
ered calendar month. For the 4-monthly series, which are field model Blith are also removed:
constructed using data from within the intervals Janu-
ary–April, May–August, and September–December, the δBcore,1month = Bobs − BMF − Blith . (2)
GVO estimates have been allocated timestamps of the 1st Here, we calculate Blith using SH degrees n ∈ [14, 185]
of March, the 1st of July and the 1st of November. The from the LCS-1 model (Olsen et al. 2017).

Fig. 2 Overview of the Swarm GVO data product processing and model estimation algorithm
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 5 of 22

To derive 4-monthly core field GVO time series from

data chain (b), models of the magnetospheric and iono-
spheric fields are also removed during the pre-processing:

δBcore,4month = Bobs − BMF − Blith − Bmag − Biono ,

(3)
where B mag is the predicted large-scale magnetospheric
field and its Earth-induced counterpart field, as given by
the CHAOS model (Finlay et al. 2020), and Biono is the
predicted ionospheric field and its Earth-induced field as
given by the CI model (Sabaka et al. 2018).
Estimates of the main field are in all cases removed
from the data before carrying out the GVO estimation,
and then afterwards added back at the target location
and time; this step is necessary in order to pre-whiten the
data before carrying out the GVOestimation. Previous
studies have shown that such pre-whitening by remov-
ing a main field model is necessary in order to avoid noisy
GVO estimates. Hammer (2018, p.74, Fig 4.6) presents
examples of GVO series computed with and without
pre-whitening applied. Without pre-whitening robust
estimation schemes, based on iteratively reweighted
least squares, are unable to correctly identify and down-
weight disturbed data resulting in noisier estimates. The
specific main field model used for the pre-whitening is Fig. 3 Top: distribution of the 300 GVOs (red dots) and associated
however not crucial. For example, comparing GVO esti- cylinder bins (in green) using a Hammer projection. Bottom:
mates constructed using IGRF-13 and the CHAOS-7.2 illustration of the geocentric coordinate system and the local
main field model, we found rms differences compared to topocentric Cartesian coordinate system used in the GVO estimation
the benchmark ground observatories described in Sect. (in red)
"Validation tests" that agreed to within 0.05 nT across all
three components, for both 4-monthly and 1-monthly
GVO estimates. We therefore chose to use IGRF-13 for The residual magnetic field and the associated locations
the pre-whitening step in for producing the Swarm GVO within a specific target cylinder are transformed from
series in order to emphasize that the results obtained are the spherical coordinate system to a right-handed local
not simply a result of biasing towards the CHAOS model. topocentric Cartesian frame centered on (and constant
Note that when considering time windows of 1 or 4 within) each target cylinder, where at the GVO location
months for the GVO estimates, any information on time x points towards geographic south, y points towards East
variations with periods shorter than these intervals is, of and z points upwards. The bottom panel in Fig. 3 illus-
course, lost. trates the geocentric spherical and local topocentric
frames. Note that the unit vectors of the local Cartesian
GVO model parameterization and estimation frame, (êx , êy , êz ), coincide with the spherical unit vec-
A Cartesian potential forward model tors, (êθ , êφ , êr ), at the target location but not elsewhere.
We assume that magnetic field measurements are made The magnetic scalar potential, V, associated with the
in a source-free region such that the residual magnetic residual magnetic field in a source-free region must sat-
field is a Laplacian potential field, which fulfills the quasi- isfy Laplace’s equation ∇ 2 V = 0 and the potential is
stationary approximation (Backus et al. 1996). In the related to the residual field by δB = −∇V . The solution
following, we will use the general notation δB for the to Laplace’s equation in Cartesian coordinates can be
residual fields of Eqs. (1, 2, 3) and refer to the position written as a sum of harmonic polynomials (e.g., Backus
of the Geomagnetic Virtual Observatory as the target et al. 1996)
location.
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 6 of 22

L

L

a b c
V (x, y, z) = Cabc xa yb z c
V (x, y, z) = Cabc x y z l=1
l=1
= C100 x + C010 y + C001 z + C200 x2 + C020 y2
= C100 x + C010 y + C001 z + C200 x2 + C020 y2 + C002 z 2
+ C002 z 2 + C110 xy + C101 xz + C011 yz + C300 x3 (4)
+ C110 xy + C101 xz + C011 yz + C300 x3 + C030 y3 + C003 z 3 3 3 2 2
+ C030 y + C003 z + C210 x y + C201 x z
+ C210 x2 y + C201 x2 z + C120 y2 x + C021 y2 z + C102 z 2 x + C012 z 2 y
+ C120 y2 x + C021 y2 z + C102 z 2 x + C012 z 2 y
+ C111 xyz + · · · ,
+ C111 xyz + · · · ,

The potential Eq. (6) also fulfills the curl-free criteria.

where l = a + b + c , and Cabc are the expansion coef-
With an expansion for the magnetic potential estab-
ficients, a, b, c are non-negative integers, and L is the
lished, we can now write a linear forward problem
expansion order. In tests we found that for a GVO cylin-
relating a vector m , consisting of the model coefficients
der radius of 700 km, it was sufficient to expand the mag-
Cabc , to a vector dvec of predictions for the residual mag-
netic scalar potential to cubic order L = 3; this involves
netic field components in the local Cartesian system,
the 19 parameters in the above expansion.
(δBx , δBy , δBz ), at the positions of satellite observations
To be a valid potential field requires irrotational
that fall with the GVO target cylinder for the time win-
(∇ × δB = 0) and solenoidal (∇ · δB = 0) conditions be
dow under consideration,
satisfied. First we consider the solenoidal divergence-free
criteria. This requires ∇ 2 V = 0 which on inserting for dvec = Gvec m. (7)
the potential from Eq. (4) implies
Here Gvec is then a design matrix constructed from
∂ 2V ∂ 2V ∂ 2V
0= + + appropriate spatial derivatives of the potential.
∂x 2 ∂y 2 ∂z 2
Olsen et al. (2015) and Sabaka et al. (2018) have dem-
= −2C200 − 6C300 x − 2C210 y − 2C201 z − 2C020 − 6C030 y
onstrated that using East–West differences (between
− 2C120 x − 2C021 z − 2C002 − 6C003 z − 2C102 x − 2C012 y
Swarm A and C) and along-track differences (for all
= −(C200 + C020 + C002 ) − x(3C300 + C120 + C102 )
three Swarm satellites) improves the resolution of both
− y(3C030 + C210 + C012 ) − z(3C003 + C201 + C021 ).
the lithospheric field and the core field. They argued that
(5)
correlated errors due to incompletely modeled large-
Each of the terms in parenthesis must equal zero for this scale magnetospheric fields are reduced when using
to be satisfied. This means that such field differences. In addition to field differences,
1 East–West and along-track sums of the measurements
C002 = −(C200 + C020 ), C300 = − (C102 + C120 )
3 also need to be included in order to adequately con-
1 1 strain the largest wavelength parts of the field (Ham-
C030 = − (C210 + C012 ), C003 = − (C201 + C021 ). mer 2018; Sabaka et al. 2013). Based on this, we have
3 3
chosen to use sums and differences of the vector com-
The cubic potential series is thereby reduced by 4 param- ponents of the residual magnetic field, constructed
eters to a total of 15 parameters along satellite tracks and considering East–West pairs
of data between Swarm A and C. This results in a data

V (x, y, z) = C100 x + C010 y + C001 z + C200 x2

2 2
V (x, y, z) = C100 x + C010 y + C001 z + C200 x + C020 y + C020 y2 − (C200 + C020 )z 2 + C110 xy
− (C200 + C020 )z 2 + C110 xy + C101 xz + C011 yz 1 (6)
+ C101 xz + C011 yz − (C102 + C120 )x3
1 3 1 3 1 3 3
− (C102 + C120 )x − (C210 + C012 )y − (C201 + C021 )z
3 3 3 1 1
− (C210 + C012 )y − (C201 + C021 )z 3
3
+ C210 x2 y + C201 x2 z + C120 y2 x + C021 y2 z + C102 z 2 x 3 3
2 2
2
+ C z y + C xyz. + C 210 x y + C 201 x z + C120 y2 x + C021 y2 z
012 111
+ C102 z 2 x + C012 z 2 y + C111 xyz.
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 7 of 22

vector d = {dxvec , dyvec , dzvec , dxvec , dyvec , dzvec }, Observed Field GVOs
where  and  denote the differences and sums We define ‘Observed Field’ GVOs as field estimates
of the residual field described in Sect. "Data". The computed from satellite observations while retaining
design matrix linking the sums and differences to all observed geomagnetic field sources. Observed Field
the coefficients of the potential is constructed as GVO time series are derived from the sums and differ-
G = {�Gxvec ; �Gyvec ; �Gzvec ; �Gxvec ; �Gyvec ; �Gzvec )} ences of the residual field computed using Eq. (1) and
where �Gkvec = [Gkvec (r1 ) − Gkvec (r2 )] and then applying the GVO method described by Eqs. (8–10).
�Gk = [Gk (r1 ) + Gkvec (r2 )]/2 with k = (x, y, z).
vec vec One-monthly observed field GVOs are computed from
data chain a) while 4-monthly observed field GVOs are
Robust least squares estimation computed from data chain b).
Based on the above definitions of d and G for sums and Error estimates, σobs, for the Observed Field GVOs are
differences of the residual magnetic field, the coefficients assumed to be time-independent and spatially uncorre-
of the GVO model can be estimated using the following lated. They are calculated separately for each GVO times
robust least-squares inversion scheme series (i.e., for each field component at each GVO loca-
tion) based on a robust version of the total mean square
m = (GT W G)−1 GT Wd. (8) error (e.g., Bendat and Piersol 2010), that includes both
mean square residual and the mean residual squared,
Here W is a diagonal weight matrix, consisting of robust between the input data di and the GVO estimates d̂i for a
(Huber) weights for each entry in the data vector (e.g., given series. With ei = di − d̂i this is calculated as
Constable 1988), and an additional down-weighting 
factor of 1/2 for data from satellites Alpha and Charlie (ei − µw )2
i wi
which takes into account that these two satellites fly side- σobs = + µ2w , (11)
by-side and therefore provide similar measurements. i wi

Having determined the potential, estimates of the resid- where the index i runs
ual magnetic field components, in local cartesian coordi-  over all data contributing to a
given series, µw = i wi ei / i wi is the robust mean
nates, at GVO target location (i.e., x = 0, y = 0, z = 0) residual and the robust weights wi are calculated itera-
are computed as follows: tively assuming a long-tailed Huber distribution (Consta-
  ble 1988):
C100
δBGVO (x, y, z) = −∇V (0, 0, 0) = − C010 . 
(9) 1 if ǫi ≤ cw
C001 wi =
cw /ǫi if ǫi > cw , (12)

At the GVO target location, the local Cartesian field with ǫi = abs(ei )/std(e) and cw = 1.5 is the chosen break-
components are directly related to spherical polar point for the Huber distribution.
field components (see Fig. 3) with δBGVO,r = δBGVO,z ,
δBGVO,θ = δBGVO,x and δBGVO,φ = δBGVO,y . Each esti- Core Field GVOs and Secular Variation
mate is for a specific target GVO location r and epoch t, We define ‘Core Field’ GVOs as field estimates computed
which is the center of the considered time window. The from satellite observations with non-core fields removed
above procedure is repeated for each time window at (as far as possible). The core field and associated secular
each target location to obtain time series of estimates of variation (SV) GVO time series are produced as follows.
the residual vector field at all GVO target locations. First, 1- and 4-monthly GVO data files are produced,
A final step is needed to obtain the GVO estimates for after which the 1-monthly GVOs are de-noised by a
the field. This is to add back the prediction of the main principal component analysis. Next, an epoch-by-epoch
field model BMF GVO (r, t), at each target point and epoch, spherical harmonic analysis is carried out and the result-
using the same model (here IGRF-13) that was removed ing external and toroidal magnetic fields (i.e., non-inter-
from each satellite measurement during the pre-process- nal parts) are removed. Finally, annual differences of each
ing. This step is carried out separately for each compo- series are computed in order to obtain the GVO core
nent at each GVO location and each epoch, such that we field SV time series.
finally obtain the GVO vector field time series For the 1-monthly Core Field GVOs, GVO estimates
BGVO (r, t) = δBGVO (r, t) + BMF are computed from sums and differences of the field
GVO (r, t). (10)
residuals using Eq. (2) based on data chain (a) (i.e., with-
The estimated GVO magnetic field is provided in spheri- out data selection criteria). For the 4-monthly GVOs
cal polar (r, θ , φ) vector components.
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 8 of 22

Core Field GVOs, GVO estimates are computed from the 4-month local time sampling bias that remains in the
sums and differences of the field residuals using Eq. (3) 1-month series where PCA de-noising is applied.
based on data chain (b) (i.e., with dark geomagnetically The key premise to our approach is that the SV residu-
quiet-time criteria applied). als (the difference between observed GVO SV and that
predicted by an internal magnetic field model) provide
Principal component analysis information about contaminating signals that are present
Since the 1-monthly Core Field GVOs were derived with- in the GVO data but not in the internal model. The PCA
out data selection and having no model estimates of the of the SV residual covariance matrix leads to a proxies for
ionosphere nor magnetosphere removed, external mag- these contaminating signals that are then removed from
netic field signals remain. Such signals are considered as the GVO data. We approximate the GVO SV series using
contamination (‘noise’) in the current context because annual differences and the SV residuals are calculated as
our goal is to produce GVO estimates of the core field the difference between the GVO SV estimates and the SV
only. The monthly sampling rate means that a local time predicted by the CHAOS-7.2 model (Finlay et al. 2020)
sampling bias also contaminates the GVO estimates, as it evaluated up to SH degree 13 at the same times and loca-
takes approximately 4 months for each satellite to revisit tions. Comparable results can be achieved using alterna-
the same local time on Earth’s surface when considering tive field models, provided they represent time variation
both ascending and descending orbit tracks (see Shore of the main field in a continuous manner when detrend-
2013). To produce 1-monthly Core Field GVOs we there- ing each GVO SV series.
fore employ the principal component analysis (PCA) In their application to ground magnetic data, Cox et al.
method and Python package (MagPySV) described in (2018) found that this method is most effective when
Cox et al. (2018), to separate out and remove the various considering groups of observatories at similar magnetic
contaminating signals from the 1-monthly GVO esti- latitudes because the dominant external magnetic field
mates. This procedure is based on earlier work by War- source varies with magnetic latitude. Suitably grouped
dinski and Holme (2011) and Brown et al. (2013), who observatories experience similar noise at the same times
used the PCA method to de-noise ground observatory and these correlated signals show clearly in the dominant
data one observatory at a time, rather than de-noising principle components (PCs) of the SV residuals. On that
time series from several locations simultaneously, as in basis, we estimated the mean magnetic latitude at GVO
Cox et al. (2018) and this work. A brief summary of this locations using the AACGM-v2 Python package (Burrell
method is provided here; the reader is referred to Cox et al. 2020; Shepherd 2014) and assigned them to one of
et al. (2018), Cox et al. (2020) and the companion to this five magnetic latitude regions: Polar North, Polar South,
paper Brown et al., in preparation for further details. Auroral North, Auroral South and Low- to Mid-magnetic
Domingos et al. (2019) applied PCA to an earlier ver- latitudes (see Table 1).
sion of the 4-monthly GVO data series, considering both For N GVO locations, the SV residual covariance
CHAMP and Swarm measurements. They performed matrix for the vector time series is 3N by 3N, and can
PCA directly on GVO data series, rather than on annual be decomposed into 3N eigenvalues and eigenvectors,
differences of GVO series after subtracting predic- describing the PCs of the SV residual data set. The con-
tions from a core field model as we do. Hence, our PCA tributions of the K dominant PCs, corresponding to the
analysis looks for coherent signals that remain once fea- K largest eigenvalues, are removed from the SV residuals,
tures like the large-scale internal variations identified and afterwards the internal model SV from the CHAOS-
by Domingos et al. (2019) have been removed. Whilst 7-2 model is added back to the corrected residuals to
their focus was on modes associated with the vari- form the de-noised SV.
ance of internal field, they also identified an interesting In this application, we remove the most significant K
mode associated with annual variations of the external PCs entirely, as opposed to removing the scalar projec-
field in their Swarm GVO series. Our analysis is not well tion of a proxy signal for the PC content as described in
suited to studying annual variations since we apply PCA Cox et al. (2018) and earlier related works. Our removal
to annual difference estimates of SV. After carrying out of PCs here involves the removal of the associated eigen-
tests with our PCA procedure we decided there was not vectors and the component of signal at each GVO loca-
much advantage in applying it to the 4-monthly GVOs. tion in the projected directions of these eigenvectors.
Our 4-monthly GVO SV series contain fewer identifi- Note the number K differs by region, depending on how
able coherent external signals on which we would apply many PCs can be confidently identified as arising from
the PCA. This is due to the dark quiet-time data selec- one of the expected contaminating sources described
tion criteria, the applied corrections for magnetospheric above.
and ionospheric fields, and the absence (by design) of
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 9 of 22

We identify PCs as noise sources based on their geo- et al. 2010), especially in the auroral regions. Due to
graphic distributions, correlations to annual differences space–time aliasing, these non-potential fields can leak
of external magnetic field proxies (e.g., Dst, Polar Cap into the GVO estimates (Olsen and Mandea 2007).
North/South, Em, AE (Kauristie et al. 2017)), or peaks in In the situation of non-vanishing, but purely radial,
their discrete Fourier transform (DFT) at the local time currents within the shell of measurements the magnetic
bias frequency. Table 1 gives the number of PCs identi- field can written in terms of poloidal, V int , V ext , toroidal,
fied as noise, along with the percentage of variance in the T sh, and scalar potentials (e.g., Backus 1986; Olsen 1997;
SV residuals accounted for by each of these PCs and the Olsen and Mandea 2007):
total percentage variance removed in each region.
In a last step, the de-noised SV are numerically inte- B = −∇V int − ∇V ext + ∇ × r̂T sh , (13)
grated to produce de-noised one-monthly magnetic field
where each of the potentials can be represented by
time series, again treating SV as annual differences. The
expansions up to some maximum SH degree N:

N 
 n
 m  ra n+1 m
V int (r, θ , φ, t) = ra gn (t)cos mφ + hnm (t)sin mφ Pn (θ), (14)
r
n=1 m=0

n
N   
  m  r n m
V ext (r, θ , φ, t) = ra qn (t)cos mφ + sm
n (t)sin mφ Pn (θ), (15)
ra
n=1 m=0

N 
 n
 m,c 
T sh (r, θ , φ, t) = ra tn (t)cos mφ + tnm,s (t)sin mφ Pnm (θ), (16)
n=1 m=0

where ra = 6371.2km is the reference value for the

Earth’s mean spherical radius, n and m are here the SH
de-noised magnetic field must be re-leveled at the start of degree and order, respectively, and Pnm are the associated
this calculation. We use the original GVO field values for Schmidt semi-normalized Legendre functions. In the
the first 12 time samples for this purpose, meaning that three expansions, {gnm , hm n } are the internal coefficients,
the de-noised field values start 12 months after the origi- {qnm , snm } are the external coefficients and {tnm,c , tnm,s } are
nal GVO time series begins. the expansion coefficients associated with the toroidal
scalar potential.
Spherical harmonic analysis Predictions of the geomagnetic field components at the
The magnetic field time series produced by the GVO GVO locations are linearly related to the above expansion
method assumes a potential field description. This implies coefficients such that a forward problem can be written
that no electrical currents exist within the measurement
region. In reality however, satellite magnetic measure- dGVO = GSH mSH , (17)
ments are made in the ionospheric F-region where in situ
electrical currents may be present (Olsen 1997; Sabaka

Table 1 Magnetic latitude boundaries for each of the five regions de-noised separately using PCA
Region Magnetic latitude N K Variance in kth PC (%) Total (%)

Polar North 70◦ ≤ m ≤ 90◦ 11 4 32.7, 28.6, 16.7, 5.3 83.3

Auroral North 50◦ ≤ m < 70◦ 30 3 68.6, 7.8, 6.3 82.7
Low/Mid Lats −50◦ ≤ m < 50◦ 216 4 69.3, 9.3, 5.2, 3.9 87.7
Auroral South −70◦ ≤ m < −50◦ 33 3 61.8, 10.5, 5.9 78.2
Polar South −90◦ ≤ m < −70◦ 10 4 30.7, 24.1, 15.3, 13.0 83.1
The number (N) of GVOs per region, the number (K) of PCs removed during each analysis, the percentage variance accounted for by each of the K PCs and the total
percentage variance removed
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 10 of 22

where the data for a given epoch, t are given Error estimates of the secular variation GVO time
by dGVO = {Br (r1 , t), ..., Br (rNGVO , t), ... series are computed in a similar manner as described
Bθ (r1 , t), ..., Bθ (rNGVO , t), Bφ (r1 , t), ..., Bφ (rNGVO , t)}, where above but using residuals between the SV GVO data,
NGVO is the number of GVOs, related to the expansion SVGVO , and the SV predictions of the CHAOS time-
coefficients mSH = {gnm , hm n , qn , sn , tn , tn } via a design
m m m,c m,s dependent internal field model.
matrix GSH , which is constructed from the spatial deriva-
tives of Eqs. (14, 15 and 16). Here, we truncated the inter- Validation tests
nal, external and toroidal expansions at SH degree 13 and Comparison of GVO series with ground magnetic
the model coefficients were determined epoch by epoch observatories
from the GVO estimates using a simple least-squares Validation tests were performed by comparing the GVOs
solution: and independent ground observatory (GObs) records,
which are the established standard reference data series
mSH = (GTSH GSH )−1 GTSH dGVO . (18) for monitoring long-term variations of the geomag-
netic field. Our validation tests considered data from 28
At epochs where an insufficient number of GVOs are INTERMAGNET (International Real-time Magnetic
available to ensure a stable solution, the external and Observatory Network) ground observatories, listed in
toroidal coefficients were determined by a linear interpo- Table 2. These were chosen for their representative geo-
lation between nearby epochs. Following the SHA, exter- graphic coverage, spanning both polar and non-polar lat-
nal and toroidal field estimates are removed epoch by itudes and all longitude sectors. Below we refer to polar
epoch from the 1- and 4-monthly time series to produce stations as being the 13 stations with colatitudes 0◦ to 36◦
final Core Field GVO time series. and 144◦ to 180◦, with the remaining 15 stations referred
to as non-polar stations. From these stations we further
Secular variation estimates selected six ‘benchmark’ stations (Chambon la Forêt,
The secular variation of the Core Field series at a par- Kakioka, Honolulu, Guam, Hermanus and Canberra)
ticular GVO location, r , for a given epoch t, is computed from mid-to-low latitudes that are well known for their
using annual differences between field values at time high quality. We use these in an attempt to establish, in
t + 0.5 yr and at time t − 0.5 yr: well-understood conditions, the extent to which Swarm
SVGVO (r, t) = BGVO (r, t + 0.5 yr) − BGVO (r, t − 0.5 yr). GVO series agree with ground records, with an emphasis
(19) on how well the core field secular variation is captured.
Annual differences are a well established way to estimate We used the Swarm AUX_OBS_2_ hourly mean
the core field secular variation since they remove annual ground observatory dataset, version 0122 from February
signals from ionospheric and magnetospheric signals 2020, maintained by the British Geological Survey (BGS),
that are otherwise difficult to isolate. Note however that retrieved from ftp://ftp.nerc-murch​ison.ac.uk/geoma​g/
such annual signals do remain in the GVO field series Swarm​/AUX_OBS. These data have been checked and
themselves. corrected for known baseline jumps (Macmillan and
Olsen 2013). From these hourly mean values for each
Error estimates selected observatory we compute:
The error estimates, σcore , for each Core Field GVO time (i) One-monthly and four-monthly simple mean field
series are assumed to be time-independent and spatially values, for each of the three spherical polar com-
uncorrelated. They are computed separately for each field ponents. These are used for comparisons with the
component at each GVO based on the residuals between Observed Field GVO products.
the GVO data and the corresponding predictions of the (ii) One-monthly and four-monthly versions of revised
time-dependent internal part of the CHAOS field model means (Olsen et al. 2014), wherein the CHAOS
for SH degrees n ∈ [1, 20]. Denoting the residuals by magnetospheric field (Finlay et al. 2016) and CM4
e = dGVO − dCHAOS the error estimates are given by ionospheric field predictions (Sabaka et al. 2004)
 (and their induced counterparts) are first removed
2 from the hourly means for each of the three
i (ei − µ) (20)
σcore = + µ2 , spherical polar field components and then robust
M
(Huber-weighted) means are computed over 1- or
where i = 1, ..., M denotes the ith data element, and M is 4-monthly non-overlapping windows. These series
the number of data in a given series and µ is the residual are used for comparisons with the Core Field and
mean. Secular Variation GVO products.
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 11 of 22

To enable direct comparisons with these ground obser- Table 2 List of the selected ground observatories used
vatory series, we computed dedicated GVO time series for validation tests, listed in alphabetic order
directly above each selected ground observatory, using Observatory name IAGA code Colatitude East longitude
the approach described in the Sect. "GVO model param-
eterization and estimation". We removed crustal bias Abisko ABK 21.64◦ 18.82◦
estimates from each series (computed as the median Ascension Island ASC 97.95◦ 345.62◦
residual from the CHAOS-7.2 internal field model to up Alice Springs ASP 113.77◦ 133.88◦
SH degree 16) and mapped the GVO estimates down- Boulder BOU 49.87◦ 254.77◦
wards to the position of the ground observatory at Earth’s Barrow BRW 18.68◦ 203.38◦
surface by removing the difference between CHAOS-7.2 Chambon la Forêt CLF 41.98◦ 2.27◦
model predictions at the GVO location and the ground College CMO 25.13◦ 212.14◦
observatory location. This results in series we refer to as Canberra CNB 125.32◦ 149.36◦
GObs (ti ) for the ground observatories and B
B GVO,map (ti ) Gan GAN 89.30◦ 73.15◦
j j
for the GVOs, respectively, both at the ground observa- Qeqertarsuaq (Godhavn) GDH 20.748◦ 306.467◦
tory location. The subscript j indicates either the r, θ or Guam GUA​ 76.41◦ 144.87◦
φ component, or the scalar field intensity F (computed Hermanus HER 124.43◦ 19.23◦
by taking the square-root of the sum of the squares of Honolulu HON 68.68◦ 202.0◦
the three spherical polar components). The root-mean- Hornsund HRN 12.99◦ 15.54◦
square (rms) deviation between the correspond ground Huancayo HUA 102.05◦ 284.67◦
observatory and GVO series was then computed as Kakioka KAK 53.77◦ 140.18◦
 Kourou KOU 84.79◦ 307.27◦
 Nd  2
 1  Learmonth LRM 112.22◦ 114.1◦
obs
rmsj =  B GVO,map (ti ) ,
GObs (ti ) − B
j j Mawson MAW 157.60◦ 62.88◦
Nd
i=1 Mbour MBO 75.62◦ 343.03◦
(21) Macquarie Island MCQ 144.5◦ 158.95◦
where the summation runs over the length of the time Narsarsuaq NAQ 28.84◦ 314.56◦
series i = 1, ..., Nd where data are available from both Niemegk NGK 37.93◦ 12.68◦
series. The rms differences for secular variation series are Novosibirsk NSV 35.15◦ 83.23◦
computed in the same fashion, using annual differences Port Stanley PST 141.70◦ 302.11◦
of the ground observatory field, BGObs (ti ), and Core Field Resolute Bay RES 15.31◦ 265.10◦
j
GVOs mapped to the ground observatory positions, Shumagin SHU 34.65◦ 199.54◦
GVO,map (ti ). We computed summary means over these
B Tristan da Cunha TDC 127.07◦ 347.68◦
j
rms values for groups of series from the polar regions, Source http://www.inter​magne​t.org/
the non-polar region and benchmark observatories. For
these tests we used the time interval 2015–2018 when
there is good availability of both definitive observatory
for this difference, and so it is a source of discrepancy
data and Swarm data.
between the two series, particularly for the horizontal
We note that despite being the best available infor-
components.
mation concerning secular variation, the ground obser-
vatory records are themselves inherently imperfect.
Comparisons of GVO series with field model predictions
INTERMAGNET standards require that long-term accu-
A second set of validation tests involved comparisons
racy of main field series be better than 5 nT, with the best
between the GVO products and predictions from geomag-
observatories having an estimated baseline uncertainty
netic field models. These have the advantage that the GVO
of up to 0.4 nT (Lesur et al. 2017). Beyond observatory
product, provided on a global grid, can be tested directly
measurement uncertainties, a further source of differ-
(without any mapping) and the global quality of the prod-
ences between ground observatory data and GVO esti-
ucts can be assessed. However, unlike the comparisons
mates is that the latter use data above the ionospheric
with ground observatories, tests against field model pre-
E-layer, while ground data are collected at the Earth’s
dictions are not fully independent as Swarm data were also
surface. They therefore observe ionospheric and mag-
used in the construction of the field models.
netosphere-ionosphere coupling currents differently.
Comparisons to models are based on the rms devia-
Our potential field mapping used to downward continue
tion between a given GVO time series, BjGVO (ti ), and field
the GVO estimates to Earth’s surface does not account
model predictions, Bjmod (ti ), at the GVO data location
and times
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 12 of 22

180
1M SV GVO-CORE
1M SV GVO-OBS
GVO locations

150
120
90
60 30
Longitude [deg]
dBr /dt

-30
-60
-90
-120
-150 0
2014.5-2019.5
20 nT/yr

-180
90

60

30

-30

-60

-90
0

Latitude [deg]

Fig. 4 Time series of Secular Variation from the Swarm satellites: annual differences of the 1-monthly Observed Field GVOs (blue dots) and the Core
Field GVOs (red dots). Shown here is the radial field component. GVO locations are marked with a black cross
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 13 of 22

180
4M SV GVO-CORE
4M SV GVO-OBS
GVO locations

150
120
90
60
30
Longitude [deg]
dBr /dt

0
-30
-60
-90
-120
-150
2014.7-2019.7
20 nT/yr

-180
-30

-60

-90
90

60

30

Latitude [deg]
Fig. 5 Time series of Secular Variation from the Swarm satellites: annual differences of the 4-monthly Observed Field GVOs (blue dots) and the Core
Field GVOs (red dots). Shown here is the radial field component. GVO locations are marked with a black cross
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 14 of 22


 N
GVO  2 or 4-monthly time window, considering the times of the
 1
rmsmod = Bmod (ti ) − BGVO (ti ) , actual data used to derive the GVO estimates. We note
j j j
NGVO
i=1 the model values compared to the Observed Field GVOs
(22) are not fully representative of all the fields contributing
to the GVOs, in particular they do not include realistic
where the summation runs over the length of the GVO
ionospheric fields in the polar region, or magnetosphere–
time series i = 1, ..., NGVO and j indicates a specific spher-
ionosphere coupling currents.
ical polar component r, θ , φ of the vector field or the sca-
For comparisons with the Core Field GVOs, Bjmod (ti )
lar field intensity F.
is computed using the time-dependent internal field
For comparisons with the Observed Field GVOs,
from the CHAOS-7.2 model (Finlay et al. 2020) using
Bjmod (ti ) is computed using the CHAOS-7.2 time-
SH degrees up to 20, with the LCS-1 lithospheric model
dependent internal field for degrees n ∈ [1, 13], the
(Olsen et al. 2017) for degrees n ∈ [14, 20] removed.
LCS-1 lithopsheric field model degrees n ∈ [14, 185],
For comparisons with the Core Field Secular Variation
as well as the CHAOS-7.2 magnetospheric field (and
GVOs, Bjmod (ti ) is computed using the first time deriva-
induced counterparts) and the CIY4 ionospheric field
tive of the time-dependent internal field from CHAOS-
(Sabaka et al. 2018) (and induced counterparts). The
7.2, again up to SH degree 20. In the global grid there
magnetospheric and ionospheric fields and their counter-
are 78 polar and 222 non-polar GVOs and Benchmark
parts are computed as mean values for each 1-monthly
Table 3 Summary of validation tests for the 1-monthly GVO products
GVO series Ground observatory comparison Model comparison Error estimates
rmsr rmsθ rmsφ rmsF rmsr rmsθ rmsφ rmsF rmsr rmsθ rmsφ

All GVO-OBS [nT] 3.00 6.98 5.25 3.91 4.16 3.59 4.00 3.83 4.17 7.18 6.92
GVO-CORE [nT] 4.12 5.17 3.55 4.37 3.84 2.83 2.16 3.67 3.42 2.78 1.85
GVO-SV [nT/yr] 3.06 4.35 3.60 3.18 1.60 1.64 1.48 1.69 1.62 1.66 1.32
Polar GVO-OBS [nT] 4.28 10.30 7.27 4.83 5.47 5.24 5.19 5.30 8.13 14.44 15.71
GVO-CORE [nT] 5.35 7.42 5.26 5.35 4.25 3.51 2.92 4.25 4.00 3.58 2.76
GVO-SV [nT/yr] 4.46 6.11 5.03 4.25 2.19 1.95 2.50 2.20 2.22 1.96 1.87
Non-polar GVO-OBS [nT] 1.90 4.11 3.51 3.12 3.70 3.01 3.59 3.31 2.78 4.63 3.83
GVO-CORE [nT] 3.05 3.23 2.08 3.52 3.69 2.59 1.89 3.46 3.22 2.50 1.53
GVO-SV [nT/yr] 1.86 2.83 2.35 2.26 1.40 1.53 1.12 1.51 1.41 1.55 1.13
Benchmark GVO-OBS [nT] 1.67 3.71 3.32 2.53 3.80 3.10 3.66 3.42 3.92 4.69 4.36
GVO-CORE [nT] 2.81 2.45 2.22 2.80 3.80 2.75 1.89 3.58 3.30 2.66 1.55
GVO-SV [nT/yr] 1.80 2.50 2.35 1.75 1.40 1.63 1.14 1.56 1.42 1.64 1.15

Table 4 Summary of validation tests for the 4-monthly GVO products

GVO series Ground observatory comparison Model comparison Error estimates
rmsr rmsθ rmsφ rmsF rmsr rmsθ rmsφ rmsF rmsr rmsθ rmsφ

All GVO-OBS [nT] 2.28 3.72 2.69 2.66 2.35 2.66 2.03 2.56 1.77 3.35 2.77
GVO-CORE [nT] 2.18 3.19 2.52 2.38 2.51 2.34 2.06 2.51 2.28 2.20 1.88
GVO-SV [nT/yr] 3.49 3.99 4.27 2.63 1.27 1.42 2.38 1.44 1.31 1.47 2.29
Polar GVO-OBS [nT] 3.16 4.50 3.29 3.07 2.69 3.78 3.27 2.55 3.16 5.83 6.13
GVO-CORE [nT] 3.25 4.25 3.39 3.12 2.96 3.36 2.96 2.87 2.65 3.29 2.78
GVO-SV [nT/yr] 5.83 6.59 6.49 3.81 2.15 2.68 3.38 2.79 2.23 2.78 2.84
Non-polar GVO-OBS [nT] 1.51 3.04 2.17 2.31 2.23 2.26 1.59 2.56 1.32 2.54 1.67
GVO-CORE [nT] 1.26 2.28 1.76 1.74 2.35 1.98 1.74 2.38 2.16 1.82 1.56
GVO-SV [nT] 1.47 1.73 2.34 1.60 0.96 0.98 2.03 0.96 0.99 1.01 2.10
Benchmark GVO-OBS [nT] 1.32 2.92 1.93 1.93 2.26 2.41 1.56 2.74 1.28 2.16 1.30
GVO-CORE [nT] 1.14 1.92 1.70 1.42 2.39 2.03 1.73 2.50 2.17 1.85 1.57
GVO-SV [nT/yr] 1.42 1.54 2.14 1.19 0.97 0.97 2.02 0.97 1.00 1.00 2.09
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 15 of 22

values were computed using GVOs ±30◦ in latitude from statistics. Table 3 collects results of the validation tests
the equator. Comparisons were made between 2014 and against independent ground observatories and field mod-
2020, throughout the time interval when GVO data were els for the 1-monthly GVO products, while Table 4 col-
available. lects similar statistics for 4-monthly GVO products. See
Sect. "Validation tests" above for details of the tests.
Results When considering the statistics presented here, it is
A global overview of the Swarm GVO time series important to recall that the number of ground obser-
To illustrate the 1-monthly GVO secular variation data vatories is split into 13 “Polar” stations, 15 “non-polar”
series, Fig. 4 presents a global map of annual differences stations and six “benchmark” stations. As mentioned in
of the radial field component of the Observed Field GVO Sect. "Validation tests" the stations in each category were
time series (blue dots) and of the Core field GVO time selected in order to obtain as far as possible reasonable
series (red dots). Fig. 5 presents a similar summary of the geographic coverage of both the polar and non-polar
global results for the 4-monthly GVO time series. Note regions. The aim with the benchmark stations was to
the small difference in the time scales shown at the bot- document and validate the performance of the GVO time
tom left of these two figures; the SV of the 1-monthly at known high-quality stations from mid-to-low latitudes
GVO-CORE time series begins in 2015.5, since the GVO- where external contributions are less prominent. The
CORE time series starts only in 2015 due to the PCA error estimates provided along with the GVO products
processing, while the SV of the 4-monthly GVO-CORE are also presented in these tables for reference. In these
begins in 2014.7 since no PCA is not performed on these. tables GVO-OBS, GVO-CORE and GVO-SV denotes
the Observed Field GVOs, the Core Field GVOs and the
Validation statistics: comparisons with ground Core Field Secular Variation GVOs, respectively.
observatories and field models
The results of the validation comparisons carried out
are presented here in the form of two summary tables of

Fig. 6 One-monthly Observed Field (blue dots) and Core Field (red dots) GVOs mapped to Earth’s surfaced with ±σ uncertainty envolopes,
together with simple monthly means (yellow stars) and revised monthly means (black stars) from three of the selected high-quality ‘benchmark’
ground observatories, left column: Kakioka (Japan), middle column: Honolulu (Hawaii,USA), right column: Canberra (Australia). Top row is the radial
field component, middle row is the southward field component, bottom row is the eastward field component, units are nT
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 16 of 22

Example comparisons of GVO and ground observatory southward θ-component, for example the peak observed
time series in the second half of 2017 at Kakioka. This feature, likely
More detailed insight comes from direct examination of of magnetospheric origin, is seen simultaneously at all
the time series of the ground observatory and associated benchmark stations in both the GVO and ground obser-
GVO series as described in Section "Validation tests". vatory series, and is particularly clear at Kakioka (KAK)
Fig. 6 presents the 1-monthly Observed Field (GVO- and Hermanus (not shown here). The amplitude of the
OBS, blue dots) and Core Field (GVO-CORE, red dots) peak is slightly lower in the GVO series. More scatter
GVO estimates, mapped down to the Earth’s surface at is seen in the eastward φ-component of the GVO series
three of the benchmark ground observatories. These compared to the ground observatory benchmark series
figures include ±σ uncertainty estimates, where we (e.g., at Honolulu, HON). The source of this scatter may
have made the assumption that these estimates remain be ionospheric or field-aligned currents seen by the satel-
unchanged when mapping the field to ground level. lites that are less prominent at ground; the amplitude of
When examining the Observed Field GVOs we present this scatter was larger 2014-2016, which may indicate a
time series of the field itself rather than the SV, so as not solar cycle dependence.
to filter out any signals that may be of interest by taking Figure 7 presents Observed Field (GVO-OBS) and Core
annual differences. Also plotted for comparison are the Field (GVO-CORE) GVO estimates along with their ±σ
ground observatory hourly monthly means (omm, yellow uncertainty, together with corresponding ordinary and
dots) and revised monthly means (rmm, black dots). revised ground observatory monthly means, from sta-
Radial field variations observed at the benchmark sta- tions in the more challenging polar regions. At these
tions are followed closely by the GVO series, for exam- locations, there are strong ionospheric E-region currents
ple at Kakioka (KAK) in Japan (left column, Fig 6) where lying between the satellites and the ground stations, and
both the trend in the field and its acceleration are in the satellites at times fly through intense field-aligned
agreement. The ability of the Observed Field GVO series currents. Nonetheless, the comparisons are encourag-
to track sub-annual field changes is illustrated by the ing and the trends seen at the ground stations are well

Fig. 7 One-monthly Observed Field (blue dots) and Core Field (red dots) GVOs mapped to Earth’s surfaced with ±σ uncertainty envolopes,
together with simple monthly means (yellow stars) and revised monthly means (black stars) from three of the selected polar ground observatories,
left column: Resolute Bay (Canada), middle column: Macquarie Island (Australia), right column: Mawson Station (Antarctica). Top row is the radial
field component, middle row is the southward field component, bottom row is the eastward field component, units are nT
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 17 of 22

Fig. 8 One-monthly Core Field SV GVOs mapped to Earth’s surface (red symbols) with ±σ uncertainties, and revised monthly means from selected
high-quality ‘benchmark’ ground observatories (black symbols), left column: Kakioka (Japan), middle column: Honolulu (Hawaii,USA), right column:
Canberra (Australia). Top row is the radial field component, middle row is the southward field component, bottom row is the eastward field
component, units are nT

captured by the GVO series. At the polar stations, the observatory in Antarctica (MAW) the southward θ-com-
amplitude of the error bars has been significantly reduced ponent has fluctuations of opposite sign to fluctuations
for the Core Field GVO series compared to the Observed seen at the same time in the GVO estimates. The relative
Field GVO series. position and orientations of the ionospheric currents and
The radial component at high northern latitudes in the ground and satellite observation points are clearly
Canada, at the Resolute Bay observatory (RES) inside the important for understanding such effects.
polar cap, shows a particularly clear annual variation in Figure 8 presents plots of the 1-monthly revised
the monthly means, peaking in the northern summer. monthly mean SV from ground observatories (black
These fluctuations, which are likely due to far-field effects dots) and the 1-monthly Core Field GVO SV series (red
of polar electrojet currents, are well tracked by the GVO dots) at the three low/mid benchmark locations. Note
estimates. the difference in scale here when looking at the secular
Larger differences between the GVO and ground variation, compared to the earlier plots that show the
observatory series are seen in the eastward φ-component Observed Field/Core Field GVO values without taking
at these stations, the difference being largest from 2014 to annual differences.
2017 (up to 25 nT seen at RES in summer months). The The absolute levels (i.e., amplitude of secular variation)
eastward φ-component in the GVO and ground stations and trends (i.e., secular acceleration) in these benchmark
agrees more closely at slightly lower latitudes in both the ground observatory records of the core field secular vari-
northern hemisphere (e.g., in Alaska at College station ation are well matched by the Core Field SV GVO series.
CMO, not shown here) and in the Southern hemisphere Peaks (secular variation impulses/geomagnetic jerks)
at Macquarie Island (MCQ), middle row Fig. 7. Ground such as that in the radial field at Honolulu (HON) in 2017
stations in the auroral zone see signals in the south- (Fig. 8, middle column, top row) are well captured and
ward θ-component that are less prominent in the GVO there is no indication of loss of temporal resolution in
estimates; these may be caused by polar electrojet cur- these annual difference secular variation series compared
rents that are closer to the ground stations. At Mawson to the ground records. This indicates time-dependent SV
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 18 of 22

Fig. 9 One-monthly Core Field SV GVOs mapped to Earth’s surface (red symbols) with ±σ uncertainties, and 1-monthly revised monthly means
from selected polar ground observatories (black symbols), left column: Resolute Bay (Canada), middle column: Macquarie Island (Australia), right
column: Mawson Station (Antarctica). Top row is the radial field component, middle row is the southward field component, bottom row is the
eastward field component, units are nT

with time scales down to 1 year is well captured in the especially in 2015 and 2016 when solar activity was
Swarm Core Field Secular Variation GVO product. The higher. This is particularly noticeable in the 4-monthly
scatter is slightly larger in the GVO series for the south- SV series in January 2015 and January 2016 and seems
ward θ-component and there are indications of remaining to be related to the fields measured by the Swarm satel-
noise (perhaps to due ionospheric or inter-hemispheric lites during summer 2015 (see e.g., Fig. 10). Comparisons
field-aligned currents) with period close to one year in with ground observatories and internal field models such
the eastward φ-component. Figure 9 shows similar com- as CHAOS show a noticeable bias in the Bφ component
parisons for a selection of the polar observatories. Here during this period which contributes to longer tails in the
the scatter is larger in both the ground and GVO data, distribution of residuals for Bφ all epochs and also results
due to the difficult of isolating the core field signal, but in enhanced rms differences for the φ-components of
again the observed trends agree well. SV in comparison to ground observatories, see Table 4.
Figures 10 and 11 present plots of the 4-monthly A similar bias is also seen when comparing the original
ground observatory SV (black dots) and 4-monthly GVO Swarm data to internal field models during summer 2015,
SV time series (red dots) at the same three low/mid and particularly for Swarm B. The residuals during this time
polar latitude benchmark locations. Considering Fig. 10, are largest in the northern polar region and seem to be
the scatter observed in the 1-monthly Core Field SV time geophysical in origin, perhaps related to strong field-
series has been reduced and the independent ground and aligned currents measured by the satellites during this
Swarm series show excellent agreement. The peak in the epoch.
SV observed in the radial component at Honolulu (HON) Despite the slightly higher scatter at the polar sta-
in 2017 is again well captured. Differences are apparent tions in Fig. 11, the agreement is again encouraging
at some epochs between the GVO series and the ground with trends seen at ground stations being captured in
observatory series in the eastward φ-components, the GVOs. Largest differences are seen in the horizontal
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 19 of 22

Fig. 10 Four-monthly Core Field SV GVOs mapped to Earth’s surface (red symbols) with ±σ uncertainties, and 4-monthly revised means from
selected high-quality ’benchmark’ ground observatories (black symbols), left column: Kakioka (Japan), middle column: Honolulu (Hawaii,USA), right
column: Canberra (Australia). Top row is the radial field component, middle row is the southward field component, bottom row is the eastward field
component, units are nT

components for Mawson station (MAW) in Antarctica currents that have different signatures at ground and sat-
where a sawtooth pattern about the ground series is visi- ellite altitude. The processing applied to obtain Core Field
ble in the 4-monthly GVO estimates. This enhanced scat- GVOs results in close agreement with ground obser-
ter is reflected in the error estimates supplied together vatory revised monthly means and with internal field
with the GVO products, but illustrates that caution is models.
needed when interpreting SV variations on interannual Taking annual differences to obtain SV estimates, fur-
and shorter timescales in the auroral zone. Further work ther improves the agreement. We find the secular vari-
is required to better understand these features. ation of the field intensity in the 1-monthly Core Field
GVOs agrees with six benchmark ground observatories
Discussion and conclusions from mid and low latitudes to a level of 1.8 nT/yr. For the
In Table 3 find that the 1-monthly Swarm GVO products 4-monthly Core Field GVOs the difference to the secular
Observed Field series agree with independent ground variation recorded at the ground observatories decreases
observatory and field model predictions to within 5 nT to 1.2 nT/yr. These numbers may be considered an upper
in all components at non-polar latitudes. Given that the bound on the accuracy of the Swarm GVO secular varia-
requirement for a good standard (INTERMAGNET) tion estimates, since they also include the measurement
ground observatory is an accuracy of 5 nT this indi- errors inherent in the ground observatories (perhaps
cates that the GVO method yields results comparable on 0.5 nT/yr at excellent observatories) as well as differences
these time scales with good ground observatories. The due to incomplete separation of non-core sources which
4-monthly estimates agree even better, to within 3 nT. will affect ground and GVO data in different ways.
Larger differences are found at polar latitudes where In this paper, we have presented a global network of
comparisons are complicated by the presence of strong Geomagnetic Virtual Observatories constructed from
ionospheric and magnetosphere–ionosphere coupling vector magnetic field measurements made by the Swarm
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 20 of 22

Fig. 11 Four-monthly Core Field SV GVOs mapped to Earth’s surface (red symbols) with ±σ uncertainties, and 4-monthly revised means from
selected polar ground observatories (black symbols), left column: Resolute Bay (Canada), middle column: Macquarie Island (Australia), right column:
Mawson Station (Antarctica). Top row is the radial field component, middle row is the southward field component, bottom row is the eastward field
component, units are nT

satellites. The series are provided in two variants, each Earlier versions of GVO series have already been used
with 1-monthly and 4-monthly cadences, and each with in inversions for the core surface flow (Kloss and Finlay
associated uncertainty estimates: 2019; Whaler and Beggan 2015) and in data assimila-
(1) ‘Observed’ magnetic field GVO series, with 1- and tion studies where the core field signals seen in GVOs
4-month cadence are combined with information from geodynamo mod-
(2) ‘Core’ magnetic field GVO series, and associated els in order to estimate the state of the core (Barrois et al.
annual difference secular variation series, with 1- and 2018). GVO series are particularly well suited for global
4-month cadence. studies of rapid core dynamics where a number of physi-
Good agreement has been demonstrated between the cal hypotheses are currently under exploration (Aubert
Swarm GVO series, ground observatory data, and exist- and Finlay 2019; Buffett and Matsui 2019; Gerick et al.
ing field models. The Swarm GVO series thus provide 2020). The Observed Field GVOs provide additional
consistent and accurate global information on geomag- information on long-period variations of magnetospheric
netic secular variation. and ionsospheric origin. Long-period magnetospheric
We recommend the Core Field GVOs along with their variations may prove useful for deep electromagnetic
supplied error estimates for use in studies of core dynam- induction studies (e.g. Harwood and Malin 1977). At
ics. Adopting the traditional approach of taking annual high latitudes signatures of the polar electrojets are
differences to obtain the SV helps avoid small annual sig- clearly seen in the 1-monthly Observed Field GVO series,
nals that can remain in the Core Field series. For future for example the distinctive annual variations in the ver-
work, we propose carrying out PCA de-noising based on tical component seen in Figure 7, reflecting seasonal
first differences of monthly GVOs, rather than annual variations of the polar electrojet current system. Both
differences, as a promising direction to further isolate applications will become increasingly attractive as the
core field signal. time series provided by the Swarm satellites lengthens.
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 21 of 22

The Swarm GVO dataset is now available as a regularly References

Alken P et al (2020a) Evaluation of candidate models for the 13th generation
updated ESA product (see ftp://swarm​-diss.eo.esa.int/ International Geomagnetic Reference Field. Earth Planets Space. https​://
Level​2long​term/ ) updates will take place approximately doi.org/10.1186/s4062​3-020-01281​-4
every 4 months. Similarly processed GVO series from Alken P et al (2020b) International geomagnetic reference field: the thirteenth
generation. Earth Planets Space. https​://doi.org/10.1186/s4062​3-020-
CHAMP, Cryosat-2 and Ørsted are also available from 01288​-x
the Swarm DISC GVO project webpage https​://www. Aubert J, Finlay C (2019) Geomagnetic jerks and rapid hydromagnetic waves
space​ . dtu.dk/engli ​ sh/resea ​ rch/proje ​ c ts/proje ​ c t-descr​ focusing at Earth’s core surface. Nat Geosci 12:393–398. https​://doi.
org/10.1038/s4156​1-019-0355-1
iptio​ns/geoma​gneti​c-virtu​al-obser​vator​ies. Backus G (1986) Poloidal and toroidal fields in geomagnetic field modelling.
Rev Geophys 24:75–109
Backus G, Parker R, Constable C (1996) Foundations of Geomagnetism. Cam-
Abbreviations bridge Univ. Press, New York
AT: Along-track; CHAOS: CHAMP, Ørsted, and Swarm field model; CI: Compre- Barrois O, Hammer MD, Finlay CC, Martin Y, Gillet N (2018) Assimilation of
hensive inversion model; DFT: Discrete Fourier transform; ECEF: Earth-Centered ground and satellite magnetic measurements: inference of core surface
Earth-Fixed reference frame; EW: East–West; GVO: Geomagnetic Virtual Obser- magnetic and velocity field changes. Geophys J Int 215:695–712
vatory; IGRF: International geomagnetic reference field; IMF: Interplanetary Beggan CD, Whaler KA, Macmillan S (2009) Biased residuals of core flow
magnetic field; INTERMAGNET: International real-time magnetic observatory models from satellite-derived virtual observatories. Geophys J Int
network; Kp: K planetary index; LCS: Lithospheric model from CHAMP and 177(2):463–475
Swarm; MF: Main field; nT: Nano-Tesla; OMM: Observed monthly mean; PC: Bendat J, Piersol A (2010) Random data, analysis and measurement proce-
Principal component; PCA: Principal component analysis; QD: Quasi-dipole; dures. Wiley, New Jersey
RC: Ring current index; rms: Root-mean square; RMM: Revised monthly mean; Brown W, Mound J, Livermore P (2013) Jerks abound: an analysis of geo-
SH: Spherical harmonic; SHA: Spherical harmonic analysis; SV: Secular variation; magnetic observatory data from 1957 to 2008. Phys Earth Planet Int
DISC: Data, innovation and science cluster; yr: Year. 223:62–76
Buffett B, Matsui H (2019) Equatorially trapped waves in Earth’s core. Geophys J
Acknowledgements Int 218(2):1210–1225
We wish to thank the European Space Agency (ESA) for the prompt avail- Burrell A, Meeren CVD, Laundal KM (2020) aburrell/aacgmv2: Version 2.6.0,
ability of Swarm L1b data. The staff of the geomagnetic observatories and https​://doi.org/10.5281/ZENOD​O.35987​05
INTERMAGNET are thanked for supplying high-quality observatory data. High- Constable C (1988) Parameter estimation in non-Gaussian noise. Geophys J Int
resolution 1-min OMNI data were provided by the Space Physics Data Facility 94(1):131–142
(SPDF), NASA Goddard Space Flight Center. We like to thank two anonymous Cox G, Brown W, Billingham L, Holme R (2018) Magpysv: a python package
reviewers for comments that helped improve the manuscript. for processing and denoising geomagnetic observatory data. Geochem
Geophys Geosyst 19(9):3347–3363
Authors’ contributions Cox G, Brown W, Beggan C, Hammer M, Finlay C (2020). Denoising Swarm
MDH developed the Swarm GVO data processing and modeling scheme geomagnetic virtual observatories using principal component analysis.
and drafted the manuscript. GAC developed the PCA de-noising scheme https​://doi.org/10.5194/egusp​here-egu20​20-9957
with assistance from WJB, applied it to the one monthly series and wrote Domingos J, Pais MA, Jault D, Mandea M (2019) Temporal resolution of internal
the related part of the manuscript. CCF and CDB designed the project and magnetic field modes from satellite data. Earth Planets Space 71(1):1–17
contributed to the writing of the manuscript. All authors read and approved Finlay CC, Olsen N, Kotsiaros S, Gillet N, Tøffner-Clausen L (2016) Recent
the final manuscript. geomagnetic secular variation from Swarm and ground observatories as
estimated in the CHAOS-6 geomagnetic field model. Earth Planets Space
Funding 68(1):1–18
This study was funded by ESA through the Swarm DISC GVO project, contract Finlay CC, Kloss C, Olsen N, Hammer MD, Tøffner-Clausen L, Grayver A, Kuvs-
no. 4000109587. hinov A (2020) The CHAOS-7 geomagnetic field model and observed
changes in the South Atlantic Anomaly. Earth Planets Space 72(1):1–31.
Availability of datasets and materials https​://doi.org/10.1186/s4062​3-020-01252​-9
Regularly updated versions of the Swarm GVO data series are available at ftp:// Gerick F, Jault D, Noir J (2020) Fast quasi-geostrophic magneto-coriolis modes
ftp.swarm​-diss.eo.esa.int/Level​2long​term/. Further details including additional in the Earth’s core, 856 Geophys. Res Lett. https​://doi.org/10.1029/2020G​
documentation, GVO datasets from other satellites and software are available L0908​03
from the Swarm DISC GVO project webpage, https​://www.space​.dtu.dk/engli​ Hammer MD (2018) Local estimation of the Earth’s core magnetic field, Ph.D.
sh/resea​rch/proje​cts/proje​ct-descr​iptio​ns/geoma​gneti​c-virtu​al-obser​vator​ thesis, Technical University of Denmark
ies. Swarm data are available at https​://earth​.esa.int/web/guest​/swarm​/data- Harwood JM, Malin SRC (1977) Sunspot cycle influence on the geomagnetic
acces​s. Ground observatory data are available from ftp://ftp.nerc-murch​ison. field. Geophys J Int 50(3):605–619. https​://doi.org/10.1111/j.1365-
ac.uk/geoma​g/Swarm​/AUX_OBS/hour/. The RC index is available from http:// 246X.1977.tb013​37.x
www.space​cente​r.dk/files​/magne​tic-model​s/RC/. The CHAOS-7 model and its Kauristie K, Morschhauser A, Olsen N, Finlay C, McPherron R, Gjerloev J,
updates are available at http://www.space​cente​r.dk/files​/magne​tic-model​s/ Opgenoorth H (2017) On the usage of geomagnetic indices for data
CHAOS​-7/. selection in internal field modelling. Space Sci Rev. 206:61–90. https​://doi.
org/10.1007/s1121​4-016-0301-0
Competing interests Kloss C, Finlay CC (2019) Time-dependent low latitude core flow and geomag-
The authors declare that they have no competing interests. netic field acceleration pulses. Geophys J Int 217:140–168
Leopardi P (2006) A partition of the unit sphere into regions of equal area and
Author details small diameter. Electronic Trans Numerical Anal 25(12):309–327
1
Division of Geomagnetism, DTU Space, Technical University of Denmark, Lesur V, Heumez B, Telali A, Lalanne X, Soloviev A (2017) Estimating error
Centrifugevej 356, Kongens Lyngby, Denmark. 2 Geomagnetism, British Geo- statistics for Chambon-la-Forêt observatory definitive data. Annales Geo-
logical Survey, Riccarton, Edinburgh, UK. physicae 35(4):939–952. https​://doi.org/10.5194/angeo​-35-939-2017
Macmillan S, Olsen N (2013) Observatory data and the Swarm mission. Earth
Received: 16 October 2020 Accepted: 11 January 2021 Planets Space 65(11):1355–1362. https​://doi.org/10.5047/eps.2013.07.011
Mandea M, Olsen N (2006) A new approach to directly determine the secular
variation from magnetic satellite observations. Geophys Res Lett
33(15):L15306
Hammer et al. Earth, Planets and Space (2021) 73:54 Page 22 of 22

Olsen N (1997) Ionospheric F region currents at middle and low latitudes esti- Sabaka TJ, Tøffner-Clausen L, Olsen N (2013) Use of the Comprehensive Inver-
mated from Magsat data. J Geophys Res Space Phys 102(A3):4563–4576 sion Method for Swarm satellite data analysis. Earth, Planets and Space
Olsen N, Mandea M (2007) Investigation of a secular variation impulse 65:1201–1222
using satellite data: the 2003 geomagnetic jerk. Earth Planet Sci Lett Sabaka TJ, Tøffner-Clausen L, Olsen N, Finlay CC (2018) A comprehensive
255(1):94–105 model of the Earth’s magnetic field determined from 4 years of Swarm
Olsen N, Mandea M, Sabaka TJ, Tøffner-Clausen L (2009) CHAOS-2—a geomag- satellite observations. Earth Planet Space 70(1):130
netic field model derived from one decade of continuous satellite data. Shepherd SG (2014) Altitude-adjusted corrected geomagnetic coordinates:
Geophys J Int 179(3):1477–1487 definition and functional approximations. J Geophys Res Space Phys
Olsen N, Mandea M, Sabaka TJ, Tøffner-Clausen L (2010) The CHAOS-3 geo- 119(9):7501–7521
magnetic field model and candidates for the 11th generation IGRF. Earth Shore RM (2013) An improved description of Earth’s external magnetic fields
Planets Space 62(10):719–727 and their source regions using satellite data, Ph.D. thesis, The University
Olsen N, Lühr H, Finlay CC, Sabaka TJ, Michaelis I, Rauberg J, Tøffner-Clausen of Edinburgh
L (2014) The CHAOS-4 geomagnetic field model. Geophys J Int Wardinski I, Holme R (2011) Signal from noise in geomagnetic field modelling:
197(2):815–827 denoising data for secular variation studies. Geophys J Int 185(2):653–662
Olsen N et al (2015) The Swarm Initial Field Model for the 2014 geomagnetic Whaler K, Beggan C (2015) Derivation and use of core surface flows for fore-
field. Geophys Res Lett 42(4):1092–1098 casting secular variation. J Geophys Res Solid Earth 120(3):1400–1414
Olsen N, Ravat D, Finlay CC, Kother LK (2017) LCS-1: a high-resolution global Whaler KA (2017) Probing the core surface flow with satellite data, IAGA Joint
model of the lithospheric magnetic field derived from CHAMP and Assembly Session A02—Earth’s core dynamics and planetary dynamos
Swarm satellite observations. Geophys J Int 211(3):1461–1477 (DIV I)
Rogers HF, Beggan CD, Whaler KA (2019) Investigation of regional variation in
core flow models using spherical slepian functions. Earth Planets Space
71(1):19 Publisher’s note
Sabaka TJ, Olsen N, Purucker ME (2004) Extending comprehensive models of Springer Nature remains neutral with regard to jurisdictional claims in pub-
the Earth’s magnetic field with Ørsted and CHAMP data. Geophys J Int lished maps and institutional affiliations.
159(2):521–547
Sabaka TJ, Hulot G, Olsen N (2010) Mathematical properties relevant to
geomagnetic field modeling, in Handbook of Geomathematics. Springer:
Berlin. pp. 503–538