ggad337

Geophys. J. Int. (2023) 235, 1888–1911 https://doi.org/10.
1093/gji/ggad337
Advance Access publication 2023 September 02
GJI Applied and Marine Geophysics
An information theoretic Bayesian uncertainty analysis of AEM

systems over Menindee Lake, Australia
Anandaroop Ray , Yusen Ley-Cooper, Ross C. Brodie, Richard Taylor,* Neil Symington
and Negin F. Moghaddam*
Geoscience Australia, Symonston, Symonston ACT 2609, Australia. E-mail: anandaroop.ray@ga.gov.au
Accepted 2023 August 23. Received 2023 June 30; in original form 2023 January 30
Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

SUMMARY
Long-range, active-source airborne electromagnetic (AEM) systems for near-surface conduc-
tivity imaging fall into two categories: helicopter (rotary-wing) borne or fixed-wing aircraft
borne. A multitude of factors such as flying height, transmitter loop area and current, source
waveforms, aerodynamic stability and data stacking times contribute to the geological resolv-
ability of the subsurface. A comprehensive comparison of the relative merits of each system
considering all such factors is difficult, but test flights over well-constrained subsurface ge-
ology with downhole induction logs are extremely useful for resolution studies. However,
given the non-linear nature of the electromagnetic inverse problem, handling transmitter–
receiver geometries in fixed-wing aircraft is especially challenging. As a consequence of this
non-linearity, inspecting the closeness of downhole conductivities to deterministic inversion
results is not sufficient for studying resolvability. A more comprehensive picture is provided by
examining the variation in probability mass of the depth-wise Bayesian posterior conductivity
distributions for each kind of AEM system within an information theoretic framework. For this
purpose, probabilistic inversions of data must be carried out. Each acquiring system should fly
over the same geology, survey noise levels must be measured and the same prior probabilities
on conductivity must be used. With both synthetic models as well as real data from over the
Menindee calibration range in New South Wales, Australia, we shed new light on the matter
of AEM inverse model uncertainty. We do this using two information theoretic attributes de-
rived from different Kullback–Leibler divergences—Bayesian information gain, and a strictly
proper scoring rule, to assess posterior probabilities estimated by a novel Bayesian inversion
scheme. The inversion marginalizes fixed-wing geometry attributes as generic nuisance pa-
rameters during Markov chain sampling. This is the first time-domain AEM study we know
of, that compares nuisance marginalized subsurface posterior conductivities from a fixed-wing
system, with rotary-wing derived posterior conductivities. We also compare field results with
induction log data where available. Finally, we estimate the information gain in each case via
a covariate shift adaptation technique that has not been used before in geophysical work. Our
findings have useful implications in AEM system selection, as well as in the design of better
deterministic AEM inversion algorithms.
Key words: Non-linear electromagnetics; Inverse theory; Probability distributions.
with the principles of electromagnetic induction, the earth response

1 I N T RO D U C T I O N
is ‘anomalous’ in the presence of electrically conducting bodies.
Airborne electromagnetic (AEM) systems have been in operation Such anomalies have often been associated with base-metal ores or
since the 1940s (Palacky & West 1991), primarily as a ‘bump- saline groundwater accumulations within resistive earth. However,
finding’ tool when qualitatively reviewing data collected over large qualitative approaches are unsatisfactory when trying to quantify the
swathes of land. Since AEM data are acquired after pulsing the earth geoelectric structure of the earth, and such a quantification opens
with an inducing, transmitted electromagnetic field, in accordance the doors to a far wider variety of applications (see Brodie 2010,
for a historical review). To mention only a few, such applications
include geotechnical investigations (e.g. Hodges 1999), mapping
∗ Formerly at: Geoscience Australia, Symonston ACT 2609, Australia. subtle subsurface conductivity contrasts (e.g. Worrall et al. 2001),
1888
C The Author(s) 2023. Published by Oxford University Press on behalf of The Royal Astronomical Society.
Bayesian analysis of AEM uncertainty 1889
palaeovalley mapping (e.g. Eberle & Siemon 2006), hydrogeologi- information gain (Lindley 1956; Chaloner & Verdinelli 1995; Ryan
cal investigations (e.g. Auken et al. 2017), national scale surveys for 2003; Valentine & Sambridge 2020). For the information gain com-
mapping subsurface architecture (e.g. Ley-Cooper et al. 2020) and putation, we use a covariate shift adaptation technique (Sugiyama
mapping detailed river valley aquifer systems (e.g. Minsley et al. et al. 2008b, a) that directly computes probability density ratios
2021b). from posterior and prior samples. We believe that this has not been
To estimate the subsurface geoelectric structure responsible for used before in near surface geophysics, and shows promise in other
the recorded earth response, we need to convert the data from a fields such as geostatistical learning and online learning from time-
time or frequency domain response to subsurface conductivity us- series (e.g. Hoffimann et al. 2021; Chen et al. 2021). Although there
ing Maxwell’s equations and inversion theory (e.g. Parker 1994; have been uncertainty analyses of AEM data using Bayesian meth-
Menke 2012). For meaningful interpretation of the inverted electri- ods (e.g. Minsley 2011; Hawkins et al. 2018; Blatter et al. 2018;
cal conductivity models and their spatial variation in terms of buried Minsley et al. 2021a) and deterministic spatial resolution investiga-
geology, we require knowledge of model uncertainties. In principle, tions (e.g. Bedrosian et al. 2016), we are not aware of studies that
these uncertainties can be found by propagating the data uncertain- have carried out fixed-wing geometry nuisance marginalization, or
ties through the inversion process all the way to the model estimates. compared the resulting subsurface uncertainties with those of a low
However, the electromagnetic inversion process is non-linear and flying helicopter system, while including deterministic inversions

unstable. Consequently, workhorse approaches such as Occam’s in- in the analysis.
version linearize the problem and iteratively produce the smoothest While it is generally desirable to fly low and slow to achieve
possible conductivity model compatible with the data noise (Con- high signal-to-noise ratios, there are challenges to doing so which
stable et al. 1987). There are various similar deterministic inversion range from the technical to the practical. Pastoral activities and
approaches which also produce one single estimate of the subsur- farm animals are disturbed by low flying aircraft, there are safety
face conductivity. They follow other regularization strategies and and pilot fatigue issues with flying low for extended periods of
can also incorporate prior information to produce acceptable con- time, and for rotary-wing surveys there are logistical difficulties in
ductivity models which fit the geophysical data (e.g. Farquharson & ferrying and storing fuel drums within isolated acquisition areas
Oldenburg 1998; Auken & Christiansen 2004). These methods can (such as much of regional Australia). Through synthetic examples
produce linearized estimates of uncertainty, which depend strongly as well as with real AEM data collected over the same stretch of
on the assumptions made in the regularization, as well as the start land, we show that the subsurface information content and inferred
model for inversion. With few exceptions (e.g. Kalscheuer et al. geological model interpretability of both fixed-wing and helicopter
2010), non linear uncertainties are difficult to obtain deterministi- AEM data compare favourably with each other as well as upscaled
cally and are usually found with Bayesian approaches (e.g. Taran- induction log data.
tola & Valette 1982; Mosegaard & Tarantola 1995; Tarantola 2005)
requiring an explicit or implicit declaration of prior earth struc-
ture, followed by posterior sampling using Markov chain Monte 2 FIXED WING AND HELICOPTER
Carlo (McMC, e.g. Minsley et al. 2002; Minsley 2011; Blatter SYSTEMS
et al. 2018; Ray & Myer 2019). An additional challenge for AEM
The details of casting the inverse problem to solve for system ge-
inversion is that the acquisition system geometries may in some
ometries in time domain systems can be found in chapters 2 and
cases not be well known—leading to inaccuracies in the forward
5 of Brodie (2010). The most important geometry elements for a
modelling of geophysical data from estimated conductivity models.
fixed-wing system were found to be the receiver pitch, as well as
As detailed by Brodie (2010), ambiguity in the transmitter or re-
the Tx–Rx inline (horizontal) and vertical separations (Fig. 1). It
ceiver position as well as their roll, pitch and yaw, will creep into
was also ascertained that sufficient information for geometry esti-
the estimates of subsurface conductivity if they are not accounted
mation is not found within the AEM provider supplied secondary
for in the inversion process. This is especially true of fixed-wing
field alone. The removed nominal primary field must be added back
AEM systems where the transmitter (Tx) is usually centred on the
into the secondary field, and the total field must be inverted for both
aircraft itself, and the receiver (Rx) ‘bird’ is towed some distance
geometry parameters as well as the earth conductivity model. In
behind the aircraft. While we are not interested in transmitter or
this work, instead of parametrizing the vertical Tx–Rx separation,
receiver rotations and positions in themselves, except as a quality
we will focus on the height of the Rx bird (zRx), as we assume
check for the data and for safety purposes, they can be treated as
that the height of the Tx (i.e. aircraft height) is well known. For
nuisances during deterministic inversion as described by Brodie
helicopter systems (Fig. 1) we will assume that the Tx loop frame
(2010). However, these geometry parameters trade-off amongst
is rigid and the Rx height relative to it is known and fixed, as is the
themselves, as well as with the subsurface conductivity since they all
case for many commercial systems. These assumptions are based
contribute to changes in the modelled electromagnetic field. These
on historical data stemming from our acquisition of hundreds of
trade-offs make it even more difficult to correctly quantify the un-
thousands of line-km using the systems described above.
certainty associated with the electrical conductivity profiles of the
A recent development for fixed-wing systems, is that we treat as
subsurface.
data the amplitude of the joint X- and Z-component magnetic fields,
In this study, through nuisance marginalization we obtain the
that is
posterior inverse model uncertainties of a technically mature, time-
domain fixed-wing system. Using information theoretic principles Bamp = Bx2 + Bz2 , (1)
based on the Kullback–Leibler divergence (henceforth abbreviated
as KLD, Kullback & Leibler 1951), we rigorously compare the in- where Bamp is a scalar. However, we note that Bamp is observed as
verse uncertainty of the fixed-wing system with that of a technically a time-series at various ‘channels’ so both Bamp and Bx , Bz will be
mature, time-domain helicopter system. The KLD naturally leads to considered as vectors henceforth, in the linear algebra sense. For a
the formulation of the logarithmic score, (a strictly proper scoring layered Earth excited by a vertical magnetic dipole (i.e. a horizontal
rule, see Good 1952; Gneiting & Raftery 2007) as well as Bayesian current loop transmitter), there are no azimuthal components to the
1890 A. Ray et al.

Figure 1. Helicopter (left-hand panel) and fixed-wing (right-hand panel) AEM systems. Helicopters typically fly lower and slower, and if the transmitter
airframe is rigid, typically do not require geometry nuisance inversion. Fixed wing aircraft fly higher and faster, and the geometry nuisance parameters of
importance have been marked in red.
Figure 2. Noisy synthetics with nominal survey specifications and noise levels for a fixed-wing time-domain AEM system. The data correspond to the X- and
Z-component magnetic fields recorded at each time channel. The synthetic model comprises 65 layers and is based on induction logs from an aquifer system
in the Permian basin, TX, USA.
observed magnetic field (e.g. Loseth & Ursin 2007). As a conse- where σ Bx and σ Bz are the data errors in the X- and Z-components,
quence, many AEM providers do not usually provide Y-component respectively.
data. The advantage of dealing with Bamp as opposed to Bx and
Bz jointly, is that irrespective of the rotation of the X and Z coils
in the X–Z plane (i.e. regardless of receiver pitch), the amplitude 3 H I E R A R C H I C A L B AY E S I A N
of the joint vector field remains invariant. This obviates the need SAMPLING OF AEM NUISANCES AND
to invert for the Rx pitch, reducing the number of unknowns in E A RT H P R O P E RT I E S
the nuisance estimation. Of course, this comes at the expense of
We now return to the matter of trade-offs between system geometry
subsurface information which the individual X- and Z-components
and earth conductivity. Using Bayesian inference in a hierarchical
provide in a conventional joint inversion. However, as we will see,
setting (e.g. Gelman 2006), we can estimate distributions over pa-
there is not an appreciable difference with the resolving capabilities
rameters we are not interested in, to ensure that inferences over
of the joint inversion, and the amplitude only inversions potentially
parameters of interest are unbiased. Nuisance estimation in this
remove troublesome conductivity artefacts at depth (Ley-Cooper &
manner has a long and rich history in the Bayesian geophysical lit-
Brodie 2020). Using the theory of propagation of errors, assuming
erature: for traveltime inversion to estimate data noise (Malinverno
independence of the data errors in the X- and Z-components, the
& Briggs 2004), in geoacoustics to estimate source waveforms and
data error in Bamp at each time channel can be derived from (1) as:
data noise (Mecklenbrauker & Gerstoft 2000; Dettmer et al. 2010),
with receiver functions to parametrize the likelihood (Bodin et al.
1 2 2
2012), with magnetotellurics to estimate water column conductiv-
σ Bamp = Bx σ Bx + Bz2 σ B2z , (2) ities (Blatter et al. 2018), to illustrate a few examples. For AEM
B amp
The likelihood function p(d|θm ) for Gaussian data noise can be

written as:
⎛ ⎞
t
1 ⎝ 1
L(θm ) = p(d|θm ) = √ exp − f(θm ) − d Cd f(θm ) − d ⎠, (4)
−1
|2πCd | 2
where [f (θm ) − d] is the residual vector of misfit for the model θm ,

between the forward model calculations for a given set of earth
conductivities as well as nuisances, and the AEM data. The covari-
ance matrix of data errors is given by Cd . A Gaussian likelihood is
generally justified by the application of stacking to increase signal-
to-noise ratios in geophysics, and AEM is no exception. Stacking
implies Central Limiting for the resulting noise estimates on the
mean data, and hence the implication of Gaussianity for the like-
lihood function. The data error or noise model used throughout

this work is based on the noise model given in Green & Lane
(2003). The total noise is assumed to be due to two independent
Figure 3. Noisy synthetics with nominal survey specifications and noise sources: multiplicative noise proportional to signal amplitude, as
levels for a dual moment, time-domain helicopter system, for the same well as high-altitude noise measured away from the effects of Earth
aquifer model as shown in Fig. 2. The data correspond to voltage in the Z conductivity, added together in quadrature. In AEM parlance, the
receiver coil to measure changes in the magnetic field due to two independent
misfit measure φ d is often used, with values close to 1 indicat-
exciting transients.
ing a reasonable fit to within data noise. The more commonly
applications, Minsley (2011) has considered reported transmitter used measure of RMS √ misfit in the broader geophysical EM lit-
heights as uncertain and sampled them as a nuisance. However, ours erature is given by φd , where the likelihood and φ d are related
is the first application we are aware of, which treats the rotations as:
of the transmitter and receiver, as well as the transmitter–receiver p(d|θm ) = √ 1 exp − n data
φd ( θ m ) , (5)
|2π Cd | 2
horizontal and vertical separations, as nuisances to be sampled for
t
a fixed-wing system. Possible reasons such literature is lacking are ⇒ φd ( θ m ) = 1
f(θm ) − d C−1 f(θm ) − d , (6)
n data d
that handling passive and active frames of transmitter and receiver
rotation is non-trivial (e.g. Fitterman & Yin 2004; Key & Lockwood or in other words, the χ 2 data error is φ d times the number of data
2010), modifying inline separations requires careful changes to pre- ndata .
allocated Hankel filters for forward evaluations (see Key 2012, for To be explicit about the model parametrization, we separate θm
details), and that a generic McMC inversion code which can han- into the earth resistivity θ ρ and the geometry nuisance parts θ n as
dle an extensible number of nuisances is not easily available. As follows:
part of this work, we have extended the generic trans-dimensional
Gaussian process (TDGP) McMC package (Ray & Myer 2019; Ray θ m = [θ ρ , θ n ]. (7)
2021) to handle AEM forward models based on the time domain Resistivity is the inverse of electromagnetic conductivity, and
AEM codes of Blatter et al. (2018) and the rotation formulations to span the various orders of magnitude of earth resistivity we
given in chapter 2 of Brodie (2010). As these nuisances and earth parametrize θ ρ through the base-10 logarithm of linear resistivities.
model parameters (EM conductivity) are sampled jointly by the For sampling θ ρ , we use a reversible jump sampler (Green 1995)
McMC, all interparameter dependencies including non-linearities or trans-dimensional (trans-D) McMC as it is often referred to in
can be captured. By marginalizing over the posterior nuisance pa- geophysics (Malinverno 2002; Sambridge et al. 2006; Bodin &
rameters, the posterior conductivities are freed from bias. Within the Sambridge 2009; Dettmer et al. 2010; Ray & Key 2012; Gehrmann
permissible prior ranges of the geometry nuisances, corresponding et al. 2015; Ghalenoei et al. 2021). However, instead of using a
earth conductivities within their prior ranges are extensively sam- piecewise constant stair-step parametrization as is usually done for
pled according to their likelihood as determined by corresponding 1-D, we use a Gaussian process basis (GP) as described in (Ray &
data misfits. All of this is done through the McMC in accordance Myer 2019). In addition to TDGP being spatial-dimension agnostic
with Bayes’ Theorem as detailed below. (e.g. Blatter et al. 2021), we have found it to be particularly well
suited to diffusion problems which require smooth parametriza-
3.1 Posterior sampling and parametrization tions. Sharp changes can be well represented with two GP inference
layers as detailed in Ray (2021). However, we have found that AEM
For observed data d and models θm it can be written that: data do not generally support such sharp changes in the Earth and
we have opted to use the vanilla TDGP, that is trans-D McMC
p(θm |d) ∝ p (d|θm ) p (θm ). (3)
with birth and death of stationary GP nuclei (where stationarity
Reading from right to left, p(θm ) is the prior probability of θm , which implies an unchanging GP length scale over the model domain) in
is known independent of the observations d. The prior importance a reversible jump framework as described in Ray & Myer (2019).
of θm is re-weighted by carrying out an AEM sounding which shows Specification of the prior for θ ρ is exactly the same as shown in
how likely it is that θm fits the observations. The probability of fit section 2.5.2 of Ray (2021) for a resistivity model and will not be
is provided by the likelihood function p(d|θm ). The result of re- repeated here. Priors for θ n are uncorrelated and uniform, ± a few
assessing or updating the prior importance of θm by the likelihood metres or degrees off from what is recorded during the AEM sur-
of θm provides the posterior probability of observing the model θm . vey. Particulars of prior probabilities are provided in the applications
This posterior probability is represented by the term p(θm |d). section.
1892 A. Ray et al.

Figure 4. Synthetic sounding conductivity posteriors with depth for (a) Fixed wing: conventional Bx and Bz joint inversion, (b) Fixed-wing: inversion of
B field amplitude in receiver X–Z plane and (c) Helicopter: dBz /dt inversion. Hotter colours are more probable, and the true model has been overlain with
a thick black line. The 5th, 50th and 95th percentiles of posterior conductivity are shown with dashed black-and-white lines. The deterministic inversion
result has been plotted with a thick red-and-white line. (d) Information theoretic divergence analysis: True model with prior probabilities in the background
(left-hand panel), logarithmic score for each AEM system/inversion and probabilistic inversion (centre panel) and Bayesian information gain (right-hand
panel). High logarithmic scores and high information gain indicate better adherence to the truth and posterior resolution, respectively. For each inversion
type, the summary plots at the bottom indicate 5th, 50th and 95th percentiles from scores and gains across all depths, with the filled triangle indicating the
median.
The process of finding the posterior probability p(θm |d) for vari- proposals, when proposing from the prior, and for uniform pri-
ous models θm admissible by the prior is repeated until an ensemble ors over the number of parameters, eq. (8) also holds for both
of models representative of the posterior probability density func- fixed-dimensional or reversible jump Metropolis-Hastings–Green
tion (PDF) p(θm |d) is obtained. Sampling proportional to the pos- McMC (Metropolis et al. 1953; Hastings 1970; Geyer 2011). This
terior probability is carried out by using the following acceptance is what we have used for sampling geometry nuisances and earth
probability α to move from an Earth resistivity model vector or conductivities. Due to this choice of proposals, the move proba-
nuisance model vector θ to proposed model θ in the McMC chain: bility terms q(.) never explicitly figure in calculation of the ac-
ceptance probability term α in our algorithm (e.g. Agostinetti &
⎡ 1/T ⎤ Malinverno 2010; Dosso et al. 2014). The exponent 1/T in (8) is an
L(θ )
α(θ |θ) = min ⎣1,
⎦. (8) annealing factor for parallel tempering (Swendsen & Wang 1987)
L(θ) as described in Dettmer & Dosso (2012). Parallel tempering sig-
nificantly accelerates the convergence to the posterior distribution
We note here, that for a uniform prior over the number of nuclei and (Dosso et al. 2012, 2014; Sambridge 2013) and is used by de-
when proposing from the prior resistivities for birth and death, for fault in TDGP. The entire McMC algorithm encapsulated within a
all TDGP moves, (8) provides the acceptance probability. All the parallel tempering framework is described in Algorithm B within
TDGP move proposal probabilities q(θ, θ ) are exactly the same as Appendix B.
described in detail in Ray (2021). For symmetric fixed-dimensional

Figure 5. Nuisance geometry posterior scatter plots and histograms for fixed-wing AEM: conventional Bx and Bz joint inversion. (a) Detailed view, with red
symbols indicating deterministic inversion estimates and black symbols denoting true values. (b) Zoomed out view, this time with dashed lines showing prior
bounds.
4 S Y N T H E T I C S T U D Y: I N F O R M AT I O N and heights are sourced from the Menindee calibration range AEM
T H E O RY A N D M A R G I N A L I Z AT I O N O F flights (Barlow 2019) that we will report on in a later section. As
G E O M E T RY N U I S A N C E S in the real data example, the fixed-wing AEM system and the he-
licopter AEM system compared are widely used and technically
While ground truth is in principle, the ultimate arbiter of the ac-
mature. Both have been found fit for purpose for various surveys
curacy of a geophysical investigation method, a synthetic study at
we have undertaken (e.g. Ley-Cooper 2021, 2022) on the basis
one sounding location, with noise levels, Tx–Rx geometries and
of competitive bidding from various commercial entities that have
flying heights typical of the systems under consideration is instruc-
included, but not been limited to these two systems.
tive. The transmitter waveforms, noise levels, nominal geometries
1894 A. Ray et al.
make exact comparison impossible. However, the surface geophys-

ical data should still reflect well-constrained geology and property
trends measured in the drillhole. We ask the reader to keep this in
mind in the context of further references to ground truth or induc-
tion logs in the text. For those wishing to investigate an additional
synthetic model, another example based on a real well-log featuring
a thin near-surface conductor and a deeper resistor can be found in
Appendix E.
Posterior uncertainties for both systems are shown in Figs 4(a)–
(c). Note how the deterministic inversions almost always lie within
the 90 per cent posterior credible intervals (CIs), the region be-
tween the 5 and 95 per cent posterior percentiles. Further, while
some deterministic inversion results are closer to the true value,
the posterior uncertainties paint a different picture of uncertainty.
The uncertainty with depth, qualitatively given by the width of the

Figure 6. Nuisance geometry posterior scatter plots and histograms for CI, is not markedly different between the amplitude only and heli-
fixed-wing AEM: Inversion of B field amplitude in receiver X–Z plane. copter AEM posteriors. As expected, the posterior uncertainty for
Again, red and black symbols denote deterministic and true values, respec-
all three inversion types is least around the conducting ‘layer’ be-
tively. Prior bounds are shown with dashed lines. Compared to the joint
inversion posteriors in Fig. 5, note that more of the prior space within the
tween 50 and 80 m depth. In fact, for this model, it would appear
dashed lines has been occupied by the posterior samples. that the fixed-wing ‘conventional’ inversion does well overall, with
the width of the CI remaining roughly similar over all depths except
The fixed-wing system (Fig. 1) compared was TEMPEST (Lane at the conductor. The helicopter AEM posterior however, seems to
et al. 2000), with the transmitter loop placed 120 m above the suggest that the top and bottom of the conductor can be separately
ground, and the Tx–Rx vertical separation was set to 38 m, that resolved. It must be noted that we avoided inverse crime: the for-
is the receiver antennae were placed at zRx = 82 m above the ward modelling grid for both the deterministic and probabilistic
ground. The Tx–Rx horizontal separation or hsep was set to 113 m, inversions was identical, but coarser (only 50 cells) than the true
with Rx pitch set to 2◦ . For the helicopter AEM comparison, we 65 cell modelling grid used to compute the synthetics. This was
used a SkyTEM 312 system based on developments of the system to emulate a measure of realism, as the true discretization of the
originally described in Sorensen & Auken (2004). For the helicopter earth is never known. As is the usual case, we overparametrized
system (Fig. 1), the transmitter loop was placed 40 m above the the modelling grid for the inversion, and an Occam regularization
ground, and the receiver was rigidly fixed at a vertical separation of with a return to a resistive reference or ‘prejudice’ model (see Key
2 m above the transmitter, at a radial distance of 13 m from the loop 2009) was used in the deterministic inversions. For the probabilistic
centre, inline to the flight direction. inversions, the prior parameter resistivity bounds for all the deter-
The synthetic earth model used for the comparative study is a re- ministic inversions were uniform between the extremal limits of the
alistic earth model for shallow hydrological investigations, based on log10 conductivities shown in Figs 4(a)–(c). A correlation length of
variations we would expect in a well-log. It features 65 layers with 2 thickness units was used, and the birth/death trans-D GP method
many small conductivity contrasts, one large conducting ‘layer’ be- was allowed to place a maximum of 40 GP nuclei uniformly between
tween 50 and 80 m depth and a resistive trend with depth. Forward the top and bottom of the model. The resulting GP nuclei resistivity
modelling and noisy data for this earth model using both systems values are interpolated onto the 50 cell grid which is fine enough
can be seen in Figs 2 and 3. We ask readers to note that from here to model accurately the geophysical features of interest in a spatial
on, we always have conductivity increasing to the right and resis- dimension-agnostic manner (e.g. Blatter et al. 2021). The correla-
tivity increasing to the left of all our resistivity axes. Further, we tion ensures that any thickness unit has an exponentially decreasing
represent resistivity in log10 units, in which a conductivity in S m−1 correlation with interpolated resistivities two thickness units above
or resistivity in m are simply the negative of the other, being in- and below it. Although it may appear that infinitesimally thin layers
verses. There are undoubtedly differences in the spectral content of are desirable to model, they are largely unresolvable when using
the source waveforms for both systems, especially given that the diffusive electromagnetic physics as has been described in detail
helicopter system considered here is a ‘dual moment’ system with by Ray & Myer (2019) and Ray (2021). The GP kernel used to
two exciting transients that are interleaved together after acquisi- impose the correlation is the sidelobe averse Ornstein–Uhlenbeck
tion. It is also impossible to examine the uncertainty of inverted kernel, which is described in detail by Rasmussen & Williams
subsurface conductivity structures at all possible wavenumbers at (2006).
all exciting frequencies. However, the utility of this synthetic study
is twofold. First, it is not a closest-to-true model contest, as posterior
uncertainties around both synthetic soundings will be inspected—
4.1 Information theoretic analysis of posterior
lower uncertainties in conductivity at depth imply greater resolution.
conductivities
Secondly, it sets the stage for examining posterior subsurface un-
certainty for real data with hundreds of soundings along a flight line While we can gather a useful, qualitative idea of posterior uncer-
with variable geology and established ground truth. However, we tainty by inspecting Figs 4(a)–(c), we have also provided a rigor-
must make clear that practically speaking, we never have exact truth ous and quantitative information theoretic analysis using Kullback–
even when we have logged the earth within a drillhole. The effects Leibler divergences (Fig. 4d). While the general geophysicist can
of drilling mud, invasion, calibration errors and temporal (climatic follow the remainder of this section without specialized knowledge,
and seasonal) changes between the acquisition of downhole induc- we refer those interested in the details to Appendix D for a math-
tion log conductivities and the above surface EM data acquisition, ematical discussion with proofs. With this in mind, we look at two

Figure 7. Test AEM line over Menindee Lakes, New South Wales, Australia, with logged boreholes shown in red. Note the transition from an arid Earth
surface to the lake, midway along the test line.
divergences, as they are known, which can heuristically but not recent use of logarithmic scores can be found in Seillé & Visser
mathematically, be thought of as ‘distances’ between probability (2020) for selecting an optimal likelihood function, and Friedli
density functions (see Beier et al. 2002). The first such divergence et al. (2022) for evaluating different McMC proposal schemes for
leads to a strictly proper scoring rule (Gneiting & Raftery 2007), a challenging high-dimensional inverse problem. This brings us to
the logarithmic score, dating at least to Good (1952). In essence, the second divergence, known as the Bayesian information gain,
for a forecast density p( · ), say a posterior density, we assign a which represents our increased knowledge of the subsurface with a
score log p(x) when the event x is actually observed. As shown with (usually) narrower posterior density than the prior density we began
eqs (D9)–(D11), on average the highest score will be obtained for a with. In other words, the information gain represents the dimin-
forecast probability density that equals the true probability density ishing overlap between the prior and posterior densities, a natural
of the observations, even if the true density is unknown. Naturally, proxy for Bayesian resolution as shown by Blatter et al. (2018) for
such scoring systems have found heavy use in forecasting and al- an Antarctic AEM survey. To examine these divergences we begin
lied decision theoretic fields such as meteorology, quantitative eco- by showing the true model and prior resistivity probability density
nomics and finance, psychology and optimal energy usage (see Car- in the leftmost column of Fig. 4(d). As discussed in Ray & Myer
valho 2016,for a review). In near surface or exploration geophysics, (2019), though the McMC model parameter priors are uniform,
1896 A. Ray et al.

Figure 8. Menindee test range: Dual moment helicopter EM system inversion. The top row shows the mean sampled φ d and one standard deviation. The next
three rows show the 5th, 50th (median) and 95th percentiles of posterior conductivity. Where all three percentiles are similar, the posterior conductivities are
more probable. The display of percentiles and their spread simplifies the task of assessing interpretive uncertainty, as discussed in detail in the text. The dashed
vertical line is nearest to a borehole with induction logs to be examined later in the text.
the resulting resistivities interpolated by the GP parametrization with the logarithmic score, and is discussed in detail by (Bröcker &
are not. However, similar to a bounded uniform distribution, the Smith 2007). The negative of the logarithmic score is known as ig-
interpolated prior resistivities do not have a focussed mode. norance, and (near) zero prediction probability of the truth implies
We follow Seillé & Visser (2020) and treat the posterior marginals (high) infinite ignorance. This situation could easily occur as shown
as a forecast probability density, and the true model as the eventu- above, or for extremely resistive media and inductive EM methods,
ating observation. We then fit a kernel density (Sheather & Jones as these methods are not sensitive to highly resistive media. Barring
1991) pz ( · ) to the marginal posterior samples at depth z and evalu- the outlier, the helicopter inversion does slightly better up shallow,
ate the logarithmic score log pz (mz ) for the true log10 resistivity mz and the fixed-wing joint inversion scores better deeper.
at depth z. In the middle column of Fig. 4(d), we plot the logarithmic In the rightmost column of Fig. 4(d) we show the Bayesian in-
score for each system, with higher scores indicating better repre- formation gain, by calculating the KLD directly from the posterior
sentation of the truth. There is however, a large amount of overlap and prior marginal samples at every depth. For this purpose, we
summarized by the quantile plots at the bottom, showing the overall use a covariate shift method (Sugiyama et al. 2008b), as detailed in
5th, 50th and 95th percentiles of the score. Since there is only one Appendix D. Such methods are adept at quantifying a shift between
score at each depth per inversion type, the summary percentiles samples from two probability densities, such as samples from the
are calculated from values across all depths. For decision theoretic prior and the posterior, or when there are sudden changes in online
problems, the forecast with the highest expected score is usually streaming data (Chen et al. 2021). A recent use of sample based
preferred (Diks et al. 2011). However, there is also a notable outlier covariate shift methods for geostatistical transfer learning can be
score for the helicopter system which will skew the average score found in Hoffimann et al. (2021). The information gain focuses
for that system downwards. This is because the posterior probability very minutely on the overlap (or the lack of it) between the poste-
for all three systems at the bottom end of the true conductor is very rior and prior probability densities. As shown in Appendix D, it is
small. However, it is not that the helicopter system does worse at always positive and only zero if the prior and posterior densities are
localizing the conductor, there is a narrow high probability region identical. To first order, the information gain is small when poste-
only slightly to the left, within half a conductivity decade (1 decade rior widths are large, and large when posterior widths are narrow,
= 1 log 10 unit of resistivity or conductivity) at the outlying depth such as within the conductor between 50 and 80 m depth. This
of ∼80 m in Fig. 4(c). This problem with locality is a known issue is an information theoretic counterpart of deterministic sensitivity,

Figure 9. Menindee test range: Fixed-wing AEM system conventional joint Bx and Bz inversion. The top row shows the mean sampled φ d and one standard
deviation. The next three rows show the 5th, 50th (median) and 95th percentiles of posterior conductivity. Where all three percentiles are similar, the posterior
conductivities are more probable. With the exception of the deep conductors at −200 m, the first-order geology of the lake system is very similar to what is
shown by the helicopter EM system in Fig. 8. The dashed vertical line is nearest to a borehole with induction logs to be examined later.
as we know that TE mode inductive EM sources (Loseth & Ursin this gain from 2–3 to 4 bits is not significant. Each bit of infor-
2007) are sensitive to conductors. Examining in detail, we see that mation gain leads to reduction of half the prior probability mass
the information gain both above and below the conductor is small, (Pinkard & Waller 2022) or equivalently, doubles the concentration
but is slightly larger at shallow depths for all three inversion types. of probability mass in the posterior. Since 80 per cent of posterior
This is as we would expect, given signal-to-noise considerations at conductivities are between 0.8 and −0.26 log10 m for the fixed-
late times as well as conductive shielding effects in electric media. wing system inversions at ∼70 m, halving the posterior probability
We would particularly like to draw attention to the fact that within mass does not add significantly to inferred knowledge of a con-
the conducting body itself, unlike for the logarithmic score, the in- ductor. For all three inversion types, at this depth we are able to
formation gain remains large. This is due to the information gain bracket a 0 log10 m conductor within less than half a decade, in a
not suffering from the aforementioned locality problem. In fact the prior range spanning nearly 4 decades of resistivity. Underneath the
helicopter system has a tight posterior distribution at ∼80 m depth conductor, all three inversions indicate a return to resistive geology
within half a decade of the true value, accordingly it has the highest in the 2.7 to 0.1 log10 m range, with the fixed-wing/joint inversion
information gain. indicating a slightly narrower high probability region. Finally, the
In the absence of ancillary information, no inversion/system box plots showing the 5th, 50th and 95th percentiles of informa-
would clearly outmatch the others. In all cases, from the marginal tion gain for each system and their large overlap, are shown at the
posterior probabilities of resistivity with depth, we would interpret bottom of the rightmost panel in Fig. 4(d). Again, there is only one
the following. Starting from the top: resistive geology, between value of information gain at each depth for every inversion type,
2.7 and 0.8 log10 m, followed by a conductor starting at ∼50 m consequently the summary percentiles are calculated from values
depth. Owing to conductivity-thickness trade-offs the fixed-wing across all depths. As before, we see a large overlap and are able to
systems/inversions do not narrow down the conductor bottom well, confirm our earlier first order inspection of posterior quantiles with
while the helicopter estimate of the bottom is slightly too shallow. this analysis. This is useful to note since most geophysicists or ge-
All three systems have high probability mass in the conductor be- ologists will not have computed information theoretic divergences
tween 50 and 80 m in the vicinity of 1 m or 1 S m−1 (i.e. 0 in readily available. Another useful measure of the overlap between
log10 ). While we could point to information gain within the conduc- probability densities is the Bhattacharyya distance and related co-
tor being higher for the helicopter system—contextually speaking, efficient (Bhattacharyya 1943), which has been used in geophysical
1898 A. Ray et al.

Figure 10. Menindee test range: Fixed-wing AEM system amplitude only inversion. The top row shows the mean sampled φ d and one standard deviation. The
next three rows show the 5th, 50th (median) and 95th percentiles of posterior conductivity. Again, with the exception of the deep conductors at −200 m in
Fig. 9, the first-order geology of the lake system is very similar to what is shown by the helicopter EM system in Fig. 8. The dashed vertical line is nearest to a
borehole with induction logs to be examined later.
work (Subašić et al. 2019), though it does not have a straightforward and we do not do so here either. We surmise this is because the
Bayesian interpretation. helicopter AEM system studied here, which we have observed in
In conclusion, if it were for this particular synthetic model, the operation, has a rigidly mounted Rx coil. The aerodynamically sta-
choice of system from a technical standpoint is largely equiva- ble Tx–Rx frame and its height are known well enough for this
lent. We can say this as the CI widths are similar noting the vari- not to be inverted for. For conventional fixed-wing inversions (Fig.
ations discussed above, and neither the scoring rule nor informa- 4a), in addition to the earth conductivity, three more parameters
tion gain indicate without qualification, a superior system/inversion need to be sampled (Fig. 5): the receiver pitch, the horizontal
type. Tx–Rx separation (labelled Tx–Rx hsep) and the vertical Tx–Rx
separation (equivalently, we have kept fixed the Tx height, and in-
verted the Rx height, labelled zRx). We make observations of note
underneath:
4.2 Inversion details and nuisance sampling In Fig. 5(a) the most probable nuisance model values and the
All inversions converged to a rms value of 1. The probabilistic in- truth do not coincide. Although other workers have encountered
versions for both kinds of fixed-wing inversion were run with 7 similar phenomena (e.g. Dettmer et al. 2015, most probable ver-
log-spaced parallel tempering chains, and for helicopter AEM in- sus true values in their fig. 4), we decided to investigate further.
version with 5, with a maximum annealing temperature of 2.5. A We ran the Bayesian inversion for 200 000 more samples with the
greater number of parallel chains are required for the fixed-wing same noise realization as in Fig. 2, then did an independent run for
inversions, as the inference problem with geometry nuisances is 1 000 000 samples. The longer run and independent restarts with
harder to sample. We had initially achieved stationarity well within the same data as in Fig. 2 persistently produced histograms where
400 000 samples in the target McMC chain at T = 1. However, the true nuisance values are in the tail region. Further, the poste-
to sample the near-zero probability regions and establish stable rior conductivities among these runs are virtually indistinguishable
score estimates that avoid infinite ignorance, we ran each inversion from Fig. 4(a). When compared with the prior extents (Fig. 5b), the
type for 1 000 000 samples, discarding the first fifth to preclude posterior distributions do appear more generally in the vicinity of
the possibility of biased inference. Within the legacy survey noise the true values. Given that the likelihood (Fig. 4) depends on the
levels we have accumulated over the years, inverting for height data noise—on running with a different random noise realization,
has not made a significant difference to the conductivity model we did indeed observe coincidence of true geometry values with

Figure 11. A comparison between probabilistic inversions (top three rows), deterministic inversions (bottom three rows) and induction logs (boxed, thin, tall
rectangles on every row). The distance from the AEM flight line, together with the name of the well is shown once for fixed-wing and once for helicopter
EM data. For example, the deepest logged borehole BHMAR23-1 is about 35 m away from both the helicopter and fixed-wing flights and 8 km from the
southeastern end of the line. All rows show good general agreement with the wells, despite the logging and flying having occurred several years apart.
high probability regions, resulting in minor differences with the of this prior volume is of posterior importance (i.e. a large part
posterior conductivities presented in Fig. 4a). of the boxed region with dashed lines in Fig. 5b is empty model
We are extremely sensitive to the Tx–Rx separation, but receiver space). Finally, the posterior sampling surface is quite rugged (i.e.
pitch and receiver height trade-off near linearly. This intuitively fat tails with sharp jagged dropoffs) as can be seen from the zoomed
makes sense—if the pitch decreases (antenna axis along flight line in crossplots and marginals for both pitch and zRx in Fig. 5(a).
tilts upwards) this could be compensated by the antenna origin being Most real data McMC AEM inversions converge to stationarity
translated closer to the ground. However, this implies that we cannot well within 200 000 samples, but in order to draw robust conclusions
resolve both receiver height and pitch, and perhaps the information from synthetics we have massively oversampled as described above.
contained in one ought to be enough for the inverse problem—with Parallel and high performance computing (HPC) considerations
a suitable rotation of the nuisance coordinate axes. Unfortunately, during sampling have been laid out in Appendix C
this rotation of the axes is data dependent and we need to do an initial For an amplitude only inversion (Fig. 4b) of the same data, we
sampling run to estimate a principle components rotation. However, do not need to estimate the Rx pitch and the posterior nuisances
we have found that it is indeed much more efficient to sample along are shown in Fig. 6. Immediately, we see that the fraction of prior
such rotated parameter axes as shown by Yardim et al. (2006) for volume required to solve the problem is greater and the poste-
radio-refractivity inversion and Dosso & Dettmer (2011) for geoa- rior surface is less rugged, denoting that the posterior is easier
coustic inversion. An efficient alternative for sampling nuisances to sample. While the deterministic estimate of the receiver height
could be pseudomarginal methods and their correlated variant (see nuisance parameter is outside the prior range, the deterministic con-
Andrieu & Roberts 2009; Friedli et al. 2022, for details). ductivity model, which is ultimately the earth feature of interest, is
Though the nuisance prior bounds are based on what we would within the 90 per cent CI as can be seen from Fig. 4(b). Undoubt-
expect for errors from the onboard inertial measurement unit (IMU) edly, tweaking the regularization and constraints for the determin-
sensor and variability within a flightline—only a very small portion istic nuisances can lead to ‘better’ estimates of conductivity. Since
1900 A. Ray et al.

Figure 12. Information gain at the Menindee test site for a) all system/inversion combinations. (b) The median information gain with depth across the surveys,
with summary bars showing the 5th and 95th percentiles at every depth. Since geology can be quite variable at the same depth, the interval between the 5th and
95th percentiles simply reflects this variability with depth. (c) The same as (b) but zoomed into the location nearest to BHGT14M1 with dashed lines indicating
the information gain at that sounding location.
we do not know the true earth model for real data scenarios, we and partly over a shallow ephemeral lake, co-incident with or very
have not opted for such tweaking to keep the synthetic exercise near boreholes with induction logs (Fig. 7). As will be shown later,
meaningful. downhole conductivities from these logs provide a useful compar-
ison of inversion results with ground truth. This in turn allows us
to assess within the limits of temporal (seasonal or climatic) vari-
5 M E N I N D E E C A L I B R AT I O N L I N E , ation, the accuracy with which different AEM systems image the
N E W S O U T H WA L E S : C O M PA R I S O N subsurface. It is unusual for the same survey line to be systemat-
W I T H B O R E H O L E S A N D I N F O R M AT I O N ically flown repeatedly by vendors (see Minsley et al. 2021a, for
GAIN another example of overflown lines), especially in the presence of
In the Broken Hill region of New South Wales, Australia, lie well-constrained geology and induction logs. Hence the Menindee
Menindee Lakes. Over one of these lakes we operate an AEM test range has become a valuable proving ground for AEM technol-
testing range. A 12-km-long flight line lies partly over arid ground ogy. This holds true for testing mechanical features and electronic

Figure 13. Posterior distributions of conductivity with depth at borehole BHGT14M1 for (a) conventional fixed-wing Bx and Bz joint inversion, (b) fixed-wing
amplitude only inversion and (c) helicopter dBz /dt inversion. The conductivity log has been overlain in cyan. Note the good agreement between the upscaled
well-log (thin green line) with the variation in the credible interval as well as with the the median posterior conductivity. The well and the CIs both support a
return to conductive media with depth as well as division into three zones discussed in the text.
Figure 14. Data fits for 100 randomly selected posterior models at well BHGT14M1 for (a) conventional fixed-wing Bx and Bz joint inversion and (b) fixed-wing
amplitude only inversion and (c) helicopter dBz /dt inversion.
instrumentation of the AEM systems, as well as for inversion codes the test line. The closest location to a well studied borehole which
and modelling theory. intersects a known near surface regional conductor has been marked
In Fig. 8, we show the results from probabilistic inversions of in all three figures for closer inspection later in the text. First of all,
all helicopter EM soundings inverted independently along the test it is apparent from Figs 8–10 that all three inversions show remark-
line. At each sounding, McMC was carried out for 400 000 samples ably similar posterior uncertainties, with exactly the same priors on
on multiple parallelly tempered interacting chains, with the first conductivity, using the same geometry constraints for fixed-wing
80 000 samples discarded in the burn-in phase. The first row from data as in the synthetic studies, all the while using measured high-
the top shows the mean square misfit or φ d , with a reasonable fit altitude noise levels for the test flights. From the northwest to the
to data given by the dashed line at φ d = 1. The second, third and southeast, as we descend into the lake bed, the lake clearly shows
fourth rows from the top show the 5th, 50th and 95th percentile up in all percentiles as a resistive structure relative to its surround-
posterior conductivities with depth at every sounding. Similar to ings. To the southeast, in the near surface, there appears to be a
the synthetic studies shown in Fig. 4, wherever the percentiles show layer of clays which show up as conductive. It must be noted that
similar values, the CIs are narrow and hence imply greater posterior the helicopter system was flown in 2015 and the fixed-wing system
certainty. Conversely, a large spread from red to blue among the in 2017. Differences in the posterior conductivity percentiles be-
three percentile images indicate a broad CI and greater posterior tween AEM systems, especially in the shallow tens of metres could
uncertainty. In exactly the same format as just described, in Figs 9 be due to differences in the subsurface water saturation in these
and 10, we show the results for fixed-wing AEM data flown over years.
1902 A. Ray et al.
A major point of difference between all three inversions is the mass moves towards conducting (e.g. near BHGT14M1). We look
presence of relatively certain, deep conductors shown by the joint next in detail at borehole BHGT14M1, after examining the Bayesian
inversion of fixed-wing data at −200 m relative to the Australian information gain at all sounding locations for context.
Height Datum (Fig. 9). Notably, these deep features are missing from Fig. 12(a) shows the Bayesian information gain for all surveys
Fig. 10 which uses the same input AEM data as the joint inversion. and inversions at all locations, calculated using the methods de-
The presence of these features cannot be validated as there are no tailed in Appendix D. On average, the helicopter system shows a
induction logs in their vicinity. Whether the helicopter system and higher information gain nearer the surface, and the fixed-wing/joint
the amplitude-only inversions fail to see these deep conductors, or inversion shows higher gain as we go deeper (Fig. 12b). All systems
whether they are artefacts, we are unable to say at this point. A more and inversions show increasing information gain in the near surface,
detailed examination of the fixed-wing nuisances is carried out in where conductors are inferred. Keeping in mind the discussion in
Appendix A. Section 4.1 on the synthetic example, we now examine the infor-
mation gain (Fig. 12c) and resolving capability of all systems in the
near surface at BHGT14M1 (Fig. 13), where there is a known strong
conductor within the first 25 m, with well established knowledge of
5.1 Comparison with borehole conductivities and
the aquifer system.

deterministic inversion results
From Fig. 13 it is apparent that for all three probabilistic in-
Since it is difficult to show probabilistic results and induction log versions, the posterior percentiles of conductivity follow the same
conductivities all together, we have taken the following approach: trends as the induction log shown with a cyan trace. As mentioned
We display in each panel of Fig. 11 the median model (i.e. 50th earlier, owing to the temporal changes that may have occurred in
percentile) from the probabilistic inversions. For the same data, in subsurface water content, as well the difference in induction log-
the remaining rows we show the results from a deterministic Oc- ger and AEM system footprints, comparing more than the long-
cam’s inversion. Please note that the comparison is only made in wavelength trends is inadvisable. To facilitate qualitative compari-
the shallow region as the boreholes do not extend deeper. The depth son, the borehole conductivity has been upscaled to the resolution
averaged induction-log conductivities have also been displayed on of the AEM data taking into consideration both the AEM system
every panel. While comparing AEM bulk conductivities with down- physics as well as the prior correlation length, as detailed in Ap-
hole induction-log conductivities we need to keep in mind that the pendix E, using the methodology of Davis & Hauser (2020). The
footprint of the AEM system is far larger. Therefore, it is sufficient upscaled log takes into account different physics and data types,
that the long-wavelength trends from the induction-log match the but uses the same priors on resistivity and a depth length scale of
inverted AEM conductivities (see Davis & Hauser 2020, and Ap- 2 units. This is the same length scale as is used in the real data
pendix E for a detailed comparison workflow). We should also note inversions, and the upscaled log is shown with a green line in each
that the wells were logged in different years from the AEM data ac- case (Figs 13a–c). The induction log (cyan) intersects at least three
quisition. These caveats notwithstanding, there is good agreement distinct geological zones, which can be interpreted alongside the
between the probabilistic median model, deterministic inversions lithological log and groundwater salinity measurements from dif-
and the downhole conductivity logs, for all systems. The fact that ferent intervals in the sequence. Zone 1 from ∼3 to 29 m coincides
the deterministic inversion results are smoother than the median with variable conductivity saturated clays of the Willotia Formation
conductivity section for all systems can be attributed to the fact and upper Calivil Formation. Zone 2 from 25 to 49 m is of low
that the Occam method produces the smoothest model compatible conductivity down to about 40 m depth (reaching ∼1.4 log10 m)
with the data. The fixed-wing Occam inversions are smoother than as it intersects stacked sand facies with fresh water from the Calivil
their helicopter Occam counterparts, while the difference between aquifer, below which the resistivity decreases gradually to ∼0.4
systems is not so apparent in the probabilistic inversions. We conjec- log10 m due to the presence of brackish water. Although induction
ture that this is due to the all-at-once nature of the updating scheme log data was not acquired in the underlying Renmark formation
for geometry nuisances and conductivity values in the non-linear (Zone 3), this unit is known to be more conductive than the overly-
deterministic inversion (see Brodie 2010, for details). This presents ing Calivil formation. While the helicopter data do indeed narrow
itself as an opportunity to improve the existing update and/or down the posterior range at ∼10 m by about 1 bit as evidenced by
regularization schemes that are currently in use for fixed-wing both Figs 12(c) and 13(c), this could be due to temporal variation
data. in near surface conductivity. More importantly at this depth, with
Another point of importance for geological interpretation when all three inversion types, 0.1 log10 m is bracketed at most within
examining Figs 8–11, is that in the northwest half of the line the pos- an 80 per cent credible width of half a conductivity decade—giving
terior percentiles indicate a return to resistive geology with depth good indication of a conductor. To summarize, the posterior resis-
as supported by the depth averaged conductivities in BHMAR63-1 tivity from all three inversion types are consistent with and well
(Fig. 11). However, this is not clear from the deterministic inver- supported by the borehole data as well as known hydrogeology.
sion (bottom three rows of Fig. 11) since the inverted models in Fig. 14 assures us that the fits to AEM data at this spatial location
this region could be returning to the resistive background model nearest the borehole for all inversions are good and that the posterior
at depth and there is no good way to tell from the deterministic conclusions drawn are sound.
inversion alone. Similarly, within the lake itself to the southeast,
borehole BHGT14M1 indicates a return to conductive formations
at the bottom, as do all the deterministic inversions. Although the
6 C O N C LU S I O N S
deterministic inversions shown here do not do so, a gradient-based
inversion can prefer conductivities at 0log10 m due to minimum From our study over the Menindee test site in New South Wales,
norm updates, and can skew conductive underneath shallow conduc- Australia, as well as from synthetic experiments, we conclude
tors due to the absence of a resistive prejudice model. A probabilistic the following: both the low flying helicopter and higher flying
approach can preclude this scenario as well, if posterior probability fixed-wing time domain AEM systems studied contain comparable
subsurface information. This information largely reflects the true Andrieu, C. & Roberts, G.O., 2009. The pseudo-marginal approach for
geoelectrical profile—evidenced from the behaviour of inverted efficient Monte Carlo computations, Ann. Stat., 37(2), 697–725.
posterior conductivities and information theoretic divergences with Auken, E. & Christiansen, A.V., 2004. Layered and laterally constrained 2D
depth, as well as the compatibility of conductivity CIs with long inversion of resistivity data, Geophysics, 69(3), 752–761.
Auken, E., Boesen, T. & Christiansen, A.V., 2017. A review of airborne
wavelength trends in upscaled induction logs. Prevalent determin-
electromagnetic methods with focus on geotechnical and hydrological
istic fixed-wing AEM inversions are carried out using the nuisance
applications from 2007 to 2017, Adv. Geophys., 58, 47–93.
updating strategies laid out in Brodie (2010), and given the sim- Barlow, M., 2019. Geoscience Australia: The year in review, Preview,
ilarities in the probabilistic inversion posteriors for all systems 2019(201), 9–10.
(Figs 8, 9 and 10), there is a possibility that this updating strat- Bedrosian, P.A., Schamper, C. & Auken, E., 2016. A comparison of
egy could be improved upon. If this possibility can be realized, helicopter-borne electromagnetic systems for hydrogeologic studies,
then the last three rows of Fig. 11 from the top will look as sim- Geophys. Prospect., 64(1), 192–215.
ilar to each other as do the first three rows. The implications of Beier, P., Burnham, K.P. & Anderson, D.R., 2002. Model selection and
such an improvement are that the fixed-wing AusAEM dataset inference: a practical information-theoretic approach, Model Selection
(https://www.ef tf .ga.gov.au/ausaem) covering nearly 60 per cent and Inference: A Practical Information-Theoretic Approach, 2nd edn,
Springer-Verlag.
of Australia can be re-inverted to provide greater detail at little

Bezanson, J., Karpinski, S., Shah, V.B. & Edelman, A., 2012. Julia: A
extra cost compared to the acquisition of this massive dataset (Ley-
Fast Dynamic Language for Technical Computing, pp. 1–27, preprint
Cooper & Brodie 2020). Finally, we formally present in this work (arXiv:1209.5145). https://doi.org/10.48550/arXiv.1209.5145.
the amplitude only inversions for fixed-wing systems, and note that Bezanson, J., Edelman, A., Karpinski, S. & Shah, V.B., 2017. Julia: a fresh
they appear to remove artefacts in the inversion process. However, approach to numerical computing, SIAM Rev., 59(1), 65–98.
we are unclear if the features removed could also be considered tar- Bhattacharyya, A., 1943. On a measure of divergence between two statistical
gets worthy of further geophysical investigation, especially if they populations defined by their probability distributions, Bull. Calcutta Math.
fit to within noise in the joint inversion. Soc., 35, 99–109.
Computational considerations for McMC based inversion have Blatter, D., Key, K., Ray, A., Foley, N., Tulaczyk, S. & Auken, E., 2018.
been made in Appendix C. Our experience is that the cost of com- Trans-dimensional Bayesian inversion of airborne transient EM data from
Taylor Glacier, Antarctica, Geophys. J. Int., 214, 1919–1936.
pute is a small fraction of the survey acquisition costs including
Blatter, D., Ray, A. & Key, K., 2021. Two-dimensional Bayesian inver-
mobilization charges. This is especially true if we carry out de-
sion of magnetotelluric data using trans-dimensional Gaussian processes,
terministic inversions for all soundings, followed by probabilistic Geophys. J. Int., 226(1), 548–563.
inversions at a decimated set of locations. This ensures we can Bodin, T. & Sambridge, M., 2009. Seismic tomography with the reversible
gather the posterior probabilistic detail to properly characterize jump algorithm, Geophys. J. Int., 178(3), 1411–1436.
the geological framework of interest and make land, water or re- Bodin, T., Sambridge, M., Tkalčić, H., Arroucau, P., Gallagher, K.
source utilization decisions under uncertainty (e.g. Symington et al. & Rawlinson, N., 2012. Transdimensional inversion of receiver
2020). functions and surface wave dispersion, J. geophys. Res., 117(B2),
doi:10.1029/2011JB008560.
Bröcker, J. & Smith, L.A., 2007. Scoring probabilistic forecasts: the impor-
AC K N OW L E D G M E N T S tance of being proper, Wea. Forecast., 22(2), 382–388.
Brodie, R.C., 2010. Holistic inversion of airborne electromagnetic data, PhD
All calculations were carried out using the Julia language (Bezanson thesis, pp. 121–127, The Australian National University.
et al. 2012, 2017), available under the MIT license. This study Carvalho, A., 2016. An overview of applications of proper scoring rules,
was carried out as part of the Exploring for the Future program Decision Anal., 13(4), 223–242.
(https://www.ef tf .ga.gov.au) under the High Quality Geophysical Chaloner, K. & Verdinelli, I., 1995. Bayesian experimental design: a review,
Analysis (HiQGA) module. Stat. Sci., 10(3), 409–435.
Chen, Y., Liu, S., Diethe, T. & Flach, P., 2021. Continual density ratio
This project was undertaken with computing resources and ser-
estimation in an online setting, Neural Informat. Process. Syst., 35, https:
vices from the National Computational Infrastructure’s (NCI) Gadi //doi.org/10.48550/arXiv.2103.05276.
cluster. The NCI is supported by the Australian Government. The Constable, S.C., Parker, R.L. & Constable, C.G., 1987. Occam’s inversion:
use of trade, product or firm names is for descriptive purposes only a practical algorithm for generating smooth models from electromagnetic
and does not imply endorsement by the Australian Government. sounding data, Geophysics, 52(3), 289–300.
This paper is published with the permission of the CEO, Geoscience Cui, S. & Luo, C., 2016. Feature-based non-parametric estimation of
Australia. Kullback–Leibler divergence for SAR image change detection, Remote
Sens. Lett., 7(11), 1102–1111.
Davis, A. & Hauser, J., 2020. Blocking borehole conductivity logs at the
D ATA AVA I L A B I L I T Y resolution of above-ground electromagnetic systems, Geophysics, 85(2),
E67–E77.
All software used in this work is under active development. Software Dettmer, J. & Dosso, S.E., 2012. Trans-dimensional matched-field geoa-
and examples can be freely cloned from Geoscience Australia’s coustic inversion with hierarchical error models and interacting Markov
official GitHub repositories: chains, J. acoust. Soc. Am., 132(4), 2239–2250.
https://github.com/GeoscienceAustralia/HiQGA.jl Dettmer, J., Dosso, S.E. & Holland, C.W., 2010. Trans-dimensional geoa-
https://github.com/GeoscienceAustralia/GA-AEM coustic inversion, J. acoust. Soc. Am., 128(6), 3393–3405.
Dettmer, J., Molnar, S., Steininger, G., Dosso, S.E. & Cassidy, J.F., 2012.
Trans-dimensional inversion of microtremor array dispersion data with
hierarchical autoregressive error models, Geophys. J. Int., 188(2), 719–
REFERENCES 734.
Agostinetti, N.P. & Malinverno, A., 2010. Receiver function inversion by Dettmer, J., Dosso, S.E., Bodin, T., Stipčević, J. & Cummins, P.R., 2015.
trans-dimensional Monte Carlo sampling, Geophys. J. Int., 181(2), 858– Direct-seismogram inversion for receiver-side structure with uncertain
872. source-time functions, Geophys. J. Int., 203(2), 1373–1387.
1904 A. Ray et al.
Diks, C., Panchenko, V. & Van Dijk, D., 2011. Likelihood-based scoring Key, K., 2012. Is the fast Hankel transform faster than quadrature?,
rules for comparing density forecasts in tails, J. Econometr., 163(2), 215– Geophysics, 77(3), F21–F30.
230. Key, K. & Lockwood, A., 2010. Determining the orientation of ma-
Dosso, S.E. & Dettmer, J., 2011. Bayesian matched-field geoacoustic inver- rine CSEM receivers using orthogonal Procrustes rotation analysis,
sion, Inverse Problems, 27(5), doi:10.1088/0266-5611/27/5/055009. Geophysics, 75(3), doi:10.1190/1.3378765.
Dosso, S.E., Holland, C.W. & Sambridge, M., 2012. Parallel tempering for Kullback, S. & Leibler, R.A., 1951. On information and sufficiency, Ann.
strongly nonlinear geoacoustic inversion., J. acoust. Soc. Am., 132(5), Math. Stat., 22(1), 79–86.
3030–3040. Lane, R., Green, A., Golding, C., Owers, M., Pik, P., Plunkett, C., Sattel, D.
Dosso, S.E., Dettmer, J., Steininger, G. & Holland, C.W., 2014. Efficient & Thorn, B., 2000. An example of 3D conductivity mapping using the
trans-dimensional Bayesian inversion for geoacoustic profile estimation, TEMPEST airborne electromagnetic system, Explor. Geophys., 31(2),
Inverse Problems, 30(11), doi:10.1088/0266-5611/30/11/114018. 162–172.
Earl, D.J. & Deem, M.W., 2005. Parallel tempering: theory, applications, Ley-Cooper, A. & Brodie, R., 2020. AusAEM: imaging the near-surface
and new perspectives, Phys. Chem. Chem. Phys., 7(23), 3910–3916. from the world’s largest airborne electromagnetic survey, in Exploring
Eberle, D. & Siemon, B., 2006. Identification of buried valleys using the for the Future: Extended Abstracts, pp. 1–4, eds Czarnota, K., Roach, I.,
BGR helicopter-borne geophysical system, Near Surf. Geophys., 4, 125– Abbott, S., Haynes, M., Kositcin, N., Ray, A. & Slatter, E., Geoscience
133. Australia.

Farquharson, C.G. & Oldenburg, D.W., 1998. Non-linear inversion using Ley-Cooper, A.Y., 2021. Exploring for the future AusAEM eastern resources
general measures of data misfit and model structure, Geophys. J. Int., corridor : 2021 airborne electromagnetic survey, Tech. rep., Geoscience
134(1), 213–227. Australia, Symonston.
Fisher, R.A. & Yates, F., 1938. Statistical Tables: For Biological, Agricul- Ley-Cooper, A.Y., 2022. AusAEM – WA, Murchison Airborne electromag-
tural and Medical Research, Oliver and Boyd. netic survey blocks, Tech. rep., Geoscience Australia, Symonston.
Fitterman, D.V. & Yin, C., 2004. Effect of bird maneuver on frequency- Ley-Cooper, A.Y., Brodie, R.C. & Richardson, M., 2020. AusAEM: Aus-
domain helicopter EM response, Geophysics, 69(5), 1203–1215. tralia’s airborne electromagnetic continental-scale acquisition program,
Friedli, L., Linde, N., Ginsbourger, D. & Doucet, A., 2022. Lithological Explor. Geophys., 51(1), 193–202.
tomography with the correlated pseudo-marginal method, Geophys. J. Lindley, D.V., 1956. On a measure of the information provided by an exper-
Int., 228(2), 839–856. iment, Ann. Math. Stat., 27(4), 986–1005.
Gehrmann, R. A.S., Dettmer, J., Schwalenberg, K., Engels, M., Dosso, Loseth, L.O. & Ursin, B., 2007. Electromagnetic fields in planarly layered
S.E. & Özmaral, A., 2015. Trans-dimensional Bayesian inversion of anisotropic media, Geophys. J. Int., 170(1), 44–80.
controlled-source electromagnetic data in the German North Sea, Geo- MacKay, D. J.C., 2003. Information Theory, Inference and Learning Algo-
phys. Prospect., 63(6), 1314–1333. rithms, Cambridge Univ. Press.
Gelman, A., 2006. Multilevel (hierarchical) modeling: what it can and cannot Malinverno, A., 2002. Parsimonious Bayesian Markov chain Monte Carlo
do, Technometrics, 48(3), 432–435. inversion in a nonlinear geophysical problem, Geophys. J. Int., 151(3),
Geyer, C., 2011. Introduction to MCMC methods, in Handbook of Markov 675–688.
Chain Monte Carlo, 1st edn, pp. 3–48, eds Gelman, A., Jones, G., Brooks, Malinverno, A., Briggs & V.a., 2004. Expanded uncertainty quantification
S. & Meng, X.-L., Chapman & Hall/CRC. in inverse problems: hierarchical Bayes and empirical Bayes, Geophysics,
Geyer, C.J., 1991. Markov chain Monte Carlo maximum likelihood, in Pro- 69(4), doi:10.1190/1.1778243.
ceedings of the 23rd Symposium on the Interface, American Statistical Mecklenbrauker, C.F. & Gerstoft, P., 2000. Objective functions for ocean
Association, New York, pp. 156–163. acoustic inversion derived by likelihood methods, J. Comput. Acoust.,
Ghalenoei, E., Dettmer, J., Ali, M.Y. & Kim, J.W., 2021. Gravity and mag- 8(2), 259–270.
netic joint inversion for basement and salt structures with the reversible- Menke, W., 2012. Geophysical Data Analysis: Discrete Inverse Theory,
jump algorithm, Geophys. J. Int., 227(2), 746–758. Academic Press.
Gneiting, T. & Raftery, A.E., 2007. Strictly proper scoring rules, prediction, Metropolis, N., Rosenbluth, A.W., Rosenbluth, M.N., Teller, A.H. & Teller,
and estimation, J. Am. Stat. Assoc., 102(477), 359–378. E., 1953. Equation of state calculations by fast computing machines, J.
Good, I.J., 1952. Rational decisions, J. R. Stat. Soc., B, 14(1), 107–114. Chem. Phys., 21(6), 1087–1092.
Green, A. & Lane, R., 2003. Estimating noise levels in AEM data, ASEG Minsley, B.J., 2011. A trans-dimensional Bayesian Markov chain Monte
Extend. Abstr., 2003(2), 1–5. Carlo algorithm for model assessment using frequency-domain electro-
Green, P.J., 1995. Reversible jump Markov chain Monte Carlo computation magnetic data, Geophys. J. Int., 187(1), 252–272.
and Bayesian model determination, Biometrika, 82(4), 711–732. Minsley, B.J., Esfahani, A., Deszcz-pan, M., Brodie, R.C. & Survey, U. S.G.,
Hastings, W.K., 1970. Monte Carlo sampling methods using Markov chains 2002. A Bayesian approach to the interpretation of airborne electromag-
and their applications, Biometrika, 57(1), 97–109. netic surveys: quantifying data errors, model assessment, and lithology
Hawkins, R., Brodie, R.C. & Sambridge, M., 2018. Trans-dimensional classification, in Proceedings of the 6th International AEM Conference &
Bayesian inversion of airborne electromagnetic data for 2D conductiv- Exhibition, October 2013, cp-383-00021, European Association of Geo-
ity profiles, Explor. Geophys., 49(2), 134–147. scientists & Engineers.
Hodges, G., 1999. A world of applications for helicopter electromagnet- Minsley, B.J., Foks, N.L. & Bedrosian, P.A., 2021a. Quantifying model
ics to environmental and engineering problems, in Symposium on the structural uncertainty using airborne electromagnetic data, Geophys. J.
Application of Geophysics to Engineering and Environmental Prob- Int., 224(1), 590–607.
lems 1999, pp. 899–907, Environmental & Engineering Geophysical Minsley, B.J., Rigby, J.R., James, S.R., Burton, B.L., Knierim, K.J., Pace, M.
Society. D.M., Bedrosian, P.A. & Kress, W.H., 2021b. Airborne geophysical sur-
Hoffimann, J., Zortea, M., de Carvalho, B. & Zadrozny, B., 2021. Geosta- veys of the lower Mississippi Valley demonstrate system-scale mapping
tistical learning: challenges and opportunities, Front. Appl. Math. Stat., of subsurface architecture, Nat.. Commun. Earth Environ., 2(1), 1–14.
7(July), 1–15. Mosegaard, K. & Tarantola, A., 1995. Monte Carlo sampling of solutions
Kalscheuer, T., de los Ángeles Garcı́a Juanatey, M., Meqbel, N. & Pedersen, to inverse problems, J. geophys. Res., 100(B7), 12 431–12 447.
L.B., 2010. Non-linear model error and resolution properties from two- Palacky, G.J. & West, G.F., 1991. Airborne electromagnetic methods, in
dimensional single and joint inversions of direct current resistivity and Electromagnetic Methods in Applied Geophysics, Vol. 2, Application,
radiomagnetotelluric data, Geophys. J. Int., 182(3), 1174–1188. Parts A and B, Society of Exploration Geophysicists.
Key, K., 2009. 1D inversion of multicomponent, multifrequency marine Parker, R.L., 1994. Geophysical Inverse Theory, Princeton Univ. Press.
CSEM data: methodology and synthetic studies for resolving thin resistive Pinkard, H. & Waller, L., 2022. A visual introduction to information theory,
layers, Geophysics, 74(2), doi:10.1190/1.3058434. Lecture Notes Electr. Eng., 785, 517–558.
Rasmussen, C.E. & Williams, C. K.I., 2006. Gaussian Processes for Ma- Subašić, S., Agostinetti, N.P. & Bean, C.J., 2019. Estimating lateral and
chine Learning, MIT Press. vertical resolution in receiver function data for shallow crust exploration,
Ray, A., 2021. Bayesian inversion using nested trans-dimensional Gaussian Geophys. J. Int., 218(3), 2045–2053.
processes, Geophys. J. Int., 226, 302–326. Sugiyama, M., Nakajima, S., Kashima, H., Von Bünau, P. & Kawanabe,
Ray, A. & Key, K., 2012. Bayesian inversion of marine CSEM data with a M., 2008a. Direct importance estimation with model selection and its
trans-dimensional self parametrizing algorithm, Geophys. J. Int., 191(3), application to covariate shift adaptation, in Proceedings of the Advances
1135–1151. in Neural Information Processing Systems 20 Conference (NIPS 2007),
Ray, A. & Myer, D., 2019. Bayesian geophysical inversion with trans- pp. 1–8.
dimensional Gaussian Process machine learning, Geophys. J. Int., 217, Sugiyama, M., Suzuki, T., Nakajima, S., Kashima, H., Von Bünau, P. &
1706–1726. Kawanabe, M., 2008b. Direct importance estimation for covariate shift
Ray, A., Alumbaugh, D.L., Hoversten, G.M. & Key, K., 2013. Ro- adaptation, Ann. Inst. Stat. Math., 60(4), 699–746.
bust and accelerated Bayesian inversion of marine controlled-source Sugiyama, M., Suzuki, T. & Kanamori, T., 2012. Density Ratio Estimation
electromagnetic data using parallel tempering, Geophysics, 78(6), in Machine Learning, Cambridge Univ. Press.
E271–E280. Sugiyama, M., Liu, S. & Christoffel, M., 2013. Direct divergence approxi-
Ray, A., Greene, M., Edwards, M. & Cardona, R., 2022. Methods and mation between probability, J. Comput. Sci. Eng., 7(2), 99–111.
systems for calibrating depth in a well to seismic data in a subsurface Swendsen, R.H. & Wang, J.S., 1987. Nonuniversal critical dynamics in

volume of interest. Patent. Monte Carlo simulations, Phys. Rev. Lett., 58(2), 86–88.
Roulston, M.S. & Smith, L.A., 2002. Evaluating probabilistic forecasts using Symington, N., Ray, A., Harris-Pascal, C., Tan, K.P., Ley-Cooper, Y. &
information theory, Mon. Wea. Rev., 130(6), 1653–1660. Brodie, R.C., 2020. Groundwater salinity estimation using borehole and
Ryan, K.J., 2003. Estimating expected information gains for experimental AEM data: a framework for uncertainty analysis, in EFTF Extended
designs with application to the random fatigue-limit model, J. Comput. Abstracts, August, Commonwealth of Australia (Geoscience Australia),
Graph. Stat., 12(3), 585–603. doi:10.11636/135242.
Sambridge, M., 2013. A parallel tempering algorithm for probabilis- Tarantola, A., 2005. Inverse Problem Theory and Methods for Model Pa-
tic sampling and multimodal optimization, Geophys. J. Int., 196(1), rameter Estimation, SIAM.
357–374. Tarantola, A. & Valette, B., 1982. Inverse problems= quest for information,
Sambridge, M., Gallagher, K., Jackson, a. & Rickwood, P., 2006. Trans- J. Geophys., 50, 159–170.
dimensional inverse problems, model comparison and the evidence, Valentine, A.P. & Sambridge, M., 2020. Gaussian process models—I. A
Geophys. J. Int., 167(2), 528–542. framework for probabilistic continuous inverse theory, Geophys. J. Int.,
Seillé, H. & Visser, G., 2020. Bayesian inversion of magnetotelluric data 220(3), 1632–1647.
considering dimensionality discrepancies, Geophys. J. Int., 223(3), 1565– Weijs, S.V., van Nooijen, R. & van de Giesen, N., 2010. Kullback-leibler
1583. divergence as a forecast skill score with classic reliability-resolution-
Sheather, S.J. & Jones, M.C., 1991. A reliable data-based bandwidth se- uncertainty decomposition, Mon. Wea. Rev., 138(9), 3387–3399.
lection method for kernel density estimation, J. R. Stat. Soc., B, 53(3), Worrall, L., Whitaker, A., Lane, R. & Meyers, J., 2001. Exploring through
683–690. cover – the integrated interpretation of high resolution aeromagnetic,
Smith, L.A., Suckling, E.B., Thompson, E.L., Maynard, T. & Du, H., 2015. airborne electromagnetic and ground gravity data from the Grant’s Patch
Towards improving the framework for probabilistic forecast evaluation, area, Eastern Goldfields Province, Archaean Yilgarn Craton. Part A: map-
Clim. Change, 132(1), 31–45. ping geology using airb, ASEG Extended Abstracts, 2001(1), 1–4.
Sorensen, K.I. & Auken, E., 2004. SkyTEM—a new high-resolution heli- Yardim, C., Gerstoft, P. & Hodgkiss, W.S., 2006. Estimation of radio re-
copter transient electromagnetic system, Explor. Geophys., 35(3), 194– fractivity from radar clutter using Bayesian Monte Carlo analysis, IEEE
202. Trans. Antenn. Propagat., 54(4), 1318–1327.
1906 A. Ray et al.
A P P E N D I X A : N U I S A N C E S F O R F I X E D - W I N G I N V E R S I O N S OV E R M E N I N D E E
The inverted conductivities as well as acquisition geometry nuisances are shown in this section, both for the conventional fixed-wing joint
inversion (Fig. A1) and for the same input AEM data, the amplitude only inversion (Fig. A2). The fact that the inverted nuisance values
(blue lines and shaded blue regions) do not always overlap with the measured IMU-provided values (orange) is not surprising, as the IMU
readings could be inaccurate, or it is possible that the inversion has found some trade-off as was shown in the synthetic examples. What is
important however, is that the inversion be parametrized to allow for geometry nuisance inference, as in their absence, the data residual would
be propagated incorrectly into the inverted conductivities.

Figure A1. Summary of posterior nuisances for conventional fixed-wing Bx and Bz joint inversion over Menindee. The top row shows the mean sampled φ d
and one standard deviation. The next three rows show in blue the mean sampled receiver height, Tx–Rx horizontal separation and receiver pitch, as well as
their one standard deviation spread. Orange values indicate what has actually been measured during the survey. The last three rows, as before, show the 5th,
50th (median) and 95th percentiles of posterior conductivity. Over the deep conductors at −200 m, there do not appear to be any sudden changes in geometry
during the flight which may otherwise cause artefacts.
A P P E N D I X B : T R A N S - D I M E N S I O NA L A LG O R I T H M W I T H N U I S A N C E S A M P L I N G
A N D PA R A L L E L T E M P E R I N G
As shown by the pseudocode provided in Algorithm B, one step of our algorithm encapsulates a reversible jump or trans-dimensional step
(starting at Line 4), followed by a nuisance sampling step (starting at Line 11), contained within a parallel tempering loop (Lines 3-20),
followed by a parallel tempering swap (Lines 21–28).
Details of the trans-dimensional birth and death moves for resistivity models are exactly the same as described for stationary Gaussian
processes (Ray & Myer 2019; Ray 2021). The nuisance sampler uses ordinary Metropolis–Hastings (Metropolis et al. 1953; Hastings 1970)
for the fixed-wing geometry nuisance parameters which are nnuisances in number.
Parallel tempering is used to exchange information between interacting McMC chains to escape local misfit minima (i.e. likelihood
maxima). Temperatures or models are exchanged at the end of each McMC step using the following Metropolis–Hastings criterion (Swendsen
& Wang 1987; Geyer 1991; Earl & Deem 2005; Dettmer et al. 2012; Sambridge 2013):
⎡ 1/Ta 1/Tb ⎤
L ( θ mb ) L ( θ ma )
αswap (a, b) = min ⎣1, ⎦. (B1)
L(θ m a ) L(θ m b )

Figure A2. Summary of posterior nuisances for fixed-wing amplitude only inversion over Menindee. The top row shows the mean sampled φ d and one standard
deviation. The next two rows show the sampled receiver height and Tx–Rx horizontal separation, as well as their one standard deviation spread. Orange values
in these rows indicate what has actually been measured during the survey. The final three rows, as before, show the 5th, 50th (median) and 95th percentiles of
posterior conductivity. The deep conductors at −200 m from the fixed-wing amplitude only inversion do not appear in these posterior summaries.
Figure A3. Posterior conductivities obtained after forward modelling the well log conductivities (yellow line) at borehole BHGT14M1 at Menindee lakes for
(a) joint Bx , Bz inversion, (b) amplitude only B field and (c) helicopter dBz /dt data. As in the main text, the dashed lines represent the 5, 50 and 95 percent
posterior percentiles of conductivity. In each case, the posterior median has been used as the upscaled log for qualitative comparison in Fig. 13.
For a description of why swapping models for escaping local likelihood maxima using (B1) is effective, see section 3.2 of Blatter et al. (2018).
For computational efficiency, temperatures are exchanged during inter-process communication to achieve the exact same effect as swapping
models. No likelihood computations are required in the calculation of the swap probability in (B1), as they have already been evaluated in the
preceding j loop on Line 3 of Algorithm B.
All Markov chains were run at log-spaced temperatures between 1 and 2.5. Details of setting a temperature ladder can be found in Dettmer
et al. (2012) and Ray et al. (2013). The helicopter AEM inversions required five temperatures and no nuisance sampling, whereas the fixed-
wing inversions used 7 temperatures with nuisance sampling. Larger numbers of temperatures are required to sample rugged likelihoods.
Posterior inference is carried out only from models that are at T = 1 after sorting the chains by temperature.
1908 A. Ray et al.

Algorithm B: Pseudocode for trans-dimensional McMC with nuisance sampling, and parallel tempering exchanges though a Fisher-Yates shuffle (Fisher &
Yates 1938).
A P P E N D I X C : H I G H P E R F O R M A N C E C O M P U T I N G C O N S I D E R AT I O N S
It can be seen from Algorithm B that the nuisance sampler does require likelihood evaluation (i.e. a forward call). Thus sampling the fixed-wing
system posterior requires twice as many forward calls as it does to sample the helicopter system posterior. For the real data examples, one
sampling step required 70-80 ms for the fixed-wing system, and 40–45 ms for the helicopter system. This evaluation time is dominated by
the forward evaluation for the 52 layer model. This is fortunate since, if each McMC sample took less time than a few ms, the overhead in
inter-process communication per McMC step would make the parallel computation inefficient. Parallel tempering is not an embarrassingly
parallel algorithm, owing to the requirement that information be exchanged between interacting chains. This can be seen from Algorithm B
on Line 3, where an independent McMC sampling step is made at each temperature, with the exchange of temperatures occurring on Line
21 of the algorithm. Care must be taken in the computer implementation of this exchange step to transfer the bare minimum information
possible, in order to the make the inter-process communication efficient.
The real data examples required 200 000 McMC samples for each chain in parallel tempering, though we extended each sampling run to
400 000 samples. For each sounding, 200 000 samples required approximately 4.5 hr for the fixed-wing system on 7 + 1 CPUs, and 2.5 hr
for the helicopter system on 5 + 1 CPUs. The +1 indicates the manager process for inter-CPU communication and each remaining CPU was
assigned to McMC computation at a specified temperature. A significant improvement has been made in the HPC implementation of parallel
tempering in this work compared to our earlier work in Blatter et al. (2018), which was based on the parallel tempering implementation of
Ray et al. (2013). As noted by Blatter et al. (2018), one single-moment helicopter AEM sounding required 5 d to invert probabilistically. In
our current work, soundings were batched in parallel such that 120 soundings for the fixed-wing system, and 160 soundings for the helicopter
system were carried out at the same time using 960 CPUs. To be explicit, an entire batch of 160 soundings were probabilistically inverted
within 2.5 hr for the helicopter system using 960 CPUs. The full set of 720 helicopter soundings over the Menindee test range required five
batches and a total of 12.5 hr. If more CPUs are made available, then a greater number of soundings can be parallelly inverted in a batch, and
fewer numbers of batches will be required in total. Similar calculations can be made for the fixed-wing system. On average, the fixed-wing
posterior sampling results require twice as much time with the same computational resources, due to nuisance sampling needing an additional
forward evaluation—though this can perhaps be avoided by reusing precomputed Hankel transform evaluations (future work).
The Bayesian inversion including the forward modeller code is written purely in Julia (v1.6, tested on v1.7-1.8). The deterministic inversion
code including the forward modeller is written in modern C++. The parallel computation in this study used 48-core @3.2 GHz Intel Xeon
Cascade Lake nodes in the NCI’s Gadi cluster.

A P P E N D I X D : I N F O R M AT I O N T H E O R E T I C B A C KG R O U N D
D1 Relative entropy and information gain

Uncertainty analysis in a Bayesian sense naturally leads to the formulation of information gain (Lindley 1956), in terms of the difference in
information when comparing the posterior and prior distributions. This is provided by the relative entropy between the two densities, known
as the Kullback–Leibler divergence (Kullback & Leibler 1951). The KLD is not symmetric between the densities that differ, and to avoid
confusion, for information gain we will use the definition of the KLD for prior density q(x) and posterior density p(x) as defined in eq. (22)
of Valentine & Sambridge (2020):

p( x )
DKL ( p||q) = p( x ) log dx . (D1)
q( x )
When this is computed with the natural logarithm as given in (D1), the information gain units are known as ‘nats’ and a change of base to
2 (i.e. multiplying (D1) by log2 e) changes the units to bits. As we will prove shortly, the KLD is zero if and only if the densities p and q
are identical, otherwise it is always an unbounded positive quantity. We can gain a useful perspective on the KLD by looking at it as the
expectation of the log ratio of the posterior to the prior when x is distributed according to the posterior,

p( x)
DKL ( p||q) = Ex ∼ p(x ) log . (D2)
q( x)
Eq. (D2) sets the stage for both the estimation of the KLD as information gain from samples, and showing how the KLD leads to the derivation
of the logarithmic score, a strictly proper scoring rule (Gneiting & Raftery 2007).
D2 Density ratio estimation for Bayesian information gain

If the densities p and q, respectively, can be parametrically estimated from their samples, it may be possible to analytically compute the KLD.
This is the approach followed by (Friedli et al. 2022) in assessing how well different methods approximate a known, Gaussian posterior
for a linear inverse problem with a Gaussian prior. However, calculation of the KLD is subject to the usual difficulties (Cui & Luo 2016)
when calculating entropy, if parametric approaches are not suitable, as in our AEM inversion case. Histogram binning for p and q may be
problematic as this will depend on the number and location of histogram bins. Another possibility is separately estimating densities for p and
q and then carrying out numerical or Monte Carlo integration, but the derived quantity p/q is unstable and may be be difficult to use. In cases
where a density ratio p/q is the quantity of interest (as in (D2)), it is advantageous to resort to direct methods for density ratio estimation (see
Sugiyama et al. 2012, for a comprehensive review). A recent example requiring direct density ratio estimation in the geosciences can be found
in Hoffimann et al. (2021), for geostatistical transfer learning. In the particular case of the Kullback-Leibler divergence, the Kullback–Leibler
Importance Estimation Procedure (KLIEP, Sugiyama et al. 2008a) can be used effectively for direct divergence estimation as detailed by
Sugiyama et al. (2013). As can be seen from (D2), the KLD is given by the expected value under the posterior density, of the logarithm
of the density ratio of the posterior to the prior. This is done in a three step process. First, with McMC realizations of prior and posterior
samples, KLIEP is used to fit a linear combination of Gaussian kernels directly to the ratio r ( x) = q( p( x)
x)
. In a second step, r(x) is computed
from the KLIEP estimated density ratio function when x is simulated from stored posterior samples. Finally, the mean of log r(x) provides the
KLD as Bayesian information gain according to (D2). The KLIEP procedure selects the Gaussian kernel bandwidth to approximate r through
cross-validation, and a constrained, convex optimisation problem is solved to obtain the density ratio estimate (see Sugiyama et al. 2008a,
for details). Sampling the prior is inexpensive as it requires no likelihood evaluations (i.e. forward calls), and the posterior samples must be
obtained as usual to solve the AEM Bayesian inverse problem.
D3 Logarithmic scoring rule

Scoring rules are used to evaluate the quality of probabilistic forecasts and have found heavy use in weather prediction and econometrics.
A comprehensive review can be found in Gneiting & Raftery (2007). A scoring rule is proper, if incorrect forecast densities are expected
1910 A. Ray et al.
to receive lower scores than the true forecast density, and it is strictly proper, if the maximum expected score attained is unique. While they
have not been used much in near surface geophysics, we note recent applications in Seillé & Visser (2020) for selecting the best likelihood
function for Bayesian inversion, and in Friedli et al. (2022) for evaluating the performance of different Monte Carlo methods. We shall now
prove that the logarithmic score, which dates back as far as Good (1952), is related to the KLD, by first showing that the KLD is always
greater than or equal to zero. As (D2) provides the general definition of DKL (p||q), not just for priors and posteriors, we can write for any two
valid probability densities p and q:
q( x)
− DKL ( p||q) = Ex ∼ p(x ) log . (D3)
p( x)
Using the inequality 1 + log u ≤ u, with exact equality at 1, we can substitute u → q
p
and take expectations with respect to p. We can thus
write:

q( x)
1 + Ex ∼ p(x ) log ≤ Ex ∼ p(x ) q( x)
p( x)
, (D4)
p( x)


q( x)
1 + Ex ∼ p(x ) log ≤ p( x ) q( x)
p( x)
dx. (D5)
p( x)

Now using the fact that q( x)d x = 1 and using (D3) in ( D5), we can say that
1 − DKL ( p||q) ≤ 1, (D6)
thus proving that
DKL ( p||q) ≥ 0; with exact equality only at p( x ) = q( x ) ∀ x . (D7)

The above inequality is known as Gibb’s inequality and establishes with no loss of generality that the KLD is always greater than zero unless
the two densities are identical. Now if we were to consider the logarithmic scoring rule, and p and q are the true and forecast densities
respectively, then we can formulate the Kullback–Leibler Information Criterion (see Diks et al. 2011) for q as:

DKL ( p||q) = p ( x) log p ( x) − log q( x) d x ≥ 0, (D8)

⇒ p( x ) log q( x )d x ≤ p ( x) log p ( x)d x ∀ x. (D9)
Inequality (D9) establishes that for the logarithmic score

S (q, x ) = log q( x), (D10)

p( x ) S (q, x )d x ≤ p( x ) S ( p, x )d x ∀ x . (D11)
The integral on the left hand side of (D11) is the expectation of the incorrect forecast score over all possible values x ∼ p(x) where p(x) is
the true density. This quantity is always less than or equal to the average of the true forecast score, which is the integral on the right hand
side of (D11). This shows that the requirement of propriety is satisfied by the logarithmic scoring rule. Since the maximum expected forecast
score is reached uniquely for q(x) = p(x) ∀x, the logarithmic scoring rule is also strictly proper.
D4 Summary of key information theoretic points

The KLD or relative entropy is a very useful information theoretic tool to discriminate between two probability densities (e.g. MacKay 2003).
When defined with p as posterior and q as prior, the KLD as given in (D1) represents the information gain after carrying out our geophysical
experiment. The KLD can be calculated at every location in the earth, showing the information gain at all depths considered in the geophysical
inversion (e.g. Blatter et al. 2018; Valentine & Sambridge 2020). In the context of assigning scores to posterior predictions, if p is the true
forecast distribution and q is the posterior, then S(q, x) = log q(x) represents a strictly proper score. If we were to treat the posterior resistivity
as a prediction of the truth, then at every location in the earth (D10) provides a strictly proper score, with higher values representing better
performance. For an interpretation in terms of ignorance, given by the negative of the score defined in (D10), we refer readers to Roulston &
Smith (2002) as well as Weijs et al. (2010).
It must be noted that for assessing posterior uncertainty, in the domain of the physical sciences, this is done in terms of the width of
quantities of interest. Broader ranges of probable values indicate lower certainty. In this regard, we have found the KLD for information gain
to be a more robust indicator of subsurface property resolution in terms of higher values indicating lower overlap between a narrow posterior
and broad prior. The logarithmic score is very sensitive to probability modes, and severely penalises narrow posteriors if they have missed true
values. This is a source of debate; see section 4, page 366 of Gneiting & Raftery (2007) and section 5, page 387 of Bröcker & Smith (2007).
In fact, Friedli et al. (2022) report that only a certain fraction of parameters in their geophysical simulations have true values that fall within
the range of the posterior samples. This is important to keep in mind as low probability regions will score very poorly (or even have infinite
ignorance =−log 0) and can bias the average score for a particular posterior density. While other non-local scores exist, they can be relatively
insensitive to the width of probability modes (Smith et al. 2015) and are not considered here. In sum, both logarithmic scores and information
gain are useful analysis tools derived from appropriate usage of the KLD. However, information gain can be estimated when true parameter
values are not exactly known (i.e. most of the earth) for real data. Criteria such as mean absolute error could be used to quantify closeness
to the truth, but again this is not as generally useful as information gain for similar reasons. Further, information gain is not sensitive to true
values falling adjacent to high posterior probability regions—a phenomenon which is common in non-linear geophysical inverse problems.
APPENDIX E: WELL-LOG UPSCALING

In order to populate the static (or dynamic) properties of a reservoir (flow) model, or for comparison with inversion results from surface
geophysical data, borehole logs are often upscaled or adjusted to be compatible with the surface data. Due to their parsimonious nature,
trans-dimensional methods have been found to be highly suited to this task. Using such methods, seismic well ties to adjust surface migration
imaging results with downhole sonic logs have been successfully automated (Ray et al. 2022) and conductivity logs have been represented at

the resolution of AEM systems (Davis & Hauser 2020). For the AEM problem we have followed the latter approach as it is directly applicable
to our case. Their thesis is elegant yet simple, that the best representation of the well log at the scale of above-ground geophysical systems,
is through a central tendency of synthetic posterior conductivities. The synthetic posterior is obtained after inverting the forward modelled
borehole conductivity log using the transmitter characteristics of the overflying AEM system. Using the measured survey data noise and
within the prior bounds given by borehole conductivities, this approach takes into account both the Earth filter due to AEM forward physics,
as well as any assumptions made in the inversion such as choice of prior length scale. The results of modelling the wells and sampling the
posterior conductivities per AEM system are provided in Fig. A3. We have taken the median posterior conductivity, and used it to represent
the long-wavelength features of the borehole log in Fig. 13.

C The Author(s) 2023. Published by Oxford University Press on behalf of The Royal Astronomical Society.

ggad337

Uploaded by

Copyright:

Available Formats

ggad337

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

ggad337

Uploaded by

Copyright:

Available Formats

Geophys. J. Int. (2023) 235, 1888–1911 https://doi.org/10.

An information theoretic Bayesian uncertainty analysis of AEM

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

with the principles of electromagnetic induction, the earth response

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

The likelihood function p(d|θm ) for Gaussian data noise can be

where [f (θm ) − d] is the residual vector of misfit for the model θm ,

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

make exact comparison impossible. However, the surface geophys-

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

D1 Relative entropy and information gain

D2 Density ratio estimation for Bayesian information gain

D3 Logarithmic scoring rule

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

1 − DKL ( p||q) ≤ 1, (D6)

thus proving that

DKL ( p||q) ≥ 0; with exact equality only at p( x ) = q( x ) ∀ x . (D7)

Inequality (D9) establishes that for the logarithmic score

D4 Summary of key information theoretic points

APPENDIX E: WELL-LOG UPSCALING

Downloaded from https://academic.oup.com/gji/article/235/2/1888/7259152 by guest on 05 December 2024

You might also like