Robust data processing of noisy marine controlled-sourceelectromagnetic data using independent component analysis

Naoto Imamura1,5 Tada-nori Goto2,4 Takafumi Kasaya3 Hideaki Machiyama4

1College of Earth, Ocean, and Atmospheric Sciences, Oregon State University, 104 CEOAS Administration Building,Corvallis, OR 97331-5503, USA.

2Graduate School of Engineering, Kyoto University, C1-2-216, Kyotodaigaku Katsura, Nishikyo-ku, Kyoto-shi 615-8540,Japan.

3Research and Development Center for Earthquake and Tsunami, Japan Agency for Marine-Earth Science and Technology(JAMSTEC), 2-15 Natsushima, Yokosuka, Kanagawa 237-0061, Japan.

4Research and Development Center for Submarine Resources, JAMSTEC, 2-15 Natsushima, Yokosuka,Kanagawa 237-0061, Japan.

5Corresponding author. Email: nimamura@ceoas.oregonstate.edu

Abstract. Data processing techniques are often used to estimate the noise-free response of marine controlled-sourceelectromagnetic (CSEM) data and magnetotelluric transfer functions. We have implemented a new CSEM data processingscheme that uses a robustmethod based on independent component analysis (ICA) to extract interpretable datasets fromnoisymarine CSEM data. We applied the data processing scheme to signals from a new CSEM observation system comprisinga remotely operated vehicle (ROV) and anoceanbottomelectromagnetometer (OBEM).These datasetswere obtained aroundthe IheyaNorth hydrothermalfield,OkinawaTrough, Japan. The observation systemallows a small-scaleCSEMsurvey to beconducted in areas of steep topography, such as hydrothermal fields, because the ROV can deploy the OBEM at the exactobservation site. The results show that the coherent and environment noise that exists in the raw time series is reducedsufficiently by ICA processing. It makes interpretation of the resulting electric field data possible. The results also show thatthe processed data has a higher signal-to-noise ratio in the middle-to-high-frequency band than the data without ICA. Thenormalised spectrum, obtained by normalising the observed data from the hydrothermal area, indicates that a conductiveanomaly exists in the near-offset area around the OBEM.We apply 2D inversion to the electric field data and find that a lowresistivity body exists beneath theOBEMand 50moffset from theOBEM.This resistivity structure is consistent with imagestaken by the ROV that show characteristic organisms in hydrothermal seepage around the OBEM site.

Key words: controlled-source electromagnetic method, hydrothermal vent, independent component analysis, OkinawaTrough, robust data processing.

Received 23 October 2017, accepted 25 October 2017, published online 27 November 2017

Introduction

The controlled-source electromagnetic (CSEM) method hasbeen used in the last decade as a tool for exploring marineenvironments (Constable, 2010; Streich, 2016). For example,this method has been applied to surveys for hydrocarbons(Ellingsrud et al., 2002; Constable and Srnka, 2007; Streich,2016), gas hydrates (Schwalenberg et al., 2010; Weitemeyeret al., 2011) and a fault system in bending oceanic crustbefore it sinks into the subduction (Naif et al., 2015). Mostdata analyses of the CSEM method use forward or inversionschemes to estimate the resistivity structure beneath the earth.These schemes are usually used to fit calculated electromagneticdata to observed electromagnetic data. However, the observedelectromagnetic field data are generally noisy and/or biased. Ingeneral (on land and seafloor), noise arises from the equipment(Constable, 2013), power grids (Nockles et al., 2009), sporadicnoise (Strack et al., 1989), ocean currents (Kasaya and Goto,2009), airwaves (Løseth et al., 2010;ChenandAlumbaugh, 2011;Wirianto et al., 2011) and so on. Data processing techniques areused to decrease the noise components and to extract interpretable

data from the observed electromagnetic fields and areindispensable in estimating the noise-free response of marineCSEM data. As a result, noise reduction can improve estimationsof resistivity structures beneath the surface.

Careful treatment of noise is essential for small-scale CSEMsurveys in a hydrothermal area. Exploration around hydrothermalfields is becoming important, as they provide mineral resourcesof high quality for engineering applications and hydrothermalvent organisms for scientific research. From previous research,hydrothermal vents and mounds are considered to be electricallyconductive, as they are a potential resource for metals suchas Cu, Zn, Ag and Au (Bartetzko et al., 2006; Spagnoli et al.,2016), although the resistivity structure beneath the seafloorstill remains. However, exploration in such areas is difficultfor several reasons. One reason is the topography aroundhydrothermal vents. A hydrothermal vent can create a tallchimney with height of up to 20m (Delaney et al., 1992).Figure 1 shows the seafloor topography in the Iheya Northhydrothermal field in the Okinawa Trough, Japan. The figurealso shows the tall chimneys and knolls created by hydrothermal

Originally submitted to SEGJ 1 October 2016, accepted 22 September 2017

CSIRO PUBLISHING

Exploration Geophysicshttps://doi.org/10.1071/EG17139

Journal compilation � ASEG 2017 www.publish.csiro.au/journals/eg

vents, which form amajor part of the topography in this area. Thistopography makes traditional CSEM surveying, conductedby towing cables, difficult because the cables must be towedfar above the chimneys, which reduces the amplitude of theelectromagnetic response from the subsurface. Another reasonis the limited spatial extent of the hydrothermal fields. Althoughtraditional CSEMmethods are used for conducting hydrocarbonsurveys covering several kilometres in extent, hydrothermal ventsare generally densely distributed over an area of several hundred

metres. Exploration of small spatial extent requires stable andaccurate measurement of the positions of transmitters andreceivers, which is difficult even with modern survey systems.To solve the above-mentioned problems, we have developed anew CSEM survey system based on a remotely operated vehicle(ROV) and an ocean bottom electromagnetometer (OBEM), asshown in Figure 2. In this system, the ROV brings the OBEMexactly to defined locations in the target area, irrespective of itstopography, and the offset between the ROV and the OBEM can

126.895° 126.8975°

–930

27.79°

27.7925°

0 100 200 m

Altitude (m)

Iheya North

R4

R5R6

R12

R7

R14

T1

T10

T11

T16

T20

T21

Transmitter

Hydrothermalvents

Receiver

T2T3T4

T5

–960

–990

–1020

–1050

–1080

Fig. 1. Geometry of the CSEM survey showing the topography of the seafloor with contours and colour.Yellow circles depict ROV transmitters. Red triangles depict OBEM receivers. Identification numbers areshown for several transmitters and receivers. Yellow arrows depict hydrothermal fluid venting sites shownin Kawagucci et al. (2013).

Remotely operated vehicle

TX control unit

AC/DC transformer

Transmitter withrecorder (TX)

Recorder & battery(+ magnetometer, tiltmeter)

Current electrodes

Receiverelectrodes

Ocean bottom electromagnetometer(OBEM)

Acoustic responder

Ch1

Ch2Ch3

Ch4

Plan view

Side view

Seafloor

Cable to vessel

Basket30 cm

Receiverelectrodes

Receiverdipole(1 m)

Common electrode

Acoustic responderRemotely operated vehicle Basket

Transmitterdipole(2 m)

ROVcomm. port(RS232C)

ROVpower outlet

100 V AC

50 V DC

X

Y

Fig. 2. Simplified sketch of the ROV-OBEMsystem in side and plan views. The ROV transmits electric currentafter placing the OBEM at the target location. The ROV transmits electric current every 30 s once it has settled onthe seafloor.

B Exploration Geophysics N. Imamura et al.

be determined with high accuracy. However, the ROV canhandle only small amplitudes of electric current because ofits limited output power. Also, the maximum length of theelectric dipoles of the ROV and the OBEM should be less thana few metres to ensure portability by the ROV. As such, noisemust be decreased by data processing to increase the signal-to-noise ratio.

In marine CSEM data processing, several techniques areused to recover a noise-free response function, which can be afunction in either time or frequency domains. PreviousCSEM research involves significant efforts to reduce noise tosufficiently low levels (Myer et al., 2011). CSEMdata processingtypically comprises general digital signal processing steps.However, other processes have been used for different targets,datasets and external field environments. Myer et al. (2012)pre-whitened and post-darkened observed data to suppressspectral leakage caused by the time variations in the Earth’selectromagnetic field. Streich et al. (2013) applied robustweighted least square fitting to CSEM data that was partof a magnetotelluric (MT) survey (Egbert and Booker, 1986;Chave and Thomson, 1989), while other research aims todecrease the airwave effect by data processing (Chen andAlumbaugh, 2011).

Recently, independent component analysis (ICA) has beenproposed for addressing signal separation problems (Hyvärinenand Oja, 2000; Cichocki and Amari, 2002). This method is usedto separate recorded signals that are weighted sums of sourcesignals into those original source signals. In ICA, informationabout original source signals and theirmix is unknown. ICAfindsa linear representation of the recorded signals so that theseparated source signal estimates are statistically independent,or are as independent as possible. Because we do not know themixingweights, ICA is called a ‘blind source separation’method.An interesting property of ICA is its robustness against outliersin data (Hampel et al., 1986), which is important for robustprocessing of electromagnetic survey data. ICA has beenapplied to sound, image and biomedical signal processing andhas also been used in geophysical data processing (Aires et al.,2002; Furukawa et al., 2006; Ciaramella et al., 2004; Tsuno andIwata, 2015). Murakami and Yamaguchi (2007) used ICA foranalysing geoelectric data that were subjected to a very high levelof noise during a water injection experiment at the Nojima Faultin Japan. Sato et al. (2017) also used ICA for the processing ofa self-potential survey dataset to decrease its noise level. Bothgeoelectric studies show that the noise level could be reducedsufficiently, but as yet there are few applications of ICA to thegeoelectromagnetic problem.

In this paper, we first present a new CSEM observationsystem used in the Iheya North hydrothermal field in theOkinawa Trough. We describe this observation system, whichuses a ROV and an OBEM for a small-scale CSEM survey. Next,we explain ICA and the robust data processing algorithm that isapplied to the data collected around hydrothermal areas. Theelectric potential data observed with the OBEM are shown tocontain both impulsive noise and high-frequency noise. Our aimis to extract interpretable CSEM signals from the noisy data,focusing on one OBEM dataset. We show the processing resultsof time series and spectral data without and with ICA. Theseprocessed data are normalised using a dataset obtained awayfrom the hydrothermal area to obtain the normalised electric fieldresponses. We then applied a 2D inversion scheme to the electricfield data to estimate the resistivity structure beneath the seafloor.Finally, we interpret the resistivity structure of the hydrothermalfield and discuss the effectiveness of data processing for thereduction of noise.

Remotely operated vehicle system

As we mentioned above, we have developed a CSEM surveysystem using a ROV and an OBEM as shown in Figure 2. TheROV is equipped with an electric transmitter dipole with a lengthof ~2m. The amplitude of current is limited to a maximum of50 A. Trading off desired source power and practicability, wetypically transmitted an electric current of rectangular waveformevery 30 s, only once the ROVwas stationary on the seafloor.Weshortened each of the four horizontal receiver arms of the OBEMto 0.3m to be carried by the ROV. The OBEM then carried twoorthogonal dipoles of 1m length. This arrangement allowed usto set the OBEM exactly in place inside the target area. Afterinstalling theOBEM inside the target area (e.g. at R4 in Figure 1),the ROV ran a survey line transmitting an electric current at eachstation (T1 to T5). The distance between the transmitter (ROV)and the receiver (OBEM) was accurately measured with anacoustic responder. Video and digital still cameras mounted onthe ROV captured graphical information while the ROV movedin the area. GPS signals were used for time synchronisationbetween the transmitter and the OBEM before deployment andfor measuring the drift of the OBEM and ROV clocks after theirrecovery on board.

Although there are four channels of electric potential, relativeto a central common electrode attached to the OBEM, one ofthe channels did not record properly. We therefore used threechannels to calculate the horizontal electric field components.As these channels contain high-amplitude periodic noise andhigh-frequency noise, observed data at far offsets has a verylow signal-to-noise ratio (e.g. Figure 3). This periodic noisewas thought to arise when observed data were written to therecording media in the OBEM. As each noise impulse has anindependent duration in time, removing this noise by applyinga moving average filter or a median filter is difficult, andestimation of the noise waveform by stacking the noise pulseswas unsuccessful. The observed time series is also contaminatedby high-frequency noise, which will be described later. We applyrobust processing to the time series to remove the impulsive noiseand the high-frequency noise.

Method

Weused ICAandsomebasic signal processing to reduce thenoisein the observed time series. ICA recovers signals only from themeasured data. The ICA method is based on a form of signalanalysis that uses statistical assumptions to enable blind sourceseparation. In the ICAprocess, observed data are considered to becomposed of linearly superposed mixtures of original signals:

xm�1 ¼ Am�nsn�1; ð1Þwhere x represents the components of observed data, s representsthe original signals and A is a non-singular mixing matrix. Theapplication of ICA is for the case where s and A are unknown.Additionally, we define y and B as follows:

yn�1 ¼ Bn�mxm�1; ð2Þwhere amatrixB can be built by splitting up themeasured signalsso that the resulting output vector y is the optimal approximationof vector s. If B corresponds to the inverse matrix of A, then y isthe same as s. The goal of ICA is to obtain amatrixB in such awaythat vector y is close to the original, independent source signals.An important assumption in ICA is that the components of s arestatistically independent.

One of the important aspects of ICA is that it is astatistical model. This means that the distribution of a sum ofindependent random variables becomes a Gaussian distribution

Robust processing of MCSEM data using ICA Exploration Geophysics C

from the central limit theorem. The main idea of ICA isto maximise non-Gaussianity to extract the independentcomponents of signals. In ICA, optimisation is used tomaximise non-Gaussianity. A popular optimisation method isFastICA (Hyvärinen and Oja, 2000), which we use here forprocessing data. FastICA finds a direction w, which is one ofthe rows of the matrix B, so that the projection wTx maximisesthe independence of the single estimated source y. Independenceis measured by the approximation of negentropy, which is ameasure of non-Gaussianity. We maximise function JG to findone independent component:

JGðwÞ ¼hEfGðwTxÞg � EfGðnÞg

i2; ð3Þ

wherew is aweight vector constrained so thatE{(wTx)2} = 1,E{}indicates the expectation value, G is a suitable approximatingcontrast function and n is a standardised Gaussian randomvariable. This maximisation requires a whitening of theobserved variables. This means that before the application ofthe ICA algorithm, we transform the observed vector x linearlyso that we obtain a new vector ~x that is white, i.e. its componentsare uncorrelated and their variances equal unity with zero mean.To apply the procedure described above, ICA requires thefollowing assumptions for observed data. First, all the sourcesmust be statistically independent of each other. Theseindependent components must be non-Gaussian. Second, thenumber of observed signals must be at least as large asthe number of independent components. Finally, matrix A hasto be of full column rank.

Application to field data

We apply robust data processing to the field data acquired fromthe IheyaNorth hydrothermalfield in theOkinawaTrough duringthe research cruise ‘NT13-22’ (conducted in November 2013)of R/V Natsushima, a JAMSTEC (Japan Agency for Marine-Earth Science and Technology) research vessel. These data wereobtained by using the ROV-OBEM survey system loaded onROV Hyper-Dolphin. The recorded electric potential in threechannels is shown in Figure 3 with a sampling rate of 1 kHz. Asone can see in Figure 3, a periodic impulsive noise contaminatesthe observed data. The impulses occur approximately every0.29 s. This delta-like function has a broad spectrum in thefrequency domain and causes noise in the whole responsespectrum.

Figure 4 shows aflowchart of our processing scheme. First,wedivided the whole time series into segments that each includes asignal generated by the ROV. These time series are decomposedinto several independent signals using the ICA algorithm. Inthis example, we used either two or three independent signals,chosen by a trial-and-error method. Next, we removed the noise

Divide time series into segments

Decompose time series of channel data

Remove noise components in decomposed signals

Reconstruct time series using signal components

Take a median value of each time step

Calculate electric field from channels

Apply FFT

Iterate 10 times?No

Yes

Post-processing

Fig. 4. Summary of the statistical data processing scheme. The loopimplements the statistical processing algorithm for a given time series ofthree channels. It reconstructs recovered signals after removing the noisecomponents.

(a)

2.262.4

2.2

1.6

1.5

2.40

2.35

2.30

2.25

03:14:20 03:14:25 03:14:30 03:14:35

UTC time03:14:40 03:14:45 03:24:20 03:24:25 03:24:30 03:24:35 03:24:40 03:24:45

Ch1

(m

V)

Ch2

(m

V)

Ch3

(m

V)

2.24

2.22

1.625

1.600

1.575

2.300

2.275

2.250

(b)

Fig. 3. Data examples for receiver R4 and transmitters (a) T1 and (b) T5with a transmitter–receiver distance of 50 and 10m, respectively. As one channel in theOBEMdidnot record the timeseries of the electricfieldwell, only three channelsof data are shown.The source signal is a rectangularwaveformwith aperiodof8 s.

D Exploration Geophysics N. Imamura et al.

components in the decomposed signals. A decomposed signalwas determined to be a noise component if its correlationcoefficient with the source signals was not the highest value inthe correlation coefficients between the decomposed signals andthe source signals. After removing the noise components, theestimated electric potential is recovered using a mixing matrixbased on Equation 1.

In the FastICA algorithm, the initial value of w is a randomvalue on the unit circle in the coordinate space of whitenedobserved signals. Convergence of the ICA algorithm dependson the starting value ofw. As the ICA algorithm can fail its signaldecomposition depending on the initial choice of w, we apply it10 times and take the median value for each time step to avoidfailure of the ICA algorithm. Figure 5 shows an example of thedecomposed signals for the transmitter–receiver pair T5 and R4(Figure 1). From this figure, it is obvious that the ICA algorithmcan decompose original time series into rectangular signalwaveforms and impulsive noise components, although therestill exists some small impulsive noise in the rectangular signalwaveforms. One can see that noise components include bothperiodic and fluctuating noise in the baseline. However, theamplitude of fluctuating noise is smaller than that of periodicnoise. Themixing values of the signalwaveform for channels 1, 2and 3 are 0.0096, –0.0036 and 0.0028, respectively. Thesechannel numbers can be seen in Figure 2. The values for theimpulsive noise components are –0.0005, –0.0006 and –0.0006for channels 1–3, respectively. This shows that the mixingvalues of the impulsive noise have a similar value for each

channel, which in-turn indicates that the impulsive noisecomes from an instrumental source and not from the sourcesignal. If this noise were caused by the source current, themixing value would be different for each channel because theoffset is different. This noise reduction process allows us toextract the instrument noise and evaluate its amplitude. Noise-free electric potential channel data are obtained after removing thenoise components.Whenwe decomposed the observed signals tothree independent signals for the pair T5 and R4, the originalrectangular signal waveform was decomposed into two differentwaveforms contaminated with the noise components. Therefore,we chose to decompose the observed signals to the two signalsshown in Figure 5. For the other pairs from T1 to T4 and R4, wedecomposed the observed signals to three independent signals.Twocomponents of horizontal electricfieldswere then calculatedfrom the recovered three-channel data.

Figures 6 and 7 show the electric fields without and with dataprocessing, in the time and frequency domains. These electricfields are the components parallel to the transmitter dipole of theROV. These observed electric fields and transmitter currents areoriented close to the east–west direction, which is the broadsidedipole configuration in the marine CSEM survey. Figure 7ashows that the recovered time series with ICA processingcould remove impulsive noise compared with the original timeseries for the pair T1 and R4. One can see that the amplitude ofhigh-frequency fluctuations in Figure 6a is close to the amplitudeof the step in the rectangular wave, although the high-frequencyamplitude in Figure 7a is smaller than the amplitude of step in the

0.050

0.025

0.000

–0.025

–0.050

03:24:19 03:24:24 03:24:29 03:24:34

UTC time

Sig

nal (

mV

)

03:24:39 03:24:44 03:24:49

Fig. 5. Decomposed signals for transmitter T5 and receiver R4. The blue line depicts the noisecomponent of the time series. The green line depicts the signal component of the time series.

(a)

0

03:14:20 03:14:25 03:14:30 03:14:35 03:14:40 03:14:45 03:24:20 03:24:25 03:24:30 03:24:35

Frequency (Hz)

03:24:40 03:24:45

0.2

0

–0.2

10–2

10–3

10–4

10–5

10–6

10–1 100 101 102 10–1

10–1

10–3

10–5

180

0

–180

100 101 102

10–7

180

0

–180

Ex

(mV

/m)

Am

plitu

de (

V/m

)P

hase

( )

–0.01

(b)

Fig. 6. Time series of electricfields, amplitude spectrumof electricfield in the frequencydomain andphase spectrum(in order from top to bottom). (a)Broadsidecomponent of electric field at receiver R4 from transmitter T1 and (b) broadside component of electric field at receiver R4 from transmitter T5.

Robust processing of MCSEM data using ICA Exploration Geophysics E

rectangular wave. As the original source signal does not havethesefluctuating signals, theyare considered tobehigh-frequencynoise caused during observation and are reduced successfully.Figure 7b shows that their time series do not change beforeand after data processing. This is because the mixed value ofnoise in this pair is small compared with the rectangular signalcomponents (Figure 5). This suggests that ICA processing coulddecompose the impulsive noise and high-frequency noise if itsamplitude is large enough compared with the amplitude of therectangular signal.

The spectra of the T1-R4 pair without and with ICA aredifferent across the whole frequency range (Figures 6a and 7a).It is obvious that the impulsive noise in the time series affectsa broad range of frequencies. It can be observed that the spectrumis close to flat from 1 to 100Hz in Figure 6a. This frequency bandcorresponds to the periodic impulsive noise in the time series. Onthe other hand, the spectrum with ICA is not flat in the samefrequency band since it does not have a high-amplitude impulse.

The spectrum of the T5-R4 pair without and with ICA does notchange much (Figures 6b and 7b) because each time series issimilar to the other.

We applied this ICAprocessing scheme to all datasets (T1, T2,T3, T4 and T5) for receiver R4. Figure 8 shows the amplitude andphase of the spectrum of the broadside electric field as a functionof frequency and distance between theROVand theOBEM.Notethat the spectrum of the electric field is Green’s function, which isnormalised by the transmitted source spectrum. From this figure,one can see that the spectrum is similar at low frequencieswithout(Figure 8a, b) and with (Figure 8c, d) ICA processing. However,the spectra at middle-to-high frequencies are largely differentfromFigure 8a, c. This is the same for the phase spectrum aswell.These trends are explained by the same reason given above.As the impulsive noise in the time series affects largely themiddle-to-high frequency range, the spectra in these regionsare contaminated. One can also see that the processedspectrum becomes smoother in space compared with the result

(a)

03:14:20 03:14:25 03:14:30 03:14:35 03:14:40 03:14:45 03:24:20 03:24:25 03:24:30 03:24:35

Frequency (Hz)

03:24:40 03:24:45

0.20.002

0

–0.002

0

–0.2

10–2

10–1 100 101 102 10–1 100 101 102

10–4

10–6

10–1

10–3

10–5

180

0

–180

180

0

–180

Ex

(mV

/m)

Am

plitu

de (

V/m

)

(b)

Pha

se (

)

Fig. 7. Time series of electricfields, amplitude spectrumof electricfield in the frequency domain and phase spectrum (in order from top to bottom) after the ICAalgorithm has been applied. (a) Broadside component of electric field at receiver R4 from transmitter T1 and (b) broadside component of electric field at receiverR4 from transmitter T5.

(a)

(b)

(c)

(d)

50

40

30

20

10

50–4

–5

–6

–7

180

0

–180

40

30

20

10

log 1

0 (⏐

Ex⏐

V/(

Am

2 ))

50

40

30

20

10

0.1 1 10 100 0.1 1 10 100

50

40

30

20

10

Dis

tanc

e (m

)

Frequency (Hz)

Pha

se (

)

Fig. 8. Responses at receiverR4 for all of the transmitters: (a, b)without ICAprocessing and (c, d) with ICAprocessing applied. Amplitude spectra are shownin (a, c) and phase spectra in (b, d).

F Exploration Geophysics N. Imamura et al.

without ICA processing (Figure 8). This smooth spectrum isconsidered to be reasonable because electromagnetic fields aregoverned by the diffusion equation, whichmakes the distributionof the spectrum smooth in space under the noise-free condition.

We normalised these spectrum amplitudes by the responsesobtained outside the hydrothermal area using the same surveysystem (the site location is 1.5 km east of the target area inFigure 1). The normalised spectrum is calculated by dividingthe amplitude of the spectrum in Figure 8 with the amplitude of thespectrum from the remote area. This provides uswith an apparentresistivity map because the normalised spectrum decreases thediffusive attenuation of the observed electromagnetic fields.

Figure 9a and b show the normalised spectrum without andwith ICA processing, respectively. From the results, we findthat the non-processed response has a larger normalised ratiothan the processed response at middle-to-high frequencies. Thismeans that the non-processed response indicates larger electricfields at these frequency bands. However, these apparentlylarger electric fields are artefacts, generated by the impulsiveand environmental noise. After data processing, we obtained alower normalised ratio, indicating that the resistivity in thisarea is low, which is reasonable for a hydrothermal area. Asthis normalised ratio represents apparent resistivity, a higherfrequency reflects the resistivity at shallow depth and a lowerfrequency reflects the resistivity structure at greater depth. FromFigure 9b, one can see that the resistivity is low in the deep area.

To confirm this interpretation, we applied 2D inversion to theobserved data using MARE2DEM (Key, 2016). Broadsideelectric field responses were computed at four frequencies:0.13, 0.50, 2.00 and 7.97Hz, which are similar to the periodof the transmitted rectangle waveform. Since these frequenciesare low considering the skin depth for the target area, theresolution of the structure is not high. We used the topographyaround the observation area, which has 2m resolution in thehorizontal direction, to define the model interface. The initialmodel consists of seawater with a resistivity of 0.3125 �m and ahalf-space with a resistivity of 10�m. Figure 10a and b show theinversion results without and with ICA processing, respectively.Both inversion results converged with a target root mean square(rms) misfit 1.0 after 18 and 11 iterations, respectively. Theinversion result without ICA processing depicts an extremelyhigh-resistivity body beneath the seafloor around T4 and T5(feature A). This resistive body continues to the location of theOBEM. The inversion result with ICA processing also shows ahigh-resistivity body in a similar place, but its size is smaller thanthe one without ICA processing and is limited to shallow depths.Both inversion results image low resistivity bodies under the

50

40

30

20

10

2.0

1.5

1.0

0.5

0

2.0

1.5

1.0

0.5

0

Nor

mal

ised

rat

io

50

40

30

20

10

Dis

tanc

e (m

)

0.1 1 10 100

0.1 1 10 100Frequency (Hz)

(a)

(b)

Fig. 9. Normalised responses at receiverR4 (a) without ICAprocessing and(b) with ICA processing applied.

Location (m)0 10 20 30 40 50

log 1

0 (Ω

m)

1030

1035

1040

1045 –1

0

1

2

3

4

Dep

th (

m)

4

–1

0

1

2

3

1030

1035

1040

1045

A

B

D

C

(a)

(b)

Fig. 10. Resistivitymodel of the inversion results (a) without ICAprocessing and (b) with ICA processing applied. Fivewhitedots depict the location of transmitters (T1 to T5) and a white triangle depicts the location of the receiver (R4). Labelled featuresinclude resistive shallow accumulations (A) and conductive bodies (B, C and D).

Robust processing of MCSEM data using ICA Exploration Geophysics G

resistive shallow sediment near the OBEM (features B and D).The size and the resistivity of these features are similar to eachother. It seems that this low resistive body continues up tothe seafloor. There is another low-resistivity body beneath T1and T2 in Figure 10b. Its resistivity is ~10 times lower thanthe same location in Figure 10a. Overall, it appears that theresistivities without ICA processing are larger than thosewith ICA processing. The reason for this can be explainedfrom Figure 8. This figure shows that the amplitude of thespectrum without ICA processing is larger than the one withICA processing. The extremely high-resistive zone (A inFigure 10a) is a result of the larger amplitude of the spectrumobtainedwithout ICAprocessing and is probably a false structure,perhaps due to spike noise in the observed time series.

Since the in situ resistivity is still unknown in this target region,as far as the authors know, further discussion of the distributionof conductive hydrothermal fluids or deposits should be basedon core samples. Here, we made a preliminary interpretationbased on photographs taken by the ROV. From the photos,hydrothermal fluid and gas seepage from the seafloor appearsbetween R4 and T5 (~10m wide). Characteristic bacteria andmussels, typical organisms found at hydrothermal seepages, arealso seen around the area. The inversion results from Figure 10show the low resistivity around the location of the seepage(features B and D). The photos also show that there arecharacteristic mussels in colonies on the left side, far from T1.This might support the conclusion that the inversion result withICA processing (feature C; Figure 10b) is more reasonable thanthe one without ICA processing (Figure 10a), which has a higherresistivity in the same area.

Conclusion

We have presented the ICA method for robust processing ofdata obtained around the Iheya North hydrothermal field,Okinawa Trough, Japan. The data was collected in 2013 usinga newly developed CSEM survey method based on a ROV-OBEM system, which is useful for both small-scale surveys andsurveys in areas of steep topography. In the electric potentialdata obtained, there is periodic and environmental noise,which makes the interpretation of observed data difficult. Afterrobust processing of the data, we found that impulsive andenvironmental noise components could be reduced effectively.As a result, the spectrum of electric fields with ICA looksreasonable, especially at middle-to-high frequencies, comparedwith the spectrum without ICA. The spectrum of electric fieldswas normalised using a spectrum from data acquired away fromthe hydrothermal areas. This normalised ratio shows thatthe apparent resistivity is low at shallow depths around theOBEM location. The inversion results also show that lowresistivity bodies exist beneath the OBEM and stations T1 andT2. The location of the resistivity anomaly is also consistent withimages taken by the ROV, which show seepage with typicalhydrothermal vent organisms in the same area. The applicationsof the robust processing method described here are not limited toCSEM data. Noise in time series data can be reduced in a similarmanner for different types of survey such as MT surveys, directcurrent surveys and transient electromagnetic surveys.

Conflicts of interest

The authors declare no conflicts of interest.

Acknowledgements

The authors thank the two anonymous reviewers for their fruitful suggestionsandcomments.Theyaregrateful for themanyefforts andkindnessofHiroyuki

Yamamoto, JAMSTEC, Chief Scientist on the NT13–22 cruise. The authorsalso thank theCaptain, crewand technicians ofR/VNatsushima, the operatingteam of ROV Hyper-Dolphin and supporting members for their cooperation.This project is supported by MEXT (Ministry of Education, Culture, Sports,Science and Technology of Japan), in a development programof fundamentaltools for exploration of deep seabed resources, and partly by JSPSKAKENHIGrant Numbers 26289347, 26249060, 26282101, 15H03715 and 15H05815.

References

Aires, F., Rossow, W. B., and Chédin, A., 2002, Rotation of EOFs by theindependent component analysis: toward solving the mixing problemin the decomposition of geophysical time series: Journal of theAtmospheric Sciences, 59, 111–123. doi:10.1175/1520-0469(2002)059<0111:ROEBTI>2.0.CO;2

Bartetzko, A., Klitzsch, N., Iturrino, G., Kaufhold, S., and Arnold, J., 2006,Electrical properties of hydrothermally altered dacite from thePACMANUS hydrothermal field (ODP Leg 193): Journal ofVolcanology and Geothermal Research, 152, 109–120. doi:10.1016/j.jvolgeores.2005.10.002

Chave, A. D., and Thomson, D. J., 1989, Some comments onmagnetotelluricresponse function estimation: Journal of Geophysical Research, 94,14215–14225. doi:10.1029/JB094iB10p14215

Chen, J., andAlumbaugh,D. L., 2011, Threemethods formitigating airwavesin shallow water marine controlled-source electromagnetic data:Geophysics, 76, F89–F99. doi:10.1190/1.3536641

Ciaramella, A., Lauro, E. D., Martino, S. D., Lieto, B. D., Falanga, M., andTagliaferri, R., 2004, Characterization of Strombolian events by usingindependent component analysis:NonlinearProcesses inGeophysics,11,453–461. doi:10.5194/npg-11-453-2004

Cichocki, A., and Amari, S., 2002, Adaptive blind signal and imageprocessing: John Wiley & Sons, Inc.

Constable, S., 2010, Ten years of marine CSEM for hydrocarbon exploration:Geophysics, 75, 75A67–75A81. doi:10.1190/1.3483451

Constable, S., 2013, Reviewpaper: Instrumentation formarinemagnetotelluricand controlled source electromagnetic sounding:Geophysical Prospecting,61, 505–532. doi:10.1111/j.1365-2478.2012.01117.x

Constable, S., and Srnka, L. J., 2007, An introduction to marine controlled-source electromagnetic methods for hydrocarbon exploration:Geophysics, 72, WA3–WA12. doi:10.1190/1.2432483

Delaney, J. R., Robigou, V.,McDuff, R. E., and Tivey,M. K., 1992, Geologyof a vigorous hydrothermal system on the Endeavour Segment, Juande Fuca Ridge: Journal of Geophysical Research, 97, 19 663–19 682.doi:10.1029/92JB00174

Egbert, G. D., and Booker, J. R., 1986, Robust estimation of geomagnetictransfer functions: Geophysical Journal of the Royal AstronomicalSociety, 87, 173–194. doi:10.1111/j.1365-246X.1986.tb04552.x

Ellingsrud, S., Eidesmo, T., Sinha, M. C., MacGregor, L. M., and Constable,S., 2002, Remote sensing of hydrocarbon layers by Seabed Logging(SBL): results from a cruise offshore Angola: The Leading Edge, 21,972–982. doi:10.1190/1.1518433

Furukawa, A., Kimono, J., and Otsuka, H., 2006, Development of responseextraction technique using ICA and its application to structural damageidentification: Journal of Applied Mechanics, 9, 43–54. [in Japanese]doi:10.2208/journalam.9.43

Hampel, F. R., Ronchetti, E. M., Rousseuw, P. J., and Stahel, W. A., 1986,Robust statistics: the approach based on influence functions: JohnWiley& Sons, Inc.

Hyvärinen, A., and Oja, E., 2000, Independent component analysis:algorithms and applications: Neural Networks, 13, 411–430. doi:10.1016/S0893-6080(00)00026-5

Kasaya, T., and Goto, T, 2009, A small ocean bottom electromagnetometerand ocean bottom electrometer system with an arm-folding mechanism:Exploration Geophysics, 40, 41–48. doi:10.1071/EG08118

Kawagucci, S., Miyazaki, J., Nakajima, R., Nozaki, T., Takaya, Y., Kato, Y.,Shibuya, T., Konno, U., Nakaguchi, Y., Hatada, K., Hirayama, H.,Fujikura, K., Furushima, Y., Yamamoto, H., Watsuji, T., Ishibashi, J.,and Takai, K., 2013, Post-drilling changes in fluid discharge pattern,mineral deposition, and fluid chemistry in the Iheya North hydrothermalfield, Okinawa Trough: Geochemistry, Geophysics, Geosystems, 14,4774–4790. doi:10.1002/2013GC004895

H Exploration Geophysics N. Imamura et al.

Key, K., 2016, MARE2DEM: a 2-D inversion code for controlled-sourceelectromagnetic and magnetotelluric data: Geophysical JournalInternational, 207, 571–588. doi:10.1093/gji/ggw290

Løseth, L. O., Amundsen, L., and Jenssen, A. J. K., 2010, A solution to theairwave-removal problem in shallow-water marine EM: Geophysics, 75,A37–A42. doi:10.1190/1.3475359

Murakami, H., and Yamaguchi, S., 2007, Extraction of small geoelectricsignals from noisy geoelectric data using independent componentanalysis: Geoinformatics, 18, 29–37. [in Japanese] doi:10.6010/geoinformatics.18.29

Myer,D., Constable, S., andKey,K., 2011,Broadbandwaveforms and robustprocessing formarineCSEMsurveys:Geophysical Journal International,184, 689–698. doi:10.1111/j.1365-246X.2010.04887.x

Myer, D., Constable, S., Key, K., Glinsky, M. E., and Liu, G., 2012, MarineCSEM of the Scarborough gas field, part 1: experimental design and datauncertainty: Geophysics, 77, E281–E299. doi:10.1190/geo2011-0380.1

Naif, S., Key, K., Constable, S., and Evans, R. L., 2015, Water-rich bendingfaults at the Middle America Trench: Geochemistry, Geophysics,Geosystems, 16, 2582–2597. doi:10.1002/2015GC005927

Nockles, V. A., Wright, D., and Ziolkowski, A., 2009, Power line noiseremoval techniques for transient electromagnetic data: 71st AnnualInternational Conference and Exhibition, EAGE, Extended Abstracts,P253.

Sato, S.,Goto,T.,Kasaya,T.,Kawada,Y., Iwamoto,H., andKitada,K., 2017,Noise reduction method of marine spontaneous electric field data usingindependent component analysis:Butsuri Tansa, 70, 42–55. [in Japanese]doi:10.3124/segj.70.42

Schwalenberg, K., Haeckel, M., Poort, J., and Jegen,M., 2010, Evaluation ofgas hydrate deposits in an active seep area using marine controlled source

electromagnetics: results from Opouawe Bank, Hikurangi Margin,New Zealand:Marine Geology, 272, 79–88. doi:10.1016/j.margeo.2009.07.006

Spagnoli, G., Hannington,M., Bairlein,K.,Hördt, A., Jegen,M., Petersen, S.,and Laurila, T., 2016, Electrical properties of seafloor massive sulfides:Geo-Marine Letters, 36, 235–245. doi:10.1007/s00367-016-0439-5

Strack, K. M., Hanstein, T. H., and Eilenz, H. N., 1989, LOTEM dataprocessing for areas with high cultural noise levels: Physics of theEarth and Planetary Interiors, 53, 261–269. doi:10.1016/0031-9201(89)90010-1

Streich, R., 2016, Controlled-source electromagnetic approaches forhydrocarbon exploration and monitoring on land: Surveys inGeophysics, 37, 47–80. doi:10.1007/s10712-015-9336-0

Streich, R., Becken, M., and Ritter, O., 2013, Robust processing of noisyland-based controlled-source electromagnetic data: Geophysics, 78,E237–E247. doi:10.1190/geo2013-0026.1

Tsuno, S., and Iwata, Y., 2015, Investigation of waveform separation byIndependent Component Analysis for composite signals as a target formixed signals of earthquake groundmotions and railway-induced groundvibrations: Butsuri-Tansa (Geophysical Exploration), 68, 39–47.[in Japanese with English abstract] doi:10.3124/segj.68.39

Weitemeyer, K. A., Constable, S., and Tréhu, A. M., 2011, A marineelectromagnetic survey to detect gas hydrate at Hydrate Ridge,Oregon: Geophysical Journal International, 187, 45–62. doi:10.1111/j.1365-246X.2011.05105.x

Wirianto, M., Mulder, W. A., and Slob, E. C., 2011, Exploiting the airwavefor time-lapse reservoir monitoring with CSEM on land:Geophysics, 76,A15–A19. doi:10.1190/1.3560157

Robust processing of MCSEM data using ICA Exploration Geophysics I

www.publish.csiro.au/journals/eg