remote sensing

Article
Automatic Detection of Near-Surface Targets for
Unmanned Aerial Vehicle (UAV) Magnetic Survey
Yaxin Mu 1,2,3 , Xiaojuan Zhang 1,2, *, Wupeng Xie 1,2,3 and Yaoxin Zheng 1,2,3
1 Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100094, China;
muyaxin15@mails.ucas.ac.cn (Y.M.); xiewupeng15@mails.ucas.ac.cn (W.X.);
zhengyaoxin17@mails.ucas.ac.cn (Y.Z.)
2 Key Laboratory of Electromagnetic Radiation and Sensing Technology, Chinese Academy of Sciences,
Beijing 100190, China
3 School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences,
Beijing 100049, China
* Correspondence: xjzhang@mail.ie.ac.cn; Tel.: +86-10-5888-7276




Received: 2 January 2020; Accepted: 29 January 2020; Published: 1 February 2020


Abstract: Great progress has been made in the integration of Unmanned Aerial Vehicle (UAV)
magnetic measurement systems, but the interpretation of UAV magnetic data is facing serious
challenges. This paper presents a complete workflow for the detection of the subsurface objects,
like Unexploded Ordnance (UXO), by the UAV-borne magnetic survey. The elimination of interference
field generated by the drone and an improved Euler deconvolution are emphasized. The quality
of UAV magnetic data is limited by the UAV interference field. A compensation method based on
the signal correlation is proposed to remove the UAV interference field, which lays the foundation
for the subsequent interpretation of UAV magnetic data. An improved Euler deconvolution is
developed to estimate the location of underground targets automatically, which is the combination of
YOLOv3 (You Only Look Once version 3) and Euler deconvolution. YOLOv3 is a deep convolutional
neural network (DCNN)-based image and video detector and it is applied in the context of magnetic
survey for the first time, replacing the traditional sliding window. The improved algorithm is
more satisfactory for the large-scale UAV-borne magnetic survey because of the simpler and faster
workflow, compared with the traditional sliding window (SW)-based Euler method. The field test is
conducted and the experimental results show that all procedures in the designed routine is reasonable
and effective. The UAV interference field is suppressed significantly with root mean square error
0.5391 nT and the improved Euler deconvolution outperforms the SW Euler deconvolution in terms
of positioning accuracy and reducing false targets.

Keywords: near-surface targets detection; UAV-borne magnetic survey; UAV interference field;
interpretation method of magnetic field; Euler deconvolution; YOLOv3

1. Introduction
Magnetic detection is an ancient remote sensing technology that can be traced back to the
mid-19th century for the exploration of iron ore deposits. Today, magnetic sensing is still attractive
in various applications by virtue of its passive, rapid, noninvasive detection performance [1,2],
like geological exploration [3,4], Unexploded Ordnance (UXO) detection [5,6], archaeology [7,8],
and the magnetoencephalogram (MEG) and magnetocardiogram (MCG) [9,10] in the field of biomedical
science, etc. The traditional handheld magnetic survey can acquire high-density magnetic data, but it is
time-consuming, labor-intensive, and limited by the terrain. The fixed-wing airplane-based magnetic
survey can realize fast and wide-range measurement, but the spatial resolution of magnetic is low and

Remote Sens. 2020, 12, 452; doi:10.3390/rs12030452 www.mdpi.com/journal/remotesensing

Remote Sens. 2020, 12, 452 2 of 19

only suitable for the detection of large targets. With the advance in the small Unmanned Aerial Vehicle
(sUAV), UAV system provides an excellent platform in achieving low-cost, low-altitude, long-range,
fast magnetic field monitoring [11–14], compared with the terrestrial, manned aircraft, and satellite
magnetic survey, which opens a brand-new era in using UAV-borne magnetic survey.
This paper focuses on the remote detection of near-surface targets, like UXO and steel pipe,
using the UAV magnetic measurement system. The UAV magnetic survey shows significant advantages
in the magnetic spatial resolution, work efficiency and experimental safety. However, UAV-borne
magnetic survey is facing a lot of tough challenges due to its limited load capacity, short battery life,
interference field generated by the drone, uncertainty of the position of the magnetic sensors, and
so on.
Recently, lots of researchers have devoted themselves to the development of the
UAV-magnetometer sensing devices, a variety of miniaturized, lightweight magnetic sensors and
acquisition systems have been installed to the drones, different field experiments have also been
carried out. Douglas et al. [15] developed an autonomous aeromagnetic system using a rotary-wing
UAV with a fluxgate magnetometer to detect the small ferrous minerals. A path planning algorithm
based on the lawnmower coverage pattern was proposed to cover the test region efficiently, the
magnetic contour map was obtained to reflect the distribution of subsurface anomalous targets.
Alireza Malehmir et al. [16] used the DJI S1000 rotary-wing UAV equipped with the GEM 19GW
magnetometer, a GPS antenna, and data recorder module to detect the iron-oxide deposits in central
Sweden. The field experiment showed that the UAV magnetic measurement system was able to clearly
delineate the mineralization with finer sampling, higher resolution over the conventional aeromagnetic
survey. Christoph Eck and Benedikt Imbachy [17] designed an aerial magnetic sensing scheme with the
Scout B1-100 UAV helicopter and a three-axis magnetic sensor, which was applied in the humanitarian
rescue. In [18], two Geometrics G-823A cesium vapor magnetometers, a computer, power supplies,
differential GPS, IMU, and fluxgate magnetometer are mounted on the fixed-wing UAV, called Venturer,
for the aeromagnetic surveying.
To date, considerable progress has been made in the integration and testing of the
UAV-magnetometer system. Whether it is based on the fixed-wing UAV [17,18] or multi-rotor
UAV [15,16,19], a single magnetometer [15–17] or multiple magnetometers [18], the UAV-magnetometer
system shows excellent performance over the handheld and aeromagnetic devices. However, the
research on the processing and interpretation of UAV magnetic data is quite rare and challenging
because of the low quality of magnetic data and the lack of interpretation methods for the UAV-based
magnetic survey. Most studies attempt to show the distribution of underground anomalies roughly
through the 2D magnetic contour map from the UAV magnetic data, but the 3D location of targets
cannot be obtained by the available interpretation ways. If a sophisticated processing routine for the
UAV magnetic survey is developed, the UAV-borne magnetic survey is bound to gain more attention
in scientific research and engineering applications.
One of the biggest obstacles in collecting high-quality magnetic data by the UAV-magnetometer
system is the interference magnetic field from the drone [18,19]. The drone system is not a non-magnetic
platform, some modules and structures are made of ferromagnetic materials, which have a serious
impact on the magnetometers attached to the drone and will degrade the measurement accuracy of
system. In fact, compensating the interference field originated from the flight platform is not a new
problem. Aeromagnetic compensation has been studied deeply for the traditional manned aerial
magnetic survey. The classic magnetic compensation scheme is the Tolles–Lawson model [20,21],
but this compensation scheme cannot be used in the drone systems. Because the aeromagnetic
compensation flight is performed in a few kilometers of altitude typically, providing a uniform
background field. That is not acceptable for the drones due to its general maximum flying height
is tens of meters, where the background magnetic field is spatially non-uniform for the existence of
geological and cultural sources. Now, a long rope or rod (3–5 m) is suspended below or in front
of the drone platform in [15–19] to reduce the effects of the drone on the sensors. However, this
Remote Sens. 2020, 12, 452 3 of 19

long rod structure is prone to an unbalanced system and low flight stability, which is not an ideal
solution. The compensation of UAV interference field is an open problem [18,19] and still requires
further research.
Although there have been numerous studies on the processing and interpretation of magnetic
field data for the ground-based magnetic survey, these techniques cannot be directly applied to the
UAV-based magnetic survey. Marc Munschy et al. [5] put forward a complete processing framework that
is based on the handheld multi-sensor fluxgate magnetometers for the UXO detection. The calibration
operation was firstly conducted to eliminate nine errors from the fluxgate sensor. Then, analytical signal
method and magnetic dipole inversion were adopted to interpret the magnetic data. The analytical
signal method is able to estimate the location of objects by the first-order and second order magnetic
gradient field, which requires the high signal-to-noise (SNR) magnetic field data [5,22]. The magnetic
dipole inversion is capable of determining the position and the 3D magnetic moment vector of the
targets at the same time, but the magnetic dipole inversion is not an automatic processing technique
for the survey data. It is necessary to manually extract the anomaly field generated by an isolated
target, and the nonlinear inversion method depends on the initial value, so the calculation time is
long [5,23]. By contrast, Euler deconvolution is more appropriate to achieve automatic interpretation
of survey data [24]. A fully automatic detection of UXO for the terrestrial magnetic survey has been
proposed by Kristofer Davis et al. [25,26], which was essentially the sliding window (SW)-based Euler
deconvolution process. Euler deconvolution is a rapid way to estimate the location of the isolated target,
but due to the size of sliding window is changing in an estimated range [25–27], the entire process is
complex and cumbersome, increasing the processing time. Furthermore, the result of SW-based Euler
method depends on the attenuation of the magnetic field, has nothing to do with the strength of the
magnetic field, resulting in a large number of false anomalies, leading to the rise in the number of false
targets. Usman et al. [28] proposed a few filter techniques to suppress the unreliable Euler solutions by
constraining the amplitude of analytical signal after the SW-based Euler deconvolution.
In this paper, we introduce the basic configuration of UAV-magnetometer system and put forward
a complete data processing routine for the UAV-borne magnetic survey, including the elimination
of regional background field, the removal of interference field originated from the UAV platform,
magnetic mapping, automatic detection and positioning of subsurface targets. Among them, the
elimination of UAV interference field and automatic detection and positioning method are highlighted
for the UAV-based magnetic survey. Unlike the traditional aeromagnetic compensation scheme based
on the attitude information, a two-channel linear time-invariant (LTI) system is modelled [29,30] and
the spatial correlation of magnetic field signals sampled by two closely spaced sensors is employed
to separate the magnetic anomaly signal and the UAV interference field from the total field signal.
Besides, an improved Euler deconvolution method is presented to detect the underground targets
automatically, which is the combination of YOLOv3 (You Only Look Once version 3) and Euler
deconvolution algorithm. YOLOv3 is a deep convolutional neural network (DCNN) based artificial
intelligence algorithm and excels in image and video detection [31]. Here, YOLOv3 is applied in the
context of magnetic survey for the first time, replacing the traditional sliding window process, showing
better performance in reducing false targets and improving the positioning accuracy. In addition, the
improved Euler deconvolution is simpler and faster, which is quite suitable for the large-scale UAV
magnetic survey.
The paper is organized as follows. Section 2 presents the configuration of UAV-magnetometer
system and the workflow for the UAV-borne magnetic survey, all procedures are briefly explained.
Section 3 introduces the key techniques of the magnetic data processing; the compensation of UAV
interference field and the improved Euler deconvolution method are described in detail. The field
tests and experimental results are discussed in Section 4. Section 5 gives the discussion and remarks
regarding the proposed framework. Section 6 provides the conclusions of this research.
Remote Sens. 2020, 12, 452 4 of 19

2. System and Workflow

2.1. UAV-Magnetometer System

A novel UAV magnetic measurement system is developed for the detection of near-surface targets.
The UAV-Magnetometer system consists of two magnetometers, radar altimeter, differential GPS, data
Remote Sens.
recording 2019, 11,and
module x FORpower
PEER REVIEW
module, as shown in Figure 1. The flight platform is the eight-rotor 4 of 19
DJI MG1 unmanned aerial vehicle. Two Cs optically pumped magnetometers are mounted on the
GPS,ofdata
center drone recording module
by a vertical andthe
boom, power module,
vertical as shown
distance between in Figure 1. Theisflight
two sensors 0.3 m.platform is the
The Cs optically
eight-rotor DJI MG1 unmanned aerial vehicle. Two Cs optically pumped magnetometers are
pumped magnetometer is a lightweight, small-sized CAS-L3 designed for the drones by the Aerospace
mounted on the center of drone by a vertical boom, the vertical distance between two sensors is 0.3
Information Research Institute, Chinese Academy of Sciences, China. The operating range of CAS-L3
m. The Cs optically pumped magnetometer is a lightweight, small-sized √ CAS-L3 designed for the
is 15,000 nT to 105,000 nT and the noise sensitivity is 0.6 pTrms Hz⁄@1 Hz in the shielded room.
drones by the Aerospace Information Research Institute, Chinese Academy of Sciences, China. The
Theoperating
radar altimeter is RD2412R, manufactured by SZ DJI Technology Co., Ltd., China. It measures
range of CAS-L3 is 15, 000 nT to 105, 000 nT and the noise sensitivity is 0.6 pTrms√Hz⁄@1
theHzheight of the drone
in the shielded room. above
Thethe ground
radar ensuring
altimeter that themanufactured
is RD2412R, UAV can actively
by SZavoid obstacles and
DJI Technology Co., fly
safely in the complex terrain. Differential GPS T300, produced by Shanghai Sinan Satellite
Ltd., China. It measures the height of the drone above the ground ensuring that the UAV can actively Navigation
avoid obstacles
Technology and is
Co., Ltd., flyused
safely
toinacquire
the complex terrain.of
the location Differential GPS T300,
UAV, positioning produced
accuracy by the
is on Shanghai
order of
cm.Sinan Satellite Navigation Technology Co., Ltd., is used to acquire the location of UAV, positioning
Magnetic field data and position information are synchronized by the pps (pulse per second)
accuracy
signal. The is on the order of cm.
custom-designed Magnetic
data field module
recording data and is
position information
constructed are synchronized
and fixed on the bottom by the
of the
UAV ppssystem.
(pulse perAllsecond)
modulessignal. Thebeen
have custom-designed data recording
installed reasonably module
to ensure theisstability
constructedandand fixed of
balance
theon the bottom of the UAV
UAV-magnetometer system.
system; theAll modules
entire have been
payload weighsinstalled
4.37 kg reasonably
meeting to theensure the stability
requirement of the
and balance of
maximum load (13.7 kg). the UAV-magnetometer system; the entire payload weighs 4.37 kg meeting the
requirement of the maximum load (13.7 kg).

Figure
Figure 1. The
1. The UnmannedAerial
Unmanned AerialVehicle
Vehicle (UAV)-magnetometer
(UAV)-magnetometer system,
system,carrying
carryingtwo magnetometers,
two magnetometers,
radar
radar altimeter,
altimeter, GPS,data
GPS, datarecording
recordingand
and power
power module.
module. The
Thetop
topmagnetic
magneticsensor is is
sensor labelled as as
labelled
magnetometer
magnetometer 1 and
1 and the
the bottommagnetic
bottom magneticsensor
sensor is
is labelled
labelled as
asmagnetometer
magnetometer2.2.

2.2.2.2.
Workflow forfor
Workflow UAV-Borne
UAV-BorneMagnetic
MagneticSurvey
Survey
Applied
Applied to the
to the near-surface
near-surface targets
targets detection,
detection, the the workflow
workflow of theof UAV
the UAV magnetic
magnetic survey survey is
is divided
intodivided into three
three stages, the stages, theisfirst
first one dataone is data collection,
collection, includingincluding the preparation
the preparation of test
of test area, thearea, the of
setting
setting
flight pathofand
flight path and
drone; drone; the
the second second
phase phase
is data is data processing,
processing, involvedinvolved someprocessing
some signal signal processing
methods
methods to obtain high quality magnetic field data, the final stage is data interpretation,
to obtain high quality magnetic field data, the final stage is data interpretation, achieving the automatic achieving
the automatic
detection detection
and position and position
estimation estimation
of buried of buried
targets. targets. The
The complete complete
workflow is workflow is listed2.in
listed in Figure
Figure 2.
The mission planning is the preliminary task of the magnetic survey, comprising the setting of
test area,The missionand
profiles, planning is theFor
the drone. preliminary task of
the planning ofthe
themagnetic
test site, survey,
the regioncomprising the setting
of interest, of
topography,
test area, profiles, and the drone. For the planning of the test site, the region of
vegetation, weather and other factors need to be evaluated carefully. For the planning of the profile,interest, topography,
thevegetation, weather and other factors need to be evaluated carefully. For the planning of the profile,
flight direction, length, and spacing should be considered. In addition, the working mode of UAV
the flight direction, length, and spacing should be considered. In addition, the working mode of UAV
should be selected, such as manual or automatic flight, flying speed, above ground level (AGL), battery
should be selected, such as manual or automatic flight, flying speed, above ground level (AGL),
life. In short, the perfect mission planning lays the foundation for the UAV magnetic survey, ensuring
battery life. In short, the perfect mission planning lays the foundation for the UAV magnetic survey,
that the drone can fly normally and collect the magnetic data.
ensuring that the drone can fly normally and collect the magnetic data.
Remote Sens. 2020, 12, 452 5 of 19
Remote Sens. 2019, 11, x FOR PEER REVIEW 5 of 19

Figure2.
Figure 2. The
The workflow
workflow for
forUAV
UAVmagnetic
magneticsurvey.
survey.

The purpose
The purposeofofdata
dataprocessing
processing is
is to
to reduce the noise
reduce the noisesignal,
signal,extract
extract magnetic
magnetic anomaly
anomaly signal,
signal,
improve
improve the signal-to-noise ratio (SNR). This stage contains the following actions intended for for
the signal-to-noise ratio (SNR). This stage contains the following actions intended
different
different functions:
functions:
Background
•  Background fieldelimination:
field elimination:TheThe background
background field
fieldconsists
consistsofofthe
thelocal geomagnetic
local geomagnetic field andand
field
the magnetic field produced by the ambient sources, like power grid, traffic, buildings. We are
the magnetic field produced by the ambient sources, like power grid, traffic, buildings. We are
concerned with the magnetic anomaly signal from the underground objects, so the background
concerned with the magnetic anomaly signal from the underground objects, so the background
field should firstly be removed.
field should firstly be removed.
 UAV interference field removal: Unlike the interference field from the external environment,
• UAVUAV interference
interferencefield removal:
field Unlikenoise
is an inherent the interference
signal from field from
system, the external
which is relatedenvironment,
to the attitude UAV
interference field
of the drone. is anwe
Herein, inherent
propose noise signal from
a calibration methodsystem,
basedwhich
on the is related
signal to the attitude
correlation to separateof the
the magnetic anomaly signal from the total field signal.
drone. Herein, we propose a calibration method based on the signal correlation to separate the
 Data gridding: It is customary to perform data gridding, which is to compute the magnetic field
magnetic anomaly signal from the total field signal.
• of regular
Data gridItnodes
gridding: from theto
is customary irregularly distributed
perform data sampling
gridding, whichpoints by interpolation.
is to compute Thus, field
the magnetic a
two-dimensional
of regular grid nodes contour
from magnetic map isdistributed
the irregularly produced tosampling
reflect thepoints
abnormal distribution ofThus,
by interpolation. the a
entire test area. contour magnetic map is produced to reflect the abnormal distribution of the
two-dimensional
The test
entire aimarea.
of the magnetic data interpretation is to estimate the location of underground targets
based on the spatial magnetic field data. The input data is the gridded magnetic data and a magnetic
The aim
contour mapof the
(RGBmagnetic data
image) of theinterpretation
test area, the isYOLOv3-based
to estimate theEuler
location of underground
deconvolution methodtargets
is
developed to determine the 3D location of the targets.
based on the spatial magnetic field data. The input data is the gridded magnetic data and a magnetic
contour map (RGB image) of the test area, the YOLOv3-based Euler deconvolution method is developed
3. Materialsthe
to determine and3DMethods
location of the targets.

3.1. Background
3. Materials Field Elimination
and Methods
The Earth’s
3.1. Background magnetic
Field field is a natural magnetic field, its source is classified as two categories:
Elimination
internal and external [32]. The magnetic field originated from the internal sources is the main
component,
The Earth’swhich generally
magnetic behaves
field as a stable
is a natural dipolarfield,
magnetic field.its
The magnetic
source field originated
is classified as twofrom the
categories:
external
internal sources
and contains
external [32].geomagnetic
The magnetic pulsations, magnetic bays,
field originated frommagnetic storms,sources
the internal and so on, which
is the main
are generated
component, by generally
which the complex motionasofa charged
behaves particles
stable dipolar from
field. themagnetic
The solar wind interacting
field with
originated the the
from
Earth’ssources
external ionosphere, behaving
contains as the time-varying
geomagnetic pulsations, geomagnetic
magnetic bays, signal. The Earth’s
magnetic storms, magnetic
and so field in
on, which
arethe local region
generated is approximated
by the complex motion as a of
uniform
charged andparticles
stable background
from the solarfieldwind
for a interacting
short time. Apart
with the
from ionosphere,
Earth’s the geomagnetic field,asthe
behaving theobserved field geomagnetic
time-varying is affected bysignal.
variousThe
high-frequency
Earth’s magneticinterference
field in the
local region is approximated as a uniform and stable background field for a short time. Apart from the
Remote Sens. 2020, 12, 452 6 of 19

geomagnetic
Remote Sens.field, the
2019, 11, observed
x FOR field is affected by various high-frequency interference fields6 of
PEER REVIEW from
19
Remote Sens. 2019, 11, x FOR PEER REVIEW 6 of 19
the surrounding environment, like the 50 Hz alternating magnetic field produced by the power grid.
fields from the
Thefields
background surrounding environment, like the 50 Hz alternating magnetic field produced byabout the
from thefield should be
surrounding removed because
environment, like thethey
50 Hz doalternating
not indicate any useful
magnetic field information
produced by the
power
the power grid.
underground The background
targets. field should be removed because they do not indicate any useful
grid. The background field should be removed because they do not indicate any useful
information aboutof the underground targets.
information about thebackground
The elimination undergroundfield targets.
includes two steps, one is the low-pass filter to remove the
The elimination of background field includes two steps, one is the low-pass filter to remove the
The elimination
high-frequency interferenceof background field the
magnetic field; includes
othertwo
is steps,
the one is the
detrending inlow-pass
the filter
spatial to remove
domain the
to subtract
high-frequency interference magnetic field; the other is the detrending in the spatial domain to
the high-frequency
geomagnetic field. interference magnetic data
The raw magnetic field;collected
the otheralong
is thea profile
detrending
and thein the spatial
results afterdomain to
eliminating
subtract the geomagnetic field. The raw magnetic data collected along a profile and the results after
subtract the geomagnetic field. The
the background field are displayed in Figure 3. raw magnetic data collected along a profile and the results after
eliminating the background field are displayed in Figure 3.
eliminating the background field are displayed in Figure 3.

FigureFigure 3. The
3. The elimination
elimination of background
of background field.(a)
field. (a)The
Theraw
rawmagnetic
magnetic field along
along the
theprofile;
profile;(b)
(b)The
The
Figure 3. The elimination of background field. (a) The raw magnetic field along the profile; (b) The
filtered
filtered magnetic
magnetic fieldfield
by aby a low-pass
low-pass filter
filter with fc =
with 𝑓 3=Hz.
3 Hz.
(c)(c)
TheThe detrendedmagnetic
detrended magneticfield.
field.
filtered magnetic field by a low-pass filter with 𝑓 = 3 Hz. (c) The detrended magnetic field.
3.2. UAV Interference
3.2. UAV Field
Interference Removal
Field Removal
3.2. UAV Interference Field Removal
The interference
The interference fieldfield generated
generated by thebyUAV
the platform
UAV platform is same
is in the in thefrequency
same frequency
band asband as the
the anomaly
The interference field generated by the UAV platform is in the same frequency band as the
fieldanomaly field we are interested in, so it cannot be suppressed by the filter. A two-channel linear time-
we are interested in, so it cannot be suppressed by the filter. A two-channel linear time-invariant
anomaly field we are interested in, so it cannot be suppressed by the filter. A two-channel linear time-
(LTI)invariant
model is(LTI) model is
established established
based based on the configuration
on the configuration of UAV-magnetometer
of UAV-magnetometer system, asin
system, as illustrated
invariant (LTI) model is established based on the configuration of UAV-magnetometer system, as
Figureillustrated
4 [30]. in Figure 4 [30].
illustrated in Figure 4 [30].

Figure
Figure 4. The
4. The signal
signal modelfor
model forthe
theUAV-magnetometer
UAV-magnetometer system.
system.
Figure 4. The signal model for the UAV-magnetometer system.

All signal
All signal is sampled
is sampled at at
thethe samplingrate
sampling rateofof f 𝑓, n, nindicates
indicatesthe
thesampling
sampling points,
points, thetheinput
inputofof
All signal is sampled at the sampling rate of 𝑓s , n indicates the sampling points, the input of
LTI a(n)
a(n)
LTI LTI denotes
denotes the the magnetic
magnetic anomalyfield,
anomaly field,i(n)
i(n) denotes
denotes the the UAV
UAV interference
interferencesignal, thethe
signal, outputs of
outputs
a(n) denotes the magnetic anomaly field, i(n) denotes the UAV interference signal, the outputs of
LTI
of LTI s (𝑠n(𝑛)
) and
and s (𝑠n (𝑛)
) areare
thethe measured
measured total
total field
field signals
signals by
by the
the magnetometer
magnetometer 11and
and 2, respectively.
2, respectively.
LTI 1𝑠 (𝑛) and 2𝑠 (𝑛) are the measured total field signals by the magnetometer 1 and 2, respectively.
The magnetometer
TheThe 1 is to
closer to the interference sources than magnetometer 2; the
therefore, the
magnetometer
magnetometer 1 is closer
1 is closer to the interference
the interference sources sources
than than magnetometer
magnetometer 2; therefore,
2; therefore, the
interference
interference field measured by magnetometer 1 is k times the interference field measured by
interference field measured by magnetometer 1 is k times the interference field measured by
magnetometer 2. The signal model is expressed as follows:
magnetometer 2. The signal model is expressed as follows:
Remote Sens. 2020, 12, 452 7 of 19

field measured by magnetometer 1 is k times the interference field measured by magnetometer 2.

The signal model is expressed as follows:

s1 (n) = a(n) + k ∗ i(n)

(1)
s2 (n) = a(n) + i(n)

Remote Sens.
The differential 2019, 11,
signal x FOR PEER
between REVIEW
two sensors d(n) is: 7 of 19

 n) = a(k n− 1)k∗i(in )n 

ds1(n
(1)
(2)
s2  n   a  n   i  n 
The interference signal is uncorrelated with the magnetic anomaly signal, so the outputs of correlating
The differential signal between two sensors d(n) is:
d(n) with s1 (n). and s2 (n) are:
d  n    k  1  i  n  (2)
PN
d(n) = k (k − 1) N 2 )
P
r1 =
The interference signal is 1 s1 ( n )
uncorrelated
Pi= i=1 i (nsignal,
with the magneticPanomaly so the outputs of (3)
N N 2
= 𝑠 i(𝑛)
r2 and
correlating d(n) with 𝑠 (𝑛) 2 (n )d (n ) = (k − 1 ) i=1 i (n )
=1 sare:

r1   i  1 s1  n  d  n   k  k  1 i  1 i 2  n 
N N

Consequently, the transfer function of kNis estimated as: (3)

 s 2  n  d  n    k  1  i  1 i 2  n 
N
r2  i 1
r
k = 1as:
Consequently, the transfer function of k is estimated (4)
r2
r
k  1 (4)
Then, the magnetic anomaly field and the interference r2 field are separated from the total signal:
Then, the magnetic anomaly field and the interference field are separated from the total signal:
d(n) s (n)−s (n)
i(n) = k−1 = 1 k−12
d  n s1  n   s2  n  (5)
ai(nn) =
 s2 (n)− i(n) = ks2 (n)−s1 (n)
k 1 k 1 k−1 (5)
ks2  n   s1  n 
a  n   s2  n   i  n  
The results after removing the UAV interference field are k illustrated
1 in Figure 5, Figure 5a,b depict
two raw magnetic fieldafter
The results signals
removingsampled
the UAVsynchronously byillustrated
interference field are two magnetometers
in Figure 5, Figure(The background
5 (a) and
(b) depict
magnetic field two raw magnetic
is removed field signals
first), Figure sampled synchronously
5c,d depicts the extractedbyinterference
two magnetometers (The magnetic
field and
background
anomaly field. magnetic field
The magnetic is removed
anomaly fieldfirst),
andFigure
UAV 5interference
(c) and (d) depicts
fieldthe extracted interference
is superimposed and the UAV
field and magnetic anomaly field. The magnetic anomaly field and UAV interference field is
interference field from the magnetometer 1 even masks the anomaly signal completely. After the
superimposed and the UAV interference field from the magnetometer 1 even masks the anomaly
compensation,
signalthe magnetic
completely. anomaly
After signal and
the compensation, thethe interference
magnetic anomalyfield are
signal andseparated from field
the interference the total field,
are anomaly
the extracted separated from
fieldthe total field,two
indicates the extracted
potentialanomaly
targets,field indicates
which is intwo potential targets,
agreement with thewhich
actual scene.
is in agreement with the actual scene.

Figure 5. The removal of the UAV interference field. (a) The outputs of magnetometer 1; (b) The
outputs of magnetometer 2; (c) The separated UAV interference field; (d) The separated magnetic
anomaly field.
Remote Sens. 2020, 12, 452 8 of 19

3.3. Data Gridding

Although the flight path for the drone has been planned in advance, the sampling points are
still irregularly distributed on account of the restrictions of flight control accuracy and the effect of
wind. Gridded magnetic data is the premise of magnetic mapping. The process of data gridding is
to interpolate the magnetic field on the regular grid nodes based on the actual observation points.
There are well-established data interpolation methods for the magnetic survey, such as kriging [33–35],
spline [36], minimum curvature [37], and so forth. Ultimately, the gridded magnetic data over the test
area and a 2D magnetic contour map (RGB image) can be obtained.

3.4. YOLOv3-based Euler Deconvolution

The interpretation of magnetic data is based on an improved Euler deconvolution that is the
combination of YOLOv3 and Euler deconvolution. YOLOv3 is used to pick up the anomaly field
automatically, Euler deconvolution is used to identify the location of buried objects based on the
magnetic data chosen by YOLOv3. The complete routine of YOLOv3-based Euler deconvolution and the
comparison with the traditional sliding window (SW)-based Euler deconvolution are presented below.

3.4.1. Euler Deconvolution

The magnetic anomaly field satisfies the Euler’s homogeneity equation, described in Cartesian
coordinate as follows [24–27]:

∂T ∂T ∂T
(x − x0 ) + ( y − y0 ) + (z − z0 ) = −NT (x − xo , y − yo , z − zo ), (6)
∂x ∂y ∂z

where (x0 , y0 , z0 ) denotes the location of anomaly source, (x, y, z) denotes the location of observation
point, T,∂T/∂x,∂T/∂y,∂T/∂z denote the anomalous field and its gradient fields in -x, -y, -z direction.
N is the structure of index (SI), characterizing the attenuation rate of the amplitude of the anomalous
field with the distance, which depends on the type of the sources. For the detection of near-surface
targets, the anomaly source such as UXO is approximated to a point-like dipole, the SI N is generally
between 2.5 and 3 [24–27].
Assuming a prior value N, the location of anomaly source can be solved by establishing the linear
equation from multiple observation points above the target, as given in (7), (8):
 −1
X = AT A AT b, (7)

where
∂T1 ∂T1 ∂T1
 x1 ∂T + y1 ∂T + z1 ∂T
   
1 1 1

 ∂x ∂y ∂z

  ∂x ∂y ∂z
+ NT1   
∂T2 ∂T2 ∂T2 ∂T2 ∂T2
+ z2 ∂T  x0 

 x2 ∂x + y2 ∂y 2
+ NT
   
 ∂x ∂y ∂z
 ∂z 2 
  
A =  , b =  , X =  y0 . (8)
    
.. .. .. ..
. . . .
     
    z0
∂Tn ∂Tn ∂Tn  x ∂Tn + y ∂Tn + z ∂Tn + NT 
   

 
∂x ∂y ∂z n ∂x n ∂y n ∂z n

In theory, the observation points can be selected arbitrarily, as long as the anomaly field at the
observation points is generated by an isolated source.

3.4.2. YOLOv3
YOLOv3 is the third generation YOLO algorithm, is a more accurate, stable real-time target
detection method, which is widely used in the image and video detection. Compared with the previous
YOLO and YOLO9000, the major improvements of YOLOv3 are the category prediction changed from
the single label to the multi-label, and the prediction using multiple scale fusion methods, which are
the main reasons we choose YOLOv3, for more detailed information about the YOLOv3, we refer
to [31].
 Remote Sens. 2019, 11, x FOR PEER REVIEW 9 of 19

YOLOv3
Remote Sens. 2020, 12, 452is the third generation YOLO algorithm, is a more accurate, stable real-time target
9 of 19
detection method, which is widely used in the image and video detection. Compared with the
previous YOLO and YOLO9000, the major improvements of YOLOv3 are the category prediction
changed
Here, from the
the input imagesingle label
is the to the contour
magnetic multi-label,
mapand the survey
of the prediction area.using
Unlike multiple scale fusion
the generic object
methods, which are the main reasons we choose YOLOv3, for more detailed information about the
detection, YOLOv3 is used to detect the anomalous field generated from ferrous targets. The ferrous
target YOLOv3, we refer a
is approximately tomagnetic
[31]. dipole, when the distance between sensor and target is more than
three timesHere, the input size
the maximum image of is the magnetic
target. contour
The anomaly map
field of the survey
produced area. Unlike
by a magnetic the generic
dipole object
is the bipolar
detection,
anomalous fieldYOLOv3 is used
with positive to detect
and the anomalies
negative anomaloustypically.
field generated from ferrous
Considering targets. The
the existence ferrous
of noise or
target is approximately a magnetic dipole, when the distance between
other interference fields, the measured anomalous field might have only one positive anomaly or onesensor and target is more than
three
negative times the
anomaly maximum
in the size of
actual tests, target.
rather The
than theanomaly
bipolar field
anomalyproduced by a magneticConsequently,
field theoretically. dipole is the
bipolar
the targets to anomalous
be tested are field with positive
divided into twoandcategories:
negative anomalies
positive typically.
anomalyConsidering
(labelled asthe existence
Positive), of
and
noise or other interference fields, the measured anomalous field might have only one positive
negative anomaly (labelled as Negative).
anomaly or one negative anomaly in the actual tests, rather than the bipolar anomaly field
Besides, the size of the window that circles the anomalous field needs to be carefully assessed,
theoretically. Consequently, the targets to be tested are divided into two categories: positive anomaly
here are some strategies that need to be met:
(labelled as Positive), and negative anomaly (labelled as Negative).
• A windowBesides, the size
contains at of thethree
least window that circles
observation the anomalous
points because there fieldare
needs
three tounknown
be carefully assessed,
parameters
inhere are some
Equation (7).strategies that need to be met:
 A window contains at least three observation points because there are three unknown
• Based on the principal of Euler deconvolution, the anomaly data within the windows must be
parameters in Equation (7).
generated from an isolated object, avoiding the effects from the adjacent sources.
 Based on the principal of Euler deconvolution, the anomaly data within the windows must be
• For the real data, the higher SNR of magnetic data is, the more reliable the results of Euler
generated from an isolated object, avoiding the effects from the adjacent sources.
deconvolution
 For the real are.data,
The size of window
the higher SNR is ofdetermined
magnetic data by ais,threshold
the morethat is defined
reliable as theofratio
the results of
Euler
the maximum
deconvolution and minimum
are. The size values of the magnetic
of window is determinedfieldbyinaathreshold
window.that Theisanomaly
defined as field
the with
ratio
lower ofSNR the maximum and minimum values of the magnetic field in a window. The anomaly field
is removed by the setting of threshold, which is set to 50% here.
with lower SNR is removed by the setting of threshold, which is set to 50% here.
The key points for applying YOLOv3 to the magnetic survey are as follows:
The key points for applying YOLOv3 to the magnetic survey are as follows:
Dataset: All samples in the dataset are in picture format. The dataset is composed of a large
Dataset: All samples in the dataset are in picture format. The dataset is composed of a large
number of
number2D of magnetic
2D magneticcontour maps,
contour divided
maps, intointo
divided 85%85%training
training set,set,
15%15%test
testdata
dataset.
set.The
Thechanging
changing
number, size, position,
number, depth,
size, position, azimuth,
depth, declination,
azimuth, remanence
declination, remanence of of
thethetargets
targetsandandbackground
background fields fields
make the
makemagnetic
the magneticmapsmapsdifferent. The The
different. training set set
training contains 850850
contains images
images with
with3860
3860instances
instancesof of labeled
labeled
anomalies,
anomalies, the testthe settest set contains
contains 150 with
150 images images710with 710anomalies.
labeled labeled anomalies. Figure some
Figure 6 shows 6 shows some
examples
of the examples
dataset. of the dataset.

Figure 6. Partial samples and labelled anomalies in the dataset of You Only Look Once version 3
Figure 6. Partial samples and labelled anomalies in the dataset of You Only Look Once version 3
(YOLOv3). Each image is a 2D magnetic contour map, the unit of magnetic field in all figures is
(YOLOv3). Each image is a 2D magnetic contour map, the unit of magnetic field in all figures is nano
nano Tesla (nT), the targets to be detected are divided into two types: Positive and Negative. (a), (b),
and (c) show three magnetic contour maps generated by the same target with different azimuths and
declinations, remanence, and locations. (d), (e) and (f) show three magnetic contour maps generated
by different numbers and sizes of targets, background magnetic fields.
Remote Sens. 2020, 12, 452 10 of 19

Detection: YOLOv3 predicts bounding boxes at 3 different scales, the targets to be detected are
classified as two classes. Batch size is set to be 12, momentum and weight decay are set to be 0.9
and 0.0005, respectively. The initial learning rate is set to be 0.0001. Finally, the output for predicted
bounding box contains 6 values (top, left, bottom, right, score, class). We are more interested in the first
4 parameters, which provide the position and size of window. The ‘score’ reflects the confidence of the
box, combining location and category confidence; the ‘class’ indicates the category of the predicted box.
Evaluation: The trained YOLOv3 yields 95.92% mAP for two classes (Positive and Negative), and
the mAP is 93.87% for the test sets, indicating the reliable detection performance for the Positive and
Negative anomaly.

3.4.3. Summary of YOLOv3-based Euler Deconvolution

The processing flow of the improved Euler deconvolution is presented in Algorithm 1. Theoretically,
the Euler solutions computed by the positive anomaly field or the negative anomaly field are the same,
when the anomalous field is from the same isolated target. In the actual test, if the deviation of the
Euler solutions obtained by the two types of abnormal field is less than 30 cm, they can be clustered
into the same target.

Algorithm 1: YOLOv3-based Euler Deconvolution

Input: Gridded magnetic data and a 2D magnetic contour map
Output: The location of underground targets
1: The magnetic contour map is inputted into the trained YOLOv3
2: A total of Nwindows anomaly windows are detected by YOLOv3, extract the magnetic field data inside each
window
3: for k = 1:Nwindows do
4: Determine the Euler solution Xeulerk based on the data within the window
5: end for
j j
i
6: Cluster, if Xeuler − Xeuler < 30 cm (Xeuler
i and Xeuler are Euler solutions obtained by Positive and Negative
anomaly windows).
7: return the location of potential targets, which is same as the clustered results

For the traditional SW-based Euler deconvolution, a window slides as a fixed step and one moving
window gets one Euler solution until the test area is covered totally. All Euler solutions are clustered
into the potential targets and the fake targets are filtered out by the threshold of N, the workflow of
SW-based Euler method is summarized as Figure 7. It is worth noting that the size of window is
not a fixed value in the SW-based Euler method, it changes from small to large ensuring that targets
of different depth and size can be detected adaptively [24–27]. Because the sliding window is not
selective, the results of Euler deconvolution depends only on the decay rate of magnetic field.
Remote Sens. 2020, 12, 452 11 of 19
Remote Sens. 2019, 11, x FOR PEER REVIEW 11 of 19

Figure
Figure 7. The
7. The workflowofofsliding
workflow slidingwindow
window (SW)
(SW) Euler
Euler deconvolution
deconvolution[24],
[24],the
thesize ofof
size windows
windows varies
varies
from 3 to 25 grid points, and the “Filter” means filtering out the Euler solutions whose
from 3 to 25 grid points, and the “Filter” means filtering out the Euler solutions whose structure structure index
index
is less
N isNless thanthan
thethe threshold
threshold 2.2.2.2.

By By contrast,
contrast, thethe automaticselection
automatic selectionof
ofwindows
windows basedbasedon onYOLOv3
YOLOv3has hasconsiderable
considerable advantages
advantages
over the sliding window. Firstly, the windows obtained by YOLOv3 are related
over the sliding window. Firstly, the windows obtained by YOLOv3 are related to the strength to the strength of the
of the
magnetic field, overcoming the problem of false targets as a result of the SW method relies only on
magnetic field, overcoming the problem of false targets as a result of the SW method relies only on the
the attenuation characteristics of the magnetic field. Secondly, the size of windows chosen by
attenuation characteristics of the magnetic field. Secondly, the size of windows chosen by YOLOv3
YOLOv3 changes intelligently with the targets, a larger window is detected for the deeper objects,
changes intelligently with the targets, a larger window is detected for the deeper objects, and a smaller
and a smaller window is detected for the shallower objects. Whereas adjusting the size of windows
window is detected for the shallower objects. Whereas adjusting the size of windows is clumsy for the
is clumsy for the SW Euler deconvolution, the sliding window must cover the entire area before
SWchanging
Euler deconvolution,
the window size.the sliding
The sizewindow
of slidingmust coverchanges
window the entire area
from before
3 to changing
25 grids the window
in the SW-Euler
size. The size of sliding window changes from 3 to 25 grids in the SW-Euler
method, which is an empirical choice. The actual size of windows has a relation to the size method, which is an
of the
empirical
survey choice. The actual
area. Thirdly, size of windows
the magnetic field in has
the awindows
relation is
to high
the size
SNR ofaccording
the surveytoarea. Thirdly, the
the proposed
strategies, ensuring the accuracy of the positioning results. Furthermore, the Euler deconvolution
magnetic field in the windows is high SNR according to the proposed strategies, ensuring the is
accuracy
performed
of the positioningin the anomaly
results. windowsthe
Furthermore, predicted by YOLOv3 rather
Euler deconvolution than ininthe
is performed the entire
anomaly testwindows
area,
simplifying
predicted the processing
by YOLOv3 ratherflow
thangreatly, and thetest
in the entire YOLOv3 is able to detect
area, simplifying the the anomaly flow
processing targets in real-and
greatly,
thetime,
YOLOv3so theisimproved Eulerthe
able to detect method is faster
anomaly thaninthe
targets SW way,sowhich
real-time, is more suitable
the improved for the large-
Euler method is faster
scale magnetic survey.
than the SW way, which is more suitable for the large-scale magnetic survey.

4. Field
4. Field Experiment
Experiment

4.1.4.1. UAV-Borne
UAV-Borne Magnetic
Magnetic Survey
Survey
WeWehavehave tested
tested thetheUAV-magnetometer
UAV-magnetometer systemsystem andand carried
carriedout
outthethefield
fieldexperiments
experiments forfor
thethe
detection of near-surface targets in Hebei, China. The complete workflow proposed
detection of near-surface targets in Hebei, China. The complete workflow proposed in this paper has in this paper has
been
been performed
performed andand the
the interpretationresults
interpretation resultswere
were also
also given.
given.
A 28
A 28 mm × 33mmrectangular
× 33 rectangulararea
areawas
was selected
selected asasthe
thesurvey
surveyarea,
area,five ferrous
five ferrous objects with
objects different
with different
diameters and heights were pre-buried into the test area. The programmed
diameters and heights were pre-buried into the test area. The programmed flight profiles runflight profiles run along
along
the south-north direction, length is 62 m, spacing is 0.5 m. On a clear day, the UAV-magnetometer
the south-north direction, length is 62 m, spacing is 0.5 m. On a clear day, the UAV-magnetometer
system was set to fly automatically as the planned profiles with AGL 2 m, the flying speed was 1.9
system was set to fly automatically as the planned profiles with AGL 2 m, the flying speed was 1.9 m/s.
m/s. Figure 8 shows the flight path of the drone, the actual spacing between adjacent profiles is 0.3 -
Figure 8 shows the flight path of the drone, the actual spacing between adjacent profiles is 0.3–0.7 m,
0.7 m, limited by the effect of wind and the positioning accuracy of UAV.
limited by the effect of wind and the positioning accuracy of UAV.
The WGS-84 coordinate was transformed into the local cartesian coordinate system along the
The WGS-84
north, coordinate
east and down, was was
the origin transformed into
the starting theof
point local cartesian
the flight. The coordinate system
effective profiles along
above thethe
north,
region of interest has been obtained after cutting off the undesired flight and curved data (Figure 8).the
east and down, the origin was the starting point of the flight. The effective profiles above
region of interest has been obtained after cutting off the undesired flight and curved data (Figure 8).
Remote Sens. 2020, 12, 452 12 of 19
Remote Sens. 2019, 11, x FOR PEER REVIEW 12 of 19
Remote Sens. 2019, 11, x FOR PEER REVIEW 12 of 19

Figure
Figure 8. The
8. The planning
planning ofoftest
testsite
siteand
andflight
flightpath,
path, multiple
multiple parallel
parallel profiles
profilesalong
alongthe
thesouth-north
south-north
direction
Figure
direction cover
8. The
cover the
testtest
planning
the area.
of test
area. The
The “START”
site isisthe
and flight
“START” the starting
path, pointparallel
multiple
starting point of
of the
thedrone
droneininthe
profiles theexperiment
along and thethe
the south-north
experiment and
trimmed
direction
trimmed profiles
cover
profiles intest
thethe
in the local
area.
local coordinate
The “START”
coordinate are
are given
isgiven onthe
thepoint
the starting
on right.of the drone in the experiment and the
right.
trimmed profiles in the local coordinate are given on the right.
4.2. 4.2. Results
Results
4.2. Results
TheThe
rawraw magnetic
magnetic data
data waswas collected
collected continuously
continuously by by
twotwo magnetometers
magnetometers fs =
( 160
𝑓 =Hz),
160 Hz), and2D
and the
the 2D
magnetic Themagnetic
raw magnetic
contour mapcontour map
data was
is produced is produced
collected by theinterpolation
Krigingbyinterpolation
continuously
by the Kriging Figurein9.Figure
twoinmagnetometers 9.𝑓Four
Four(features = 160features
Hz),
are are
and
observed
the 2D magnetic
observed from contour
the map
magnetic is
maps produced
of raw by
data:the Kriging
1) The interpolation
measurement in Figure
inconsistency
from the magnetic maps of raw data: (1) The measurement inconsistency between adjacent lines results 9. Four
betweenfeatures are
adjacent
lines results
observed
in leveling frominthe
errors leveling
magnetic
[38], whicherrors
maps [38],
of which
is related raw is related
to data:
the 1) The to
attitude the attitudeinconsistency
measurement
information information
of the drone;of theAdrone;
between
(2) 2)linear
adjacent
subtle A
subtle
lines
trend linear
results
exists alongintrend exists
leveling
the profile along
errors
from[38],thewhich
south profile isfrom
to north; (3)south
related to north;
to the
The attitude
noise 3)information
signal The noise signal
measured themeasured
drone; 2) by
of magnetometer
by A1 is
subtle
higher linear
magnetometer
than trend
the noise exists
1 ismeasured
higheralongbythe
than theprofile
noise from
magnetometer 2;south
measured to magnetometer
by
(4) Only north; 3) The noise
two magnetic 2; 4) signal
Only are
anomalies measured
two by
magnetic
visible, other
magnetometer
anomalies are 1 is
visible,
magnetic anomalies are masked.higher
other than the
magnetic noise measured
anomalies are by
masked.magnetometer 2; 4) Only two magnetic
anomalies are visible, other magnetic anomalies are masked.

Figure 9. The magnetic map of raw magnetic field. (a) Magnetometer 1; (b) Magnetometer 2. (The unit
of magnetic
Figure
Figure 9. 9. The
The field
magnetic in all
magnetic figures
map
map is nano
ofofraw
raw Tesla field.
magnetic
magnetic (nT)).(a)
field. (a)Magnetometer
Magnetometer1;1;(b) (b)Magnetometer
Magnetometer 2. (The unitunit
2. (The
of magnetic field in all figures is nano Tesla (nT)).
of magnetic field in all figures is nano Tesla (nT)).
A lowpass filter with 𝑓 = 3 Hz was used to eliminate the high-frequency interference signal
A lowpass
firstly, then the
A lowpass filter
trend
filter with
with term = 3= Hz
fc 𝑓was 3removed
Hzwaswasused
used
to to eliminate
subtract
to eliminate the
the high-frequency
the regional geomagnetic field,
high-frequency interference signal
the processed
interference signal
firstly,
results
firstly, thenthen
arethethe trend
displayed
trend interm
term Figure
waswas10. removed
Compared
removed to subtract
to subtract
with the thethe regional
raw data, thegeomagnetic
regional curve field,
is smoother
geomagnetic the
in
field, theprocessed
themagnetic
processed
results
mapare
results are
formed displayed
displayed in
by theinfilteredFigure
Figure 10.Compared
magnetic
10. Compared with
field. The
with the
the raw
strips data,
in the
raw the
maps
data, curve
have
the is is
smoother
disappeared,
curve smoother in
andthe magnetic
in some
the weak
magnetic
map magnetic
map formed
formed anomaly
bybythethe signalsmagnetic
filtered
filtered are further
magnetic extracted.
field.
field. The strips
The Nonetheless,
strips in
inthe
themaps
mapsthehave
processed
have magnetic
disappeared,
disappeared, and signal
some
and someis
weakstill
weak
disturbed
magnetic by the
anomaly UAV
signalsinterference
are furtherfield, especially
extracted. for the
Nonetheless, outputs
the of magnetometer
processed
magnetic anomaly signals are further extracted. Nonetheless, the processed magnetic signal is still magnetic 1. signal is still
disturbed
disturbed bybythetheUAVUAV interferencefield,
interference field,especially
especially for for the
the outputs
outputsof ofmagnetometer
magnetometer1.1.
Remote Sens.Sens.
Remote 12, 452
2020,2019, 11, x FOR PEER REVIEW 13 13 of 19
of 19
Remote Sens. 2019, 11, x FOR PEER REVIEW 13 of 19

Figure
Figure 10. 10.
TheThe magnetic
magnetic map
map withthe
with thebackground
backgroundmagnetic
magnetic field
field removed.
removed.(a)
(a)The
Thefiltered
filteredmagnetic
magnetic
datadata from
Figure
from magnetometer
10. The magnetic map
magnetometer 1; 1; (b)
(b)with
TheThe
thefiltered
filtered magnetic
background
magnetic data from
magnetic
data magnetometer
field removed.
from 2. filtered
(a) The
magnetometer 2.(c) The magnetic
(c) Thedetrended
detrended
magnetic
data
magneticfrom data from magnetometer
magnetometer
data from 1; (b) The
magnetometer 1;
(d)(d)The
1;filtered The detrended
magnetic magnetic
data
detrended data from
from magnetometer
from magnetometer
magnetic 2.2.
2. (c) The detrended
magnetometer
magnetic data from magnetometer 1; (d) The detrended magnetic data from magnetometer 2.
UAV UAV interference
interference magnetic
magnetic fieldfield is suppressed
is suppressed through
through thethe proposed
proposed algorithmininSection
algorithm Section3.2,3.2,the
the UAV
compensated interference
compensated
magnetic magnetic
magnetic
anomaly fieldisfield
anomaly
field is suppressed
is in
given given inthrough
Figure Figure
11. The the
11. proposed
The
root algorithm
root mean
mean square square in(RMS)
(RMS) Section is 3.2,
errorerror is
4.9543
nTthe
andcompensated
4.9543 nT and
1.7714 magnetic
1.7714
nT for thenT anomaly
forsignal
total the totalfield is given
signal of in
of magnetometer Figure 11. The
magnetometer
1 and 1 root
and mean square
2, respectively,
2, respectively, the RMS(RMS) of error
the RMS isof
separated
4.9543
magnetic nT
separated and
anomaly 1.7714
magnetic
signalnTisfor
anomaly the
signal
reduced total
to signal nT
is reduced
0.5391 oftomagnetometer
0.5391
by the nT by the
proposed 1 and 2, respectively,
proposed
method, method,
indicating the RMS
indicating
that of
thethat
UAV
separated
the UAV magnetic
interferenceanomaly
field signal
has been is reduced
suppressedto 0.5391 nT by
significantly. the proposed
Compared method,
with
interference field has been suppressed significantly. Compared with the magnetic map of original data, the indicating
magnetic map thatof
the UAV
the quality interference
original of
data,
magnetic field
the quality
mapofof has been
magnetic
processed suppressed
map significantly.
of processed
signal has been signal Compared
has been
improved with
improved
a lot, the magnetic
which aislot,
able which map of
is ablethe
to reflect
original
to data,
reflect the the quality
distribution of
ofmagnetic
underground
distribution of underground anomalies clearly. map of processed
anomalies signal
clearly. has been improved a lot, which is able
to reflect the distribution of underground anomalies clearly.

Figure 11. The magnetic mapping after removing the UAV interference field.
Figure 11. The magnetic mapping after removing the UAV interference field.
Figure 11. The magnetic mapping after removing the UAV interference field.
Remote Sens. 2020, 12, 452 14 of 19

Remote Sens. 2019, 11, x FOR PEER REVIEW 14 of 19

Remote
TheSens. 2019, 11, x FOR
interpretation PEER REVIEW
results 14 of 19
by the YOLOv3-based and SW-based Euler deconvolution are presented
below. The interpretation
Firstly, the magnetic results by themap
contour YOLOv3-based
(Figure 11)and SW-based
of the surveyEuler area deconvolution
is inputted into arethe
The interpretation
presented below. Firstly, results
the by the
magnetic YOLOv3-based
contour map (Figure and
11) SW-based
of the survey Euler
area deconvolution
is inputted into are
thethe
trained YOLOv3 network, anomaly windows are detected in real-time, seven windows with
presented
trained below. Firstly,
YOLOv3 network, theanomaly
magneticwindowscontour map are (Figure
detected 11)
in ofreal-time,
the survey area is
seven inputtedwith
windows into the
the
box_confidence_score greater than 0.5 are found, including five positive anomalies and two negative
trained YOLOv3 network,
box_confidence_score greateranomaly
than 0.5windows
are found, are detectedfive in real-time, seven windows two with the
anomalies, as shown in Figure 12. The default SI Nincluding
is 2.5, seven positive anomalies
Euler solutions areand
determined negativebased
box_confidence_score
anomalies, as shown ingreater
Figure than
12. 0.5
The are
defaultfound,
SI Nincluding
is 2.5, five
seven positive
Euler anomalies
solutions are and two
determined negative
based
on anomalies,
the magnetic data ininthe
as shown
seven detected windows, and finally they are clustered into five potential
on the magnetic data in Figure
the seven 12. detected
The default SI N is 2.5,
windows, andseven
finallyEuler
they solutions
are clusteredare determined based
into five potential
underground
on the targets,
magnetic data as
in demonstrated
the seven detected in Figure
windows, 13a. At
and the same
finally they time,
are the same
clustered gridded
into five magnetic
potential
underground targets, as demonstrated in Figure 13 (a). At the same time, the same gridded magnetic
data was processed
underground by as
targets, thedemonstrated
SW-based Euler deconvolution
in Figure 13 (a). At the [25],
same the sizethe of same
sliding window changes
data was processed by the SW-based Euler deconvolution [25], thetime,
size of sliding gridded
windowmagnetic
changes
fromdata
fromwas
3
3 toprocessed
to 25 grids, by the SW-based
resulting
25 grids, resulting
in 5695 Euler deconvolution
Euler
in 5695 Euler
solutions [25],
totally.
solutions totally. the size
Then,
Then,
77 of sliding
Euler
77 Euler window
solutions
solutions changes
remain
remain
after
after
from
filtering
filtering3 toout
out 25 false
grids,anomalies
false resulting based
anomalies in 5695onEuler
based on the solutions of
threshold
the threshold totally.
of SI N
SI N isThen,
is 2.2. 77 Euler20
Finally,
2.2. Finally, solutions
20 clustered
clustered remain
targetsafter
targetsare
are
filtering
obtained
obtained out
with
withfalse
the anomalies
thecluster
clusterradius based
radius 50 on the
50 cm,
cm, thethreshold
the of results
localization
localization SIresults
N isare2.2.
are Finally,
given
given 20 Figure
in
in Figure clustered13b.
13 (b). targets
FiveFive are
white
white
obtained
circles ‘#’ with
indicate the
thecluster
true radius
position 50 cm,
of the the localization
buried targets results
measured are given
by the
circles ‘○’ indicate the true position of the buried targets measured by the Real-Time Kinematic (RTK) in Figure
Real-Time 13 (b). Five
Kinematic white(RTK)
circles ‘○’ indicate the true position of the buried targets measured by the
in advance, the positioning results of the improved Euler method are consistent with the true targets.
in advance, the positioning results of the improved Euler method are Real-Time
consistent Kinematic
with the true(RTK)
targets.
By in
By advance,
contrast,
contrast, the
thethepositioning
resultsofofthe
results results
the SW of
SWEuler thedeconvolution
Euler improved
deconvolutionEulercontain
method15
contain are
15 consistent
false
false targets,
targets, with the true
despite
despite thethetargets.
filtering
filtering
By
operationcontrast,
operation thebeen
having
having results
been of the SW Euler deconvolution contain 15 false targets, despite the filtering
done.
done.
operation having been done.

Figure 12. The anomaly windows detected by YOLOv3.

Figure 12.The
Figure12. Theanomaly windowsdetected
anomaly windows detectedby
byYOLOv3.
YOLOv3.

Figure
Figure 13.13.
TheThe estimatedpositions
estimated positionsofof two
two methods,
methods, white
whiteround
roundmarker
marker‘○’‘#’
is is
thethe
real location
real locationof of
Figure
buried
buried 13. The
targets
targets estimated
measured
measured bypositions
bythe of twoKinematic
theReal-Time
Real-Time methods, (RTK)
Kinematic white
(RTK)round marker
in advance,
in advance, ‘○’ is
black
black the ‘+’
cross
cross real location
marker
‘+’ marker ofthe
is the
is
buried
estimated targets
estimated measured
location.
location. (a)(a)Theby
The the Real-Time Kinematic
YOLOv3-based
YOLOv3-based (RTK) in advance,
Euler Deconvolution;
Euler Deconvolution; (b)
(b)The black
TheSW crossDeconvolution.
SWEuler
Euler ‘+’ marker is the
Deconvolution.
estimated location. (a) The YOLOv3-based Euler Deconvolution; (b) The SW Euler Deconvolution.
In In order
order to to evaluatethe
evaluate thepositioning
positioning accuracy
accuracy ofof the
the two
twomethods,
methods,we wecompare
compare thethe
estimated
estimated
In order
locations of to
five evaluate
real the
targets positioning
in Table 1 andaccuracy
give theof the two
quantitativemethods,
analysiswe
of compare
the the
positioningestimated
locations of five real targets in Table 1 and give the quantitative analysis of the positioning errorerror in in
locations of five real targets in Table 1 and give the quantitative analysis of the positioning error in
 related to the residual interference magnetic field and the truncation error caused by difference field
approximation to gradient field. The positioning accuracy of the proposed method is higher than the
SW-based Euler approach, especially for the depth. The depth of source is only correlated to its
vertical gradient field and N, but the depth and N are determined by the least squares at the same
time
Remote in 2020,
Sens. the SW method, there is mutual coupling between the two values, causing the increase of 15
12, 452 the
of 19
positioning error in the depth.

Table 1. The
Figure 14. The localization comparison
deviation of real
of the and estimated locations
YOLOv3-based by two methods.
Euler method is less than 0.3 m, which is
related to the residual interference magnetic field and the truncation error caused by difference field
YOLOv3-based Euler SW-based Euler Deconvolution
True Location
approximation (m) field. The positioning accuracy of the proposed method is higher than the
to gradient
Deconvolution (m) (m)
SW-based Euler approach, especially for the depth. The depth of source is only correlated to its vertical
(6.17,
gradient field and 33.96,
N, but the depth and N are determined by the least squares at the same time in the
1 (6.03, 34.19, 0.70) (5.98, 34.20, 0.11)
SW method, there 0.8)is mutual coupling between the two values, causing the increase of the positioning
error in the(6.31,
depth.26.61,
2 (6.32, 26.52, 0.75) (6.45, 26.52, 0.21)
1.0)
Table 1. The comparison of real and estimated locations by two methods.
(3.99, 22.15,
3 (4.01, 22.24,YOLOv3-based
0.10) Euler (3.85,SW-based
22.22, Euler
-0.40)
0.1) True Location (m)
Deconvolution (m) Deconvolution (m)
(17.16, 29.47,
4 1 (6.17, 33.96,(16.95,
0.8) 29.54, (6.03,
1.06)34.19, 0.70) (16.72,
(5.98,29.43, 0.11)
34.20, 0.11)
2 1.1) (6.31, 26.61, 1.0) (6.32, 26.52, 0.75) (6.45, 26.52, 0.21)
3
(18.07, 22.21, (3.99, 22.15, 0.1) (4.01, 22.24, 0.10) (3.85, 22.22, -0.40)
5 4 (17.16, 29.47,(17.78,
1.1) 22.18, (16.95,
0.56)29.54, 1.06) (17.96, 22.18,
(16.72, 29.43,-0.35)
0.11)
5 0.5) (18.07, 22.21, 0.5) (17.78, 22.18, 0.56) (17.96, 22.18, -0.35)

Figure 14. The positioning error of two methods. (a) The YOLOv3-based Euler Deconvolution; (b) The
Figure 14. The positioning error of two methods. (a) The YOLOv3-based Euler Deconvolution; (b)
SW-based Euler Deconvolution.
The SW-based Euler Deconvolution.

5. Discussion
5. Discussion
The magnetic measurement system for the UAV-borne magnetic survey is introduced, two Cs
The magnetic measurement system for the UAV-borne magnetic survey is introduced, two Cs
optically pumped
optically pumpedmagnetometers,
magnetometers, GPS,GPS,
radarradar
altimeter, powerpower
altimeter, and acquisition modulemodule
and acquisition are integrated
are
on the multi-rotor drone. This UAV-magnetometer system is superior to the traditional
integrated on the multi-rotor drone. This UAV-magnetometer system is superior to the traditional magnetic
measurement devices in devices
magnetic measurement [15,16,18,19] in two
in [15,16], ways:
[18,19] (1)ways:
in two Two (1)magnetic sensors
Two magnetic are placed
sensors in the
are placed
in the vertical direction, more magnetic field information is collected, which is beneficial for studyingthe
vertical direction, more magnetic field information is collected, which is beneficial for studying
the characteristics
characteristics of the of the magnetic
magnetic interference
interference fieldfield generated
generated by the
by the drone;
drone; (2)(2)
TheThe magneticsensors
magnetic sensorsare
aretofixed
fixed to thewith
the drone drone withrod,
a short a short
which rod, which guarantees
guarantees theofstability
the stability of system.
the flight the flight
Thesystem. The
construction
andconstruction and design of UAV-magnetometer
design of UAV-magnetometer system has been system has been
verified by verified
multiplebyflight
multiple
testsflight
and tests and be
it could
it could
applied be applied
to the remote to the remote
detection detection of near-surface
of near-surface targets. targets.

5.1. Analysis on the Processing Flow

The magnetic data is collected by the designed UAV-magnetometer system, the background
magnetic field is subtracted from the raw data first. Then, the UAV interference field in the
same frequency as the magnetic anomaly field is removed by the proposed compensation method.
Compared with the existed scheme that a long hard rod or soft cord (3–5 m) is attached to the drones,
this paper presents a more elegant and efficient operation from the perspective of signal processing,
filling the gap in the UAV magnetic compensation. The experimental results (Figure 11) show that
Remote Sens. 2020, 12, 452 16 of 19

the RMS of interference field is reduced to less than 1 nT after the compensation and the quality of
UAV magnetic data is significantly enhanced, which establishes the foundation for the subsequent
interpretation of magnetic data. The residual noise is most likely to related to the geological magnetic
background of the test site where the magnetic background near the surface is not uniform.
The magnetic data is interpreted by the YOLOv3-based Euler deconvolution method, the location of
targets is estimated automatically, no need for an experienced data interpreter to manually extract data.
The experimental results indicate that the YOLOv3-based Euler approach not only can suppress the
false targets, but also has higher positioning accuracy than the SW-based Euler way. The interpretation
results of SW Euler deconvolution have 15 fake targets, because the results of SW Euler method
depend only on the attenuation rate of magnetic field, some small interference fields that have similar
attenuation characteristics as the true anomalous field may lead to the false targets, which cannot
be eliminated by the threshold of N. By contrast, the windows picked up by YOLOv3 is selective,
which is related to the strength of the magnetic field and the Euler deconvolution depends on the
decay of the magnetic field within the selected windows. The improved Euler algorithm combines the
YOLOv3 and Euler, so the interpretation results depend on both the amplitude and attenuation of the
magnetic anomalous field, which is able to eliminate the false anomalies. On the other hand, the size
of window selected by YOLOv3 has good adaptability to different objects. The size of the window is a
key parameter in the interpretation process, an appropriate window not only contains abundant and
effective magnetic anomaly data from one target, but also avoids too much noise data and interference
field from adjacent magnetic sources. The setting of threshold in the proposed strategies makes the
size of the window adaptive to targets of different depth and size, and enables the magnetic data in the
window to have higher SNR, improving the location accuracy.
Meanwhile, the workflow of improved Euler deconvolution is faster and simpler. In the field
experiment, the Euler deconvolution was performed 7 times for the YOLOv3-based Euler algorithm,
and the Euler deconvolution was performed 5695 times for the SW-based Euler approach. If the
UAV-borne magnetic survey is carried out in a larger test area, the proposed Euler deconvolution
method will become more efficient because the sliding windows should cover the entire survey region.
Hence, the improved Euler processing based on the YOLOv3 is more satisfying for the UAV-borne
magnetic survey. Besides, this processing routine can be also extended to the terrestrial magnetic
survey and aeromagnetic survey.

5.2. Limits of the Research and the Future Work

The magnetic data processing workflow described in this paper is designed for the remote
sensing of near-surface compact targets, such as unexploded ordnance, metallic targets, and steel pipe.
These objects can be regarded as the 3D point-like magnetic dipoles; therefore, the Euler solution
indicates the position of the target, which is determined by one anomaly window solely. For the
interpretation of magnetic anomaly field generated from 2D or 1D geological sources, the distribution
of multiple Euler solutions can reflect the trend and direction of underground anomalies, so the
processing procedure herein is inappropriate. In addition, the Euler deconvolution assumes that an
anomaly field within a window is generated by an isolated target. If a large target and a small target
are close, the anomalous fields might be completely overlapping. The results of Euler deconvolution
are unreliable in this case. Some effective methods like measuring the magnetic gradient field with
higher spatial resolution to separate the magnetic anomaly field of two closed targets, should be added
to the processing pipeline.
The magnetic survey is performed in a uniform and stable magnetic environment; the background
field is removed by a lowpass filter and detrending. But if the targets to be detected are emplaced in a
strong geological noise like magnetic ore, deposits, or soils [39], separating the magnetic field generated
by the interested sources and geological magnetic background remains an ongoing challenge, because
the magnetic properties of the topographic features in the test areas should be investigated in advance.
Remote Sens. 2020, 12, 452 17 of 19

In some cases, the image filtering can be employed to extract the interested signals based on spatial
wavelength characteristics of signals [40].
In this paper, the interpretation result is the location of subsurface targets, the characteristics of the
target itself cannot be obtained. The future work is to identify the magnetic moment, size, and shape
of the objects, achieving the classification of targets, like the discrimination between UXO and clutter.

6. Conclusions
This paper presents the UAV-borne magnetic survey aimed at the remote detection of near-surface
targets, including the system design and magnetic field data processing and interpretation methods,
and a case study is given.
The system design section describes the main components of the UAV-magnetometer system.
The processing pipeline consists of three phases, the first phase ensures that the magnetic field data
over the test area is acquired successfully by the UAV-magnetometer system, the second phase is
to improve the quality of UAV magnetic data through a series of data processing techniques, the
third phase realizes the interpretation of the UAV magnetic data and provides the position of the
anomalies automatically.
One of the main contributions is a compensation method for the UAV interference field, overcoming
the problem of low quality of UAV magnetic data. Moreover, a novel interpretation method of UAV
magnetic data is proposed for the detection of metallic targets. Traditional Euler interpretation method
is based on the sliding windows that are not selective and time-consuming; the improved approach
employs the state-of-the-art YOLOv3 replacing the sliding windows, showing better performance in
shortening the calculation time, reducing the false targets, and improving the positioning accuracy.
The designed workflow in this paper will broaden the application of UAV magnetic sensing technique,
especially for the remote detection of ferromagnetic targets.

Author Contributions: Y.M. proposed the methodology; Y.M., X.Z., and Y.Z. designed the validation scheme;
Y.M., W.X., and Y.Z. performed the field tests; Y.M. and W.X analyzed the data; Y.M. wrote the paper. All authors
have read and agreed to the published version of the manuscript.
Funding: The research was funded by the research project “National Natural Science Foundation of China”,
which is supported by the Chinese government, grant number is 41604155.
Acknowledgments: The authors would like to thank Zhaonan Wang for his supports in computing resources.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Nabighian, M.N.; Grauch, V.J.S.; Hansen, R.O.; LaFehr, T.R.; Li, Y.; Peirce, J.W.; Phillips, J.D.; Ruder, M.E.
The historical development of the magnetic method in exploration. Geophysics 2005, 70, 33–61. [CrossRef]
2. Caetano, D.M.; Rabuske, T.; Fernandes, J.; Pelkner, M.; Fermon, C.; Cardoso, S.; Ribes, B.; Franco, F.; Paul, J.;
Piedade, M.; et al. High-resolution nondestructive test probes based on magnetoresistive sensors. IEEE
Trans. Ind. Electron. 2019, 66, 7326–7377. [CrossRef]
3. Hansen, R.O.; Racic, L.; Grauch, V.J.S. Magnetic methods in near-surface geophysics. In Near-surface
Geophysics, 1st ed.; Dwain, K.B., Ed.; Society of Exploration Geophysicists: Tulsa, OK, USA, 2005; pp. 151–175.
4. Meyer, U.; Siemon, B.; Becken, M.; Stolz, R.; Noell, U.; Frei, M.; Buchholz, P.; Steuer, A.; Costabel, S.; Martin, T.;
et al. New airborne methods and procedures for the exploration of mineral resources—An overview of BGR
activities. In Near Surface Geoscience 2016–22nd European Meeting of Environmental and Engineering Geophysics,
04 Sep 2016; Curran Associates: Turin, Italy, 2016.
5. Munschy, M.; Boulanger, D.; Ulrich, P.; Bouiflane, M. Magnetic mapping for the detection and characterization
of UXO: Use of multi-sensor fluxgate 3-axis magnetometers and methods of interpretation. J. Appl. Geophys.
2007, 61, 168–183. [CrossRef]
6. Billings, S.D. Discrimination and classification of buried unexploded ordnance using magnetometry. IEEE
Trans. Geosci. Remote Sensing 2004, 42, 1241–1251. [CrossRef]
Remote Sens. 2020, 12, 452 18 of 19

7. Fassbinder, J.W.E. Magnetometry for archaeology. In Encyclopedia of Geoarchaeology, 1st ed; Gilbert, A.S.,
Goldberg, P., Eds.; Springer: Dordrecht, The Netherlands, 2017; pp. 499–514.
8. Ege, Y.; Nazlibilek, S.; Kakilli, A.; Citak, H.; Kalender, O.; Erturk, K.L.; Sengul, G.; Karacor, D. A magnetic
measurement system and identification method for buried magnetic materials within wet and dry soils.
IEEE Trans. Geosci. Remote Sens. 2016, 54, 1803–1811. [CrossRef]
9. Faley, M.I.; Dammers, J.; Maslennikov, Y.V.; Schneiderman, J.F.; Winkler, D.; Koshelets, V.P.; Shah, N.J.;
Dunin-Borkowski, R.E. High-Tc SQUID biomagnetometers. Supercond. Sci. Technol. 2017, 20, 1–21.
10. Oyama, D.; Kobayashi, K.; Yoshizawa, M.; Uchikawa, Y. Mobile high-Tc SQUID system for MCG measurement.
IEEE Trans. Appl. Supercond. 2009, 19, 871–873. [CrossRef]
11. Gavazzi, B.; Maire, P.L.; Munschy, M.; Dechamp, A. Fluxgate vector magnetometers: A multisensor device
for ground, UAV, and airborne magnetic surveys. Lead. Edge 2016, 35, 795–797. [CrossRef]
12. Pajares, G. Overview and current status of remote sensing applications based on unmanned aerial vehicles
(UAVs). Photogramm. Eng. Remote Sens. 2015, 81, 281–329. [CrossRef]
13. Watts, A.C.; Ambrosia, V.G.; Hinkley, E.A. Unmanned aircraft systems in remote sensing and scientific
research: Classification and considerations of use. Remote Sens. 2012, 4, 1671–1692. [CrossRef]
14. Robert, J.; Yuleika, M.; Robert, Z.; Markku, P.; Ari, S.; Björn, H.H.; Heikki, S.; Jukka-Pekka, K.; Louis, A.;
Richard, G. Drone-borne hyperspectral and magnetic data integration: Otanmäki Fe-Ti-V deposit in Finland.
Remote Sens. 2019, 11, 2084–2111.
15. Douglas, G.M.; Héctor, I.A.; Perez, I.; Paulo, A.F.R.; Guilherme, A.P.; Luiz, C.C.B.; André, W.; Gustavo, M.F.;
Luis, G.U.G.; Mario, F.M.C. Autonomous aeromagnetic surveys using a fluxgate magnetometer. Sensors
2016, 16, 2169.
16. Malehmir, A.; Dynesius, L.; Paulusson, K.; Paulusson, A.; Johansson, H.; Bastani, M.; Wedmark, M.;
Marsden, P. The potential of rotary-wing UAV-based magnetic surveys for mineral exploration: A case study
from central Sweden. Lead. Edge 2017, 36, 552–557. [CrossRef]
17. Eck, C.; Imbach, B. Aerial magnetic sensing with an UAV helicopter. In Proceedings of the International
Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Zurich, Switzerland,
14–16 September 2011.
18. Wood, A.; Cook, I.; Doyle, B.; Cunningham, M.; Samson, C. Experimental aeromagnetic survey using an
unmanned air system. Lead. Edge 2016, 35, 270–273. [CrossRef]
19. Walter, C.A.; Braun, A.; Fotopoulos, G. Impact of 3-D Attitude Variations of a UAV Magnetometry System on
Magnetic Data Quality. Geophys. Prospect. 2018, 67, 465–479. [CrossRef]
20. Tolles, W.E.; Mineola, N.Y. Compensation of Aircraft Magnetic Fields. U.S. Patent US2692970A, 26 October
1954.
21. Noriega, G. Aeromagnetic compensation in gradiometry—performance, model stability, and robustness.
IEEE Geosci. Remote Sens. Lett. 2015, 12, 117–121. [CrossRef]
22. Oruç, B. Location and depth estimation of point-dipole and line of dipoles using analytic signals of the
magnetic gradient tensor and magnitude of vector components. J. Appl. Geophys. 2010, 70, 27–37. [CrossRef]
23. Billings, S.; Wright, D. Interpretation of high-resolution low-altitude helicopter magnetometer surveys over
sites contaminated with unexploded ordnance. J. Appl. Geophys. 2010, 72, 225–231. [CrossRef]
24. Reid, A.B.; Allsop, J.M.; Granser, H.; Millett, A.J.; Somerton, I.W. Magnetic interpretation in three dimensions
using Euler deconvolution. Geophysics 1990, 55, 80–91. [CrossRef]
25. Davis, K.; Li, Y.; Nabighian, M. Automatic detection of UXO magnetic anomalies using extended Euler
deconvolution. Geophysics 2010, 75, 13–20. [CrossRef]
26. Davis, K.; Li, Y.; Nabighian, M. Automatic detection of UXO magnetic anomalies using extended Euler
deconvolution. In 2005 SEG Annual Meeting, Houston, TX, USA, 6-11 November 2005; Society of Exploration
Geophysicists: Tulsa, OK, USA, 2005.
27. Gerovska, D.; Araúzo-Bravo, M.J. Automatic interpretation of magnetic data based on Euler deconvolution
with unprescribed structural index. Comput. Geosci. 2003, 29, 949–960. [CrossRef]
28. Usman, N.; Abdullah, K.; Nawawi, M.; Khalil, A.E. New approach of solving Euler deconvolution relation
for the automatic interpretation of magnetic data. Terr. Atmos. Ocean. Sci. 2018, 29, 243–259. [CrossRef]
29. Sheinker, A.; Moldwin, M.B. Adaptive interference cancelation using a pair of magnetometers. IEEE Trans.
Aerosp. Electron. Syst. 2016, 52, 307–318. [CrossRef]
Remote Sens. 2020, 12, 452 19 of 19

30. Mu, Y.X.; Zhang, X.J.; Chen, L.Z.; Zheng, Y.X.; Xie, W.P. Compensation method for multi-rotor
UAV-magnetometer system, IEEE International Geoscience and Remote Sensing Symposium (IGARSS
2019). in press.
31. Redmon, J.; Farhadi, A. YOLOv3: An Incremental Improvement. arXiv 2018, arXiv:1804.02767.
32. Daya, Z.A.; Birsan, M.; Holtham, P.M.; Holtham, P.M. The magnitude and distribution of ambient fluctuations
in the scalar magnetic field. In OCEANS 2005 MTS, 17–23 Sept. 2005; IEEE: Washington, DC, USA, 2005.
33. Topics in Gridding. Available online: https://www.geosoft.com/media/uploads/resources/technical-papers/
topicsingriddingworkshop.pdf (accessed on 16 June 2019).
34. Wang, X.H.; Zhang, C.C.; Liu, F.Y.; Dong, Y.N.; Xu, X.L. Exponentially weighted particle filter for simultaneous
localization and mapping based on magnetic field measurements. IEEE Trans. Instrum. Meas. 2017, 66,
1658–1667. [CrossRef]
35. Pilkington, M.; Gregotski, M.E.; Todoeschuck, J.P. Using fractal crustal magnetization models in magnetic
interpretation. Geophys. Prospect. 1994, 42, 677–692. [CrossRef]
36. Feng, Y.; Jiang, Y.; Jiang, Y.; Li, Z.; Jiang, J.; Liu, Z.W.; Ye, M.C.; Wang, H.S.; Li, X.M. Regional magnetic
anomaly fields: 3D Taylor polynomial and surface spline models. Appl. Geophys. 2016, 13, 59–68. [CrossRef]
37. Briggs, I.C. Machine contouring using minimum curvature. Geophysics 1974, 39, 39–48. [CrossRef]
38. Fan, Z.F.; Huang, L.; Zhang, X.J.; Fang, G.Y. An elaborately designed virtual frame to level aeromagnetic
data. IEEE Geosci. Remote Sens. Lett. 2016, 13, 1153–1157. [CrossRef]
39. Li, Y.G.; Krahenbuhl, R.; Meglich, T.; Oldenburg, D.; Pasion, L.; Billings, S.; Dam, R.V.; Harrison, B. Improving
UXO Detection and Discrimination in Magnetic Environments; SERDP Project MM-1414; Colorado School of
Mines Golden Dept of Geophysics: Boulder, CO, USA, May 2010.
40. Maysam, A.; Kiomars, M.; Hamid, D.; Ahmad, M.Z. Geological noise removal in geophysical magnetic
survey to detect unexploded ordnance based on image filtering. Iran. J. Geophys. 2016, 9, 11–23.

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).