Nat. Hazards Earth Syst. Sci., 11, 2341–2353, 2011www.nat-hazards-earth-syst-sci.net/11/2341/2011/doi:10.5194/nhess-11-2341-2011© Author(s) 2011. CC Attribution 3.0 License.

Natural Hazardsand Earth

System Sciences

Neural network based tomographic approach to detectearthquake-related ionospheric anomalies

S. Hirooka1, K. Hattori 1, M. Nishihashi2, and T. Takeda3

1Graduate School of Science, Chiba University, Chiba, Japan2Meteorological Research Institute, Tsukuba, Japan3Department of Computer Science, The University of Electro-Communications, Tokyo, Japan

Received: 27 November 2010 – Revised: 23 February 2011 – Accepted: 4 July 2011 – Published: 26 August 2011

Abstract. A tomographic approach is used to investigatethe fine structure of electron density in the ionosphere. Inthe present paper, the Residual Minimization Training Neu-ral Network (RMTNN) method is selected as the ionospherictomography with which to investigate the detailed structurethat may be associated with earthquakes. The 2007 South-ern Sumatra earthquake (M = 8.5) was selected because sig-nificant decreases in the Total Electron Content (TEC) havebeen confirmed by GPS and global ionosphere map (GIM)analyses. The results of the RMTNN approach are consistentwith those of TEC approaches. With respect to the analyzedearthquake, we observed significant decreases at heights of250–400 km, especially at 330 km. However, the height thatyields the maximum electron density does not change. In theobtained structures, the regions of decrease are located onthe southwest and southeast sides of the Integrated ElectronContent (IEC) (altitudes in the range of 400–550 km) and onthe southern side of the IEC (altitudes in the range of 250–400 km). The global tendency is that the decreased regionexpands to the east with increasing altitude and concentratesin the Southern hemisphere over the epicenter. These resultsindicate that the RMTNN method is applicable to the estima-tion of ionospheric electron density.

1 Introduction

Liu et al. (2006a) statistically examined the relationship be-tween foF2 by ionosonde in Taiwan and 184M ≥ 5.0 earth-quakes in Taiwan that occurred during the period from 1994to 1999. They found a significant decrease in foF2 in the

Correspondence to:S. Hirooka(s-hirooka@graduate.chiba-u.jp)

afternoon one to five days before the investigated earth-quakes. They also reported that significant decreases in To-tal Electron Content derived by GPS (GPS-TEC) appearedone to five days prior toM ≥ 6.0 earthquakes in Taiwan(Liu et al., 2004). The GPS-TEC results can also providea spatial distribution of anomalous regions that may be as-sociated with earthquakes. However, it is difficult to ob-tain information on vertical distributions of electron density.In order to clarify the physical mechanism of earthquake-related ionospheric anomalies, the investigation of three-dimensional electron density distributions is very important.This requires a tomographic approach (Garcia-Fernandez etal., 2003; Saito et al., 2007). These methods require an iono-spheric model and/or a large amount of data for computation,although model-free reconstruction is essential in the case ofdisturbed ionospheric problems. In this sense, the ResidualMinimization Training Neural Network (RMTNN) tomogra-phy proposed by Ma et al. (2005a, b) and Takeda et al. (2007)is a good algorithm because it does not require a model for aninitial value and requires only a small number of data. There-fore, the RMTNN algorithm is used in the present paper.

Sumatra island is one of the most seismically active re-gions in the world. Numerous large earthquakes occur alongthe Sumatra trench every year. A total of 70M > 6.0 earth-quakes have occurred along the Sumatra trench after the2004 Sumatra-Andaman Earthquake (Mw = 9.2). A numberof researchers have reported various anomalous electromag-netic phenomena that may be associated with large earth-quakes in the Indonesia region. These anomalous changesare classified into two types: co-/post-seismic phenomenaand pre-seismic phenomena (Hattori, 2009). Here, we in-troduce a number of aspects related to the large earthquakesaround the Sumatra island. First, co-/post-seismic phenom-ena have been reported. Iyemori et al. (2005) reported geo-magnetic pulsations that may have been caused by the 2004

Published by Copernicus Publications on behalf of the European Geosciences Union.

2342 S. Hirooka et al.: Neural network based tomographic approach

Sumatra-Andaman earthquake. The observed geomagneticpulsations suggest the existence of a large current in spacecaused by acoustic waves based on surface displacements.Otsuka et al. (2006) reported the enhancement of GPS-TECon the north side of the epicenter 14 to 40 min after the 2004Sumatra-Andaman earthquake. They concluded that the en-hancement is caused by acoustic waves generated by theearthquake. Heki et al. (2006) also suggested that the verti-cal displacement of the ruptures propagates to the ionosphereby means of an acoustic wave and generates TEC anoma-lies. Liu et al. (2006b) detected two giant ionospheric dis-turbances by a digital Doppler sounder network in Taiwan.They reported that the first disturbance was excited primar-ily by Rayleigh waves and that the second disturbance wasattributable to the coupling of the atmospheric gravity waves(AGW) excited by broad crustal uplift together with the sub-sequent large tsunami waves. Balasis and Mandea (2007)investigated magnetic and electron density data observed byCHAMP for the 2004 and 2005 Sumatra earthquakes. Theyobserved some specific features after two earthquakes hav-ing periods of approximately 16 and 30 s. The results fa-vored an external source origin for the 30 s pulsation. Hasbiet al. (2009) observed TEC oscillations with periods of 5 to10 min approximately 10 to 24 min after the 2005 SumatraEarthquakes (M = 8.6 andM = 6.7), which had horizontalpropagation velocities of 922 to 1259 m s−1. In addition,magnetic field measurements revealed rapid fluctuations of4 to 5 s shortly after the earthquake, followed by a Pc5 pulsa-tion of 4.8 min approximately 11 min after the event. A goodcorrelation was found between the filtered TEC and Pc5 pul-sation, suggesting that the acoustic pulse generated near theepicenter propagates upward to ionospheric heights and in-duces disturbances in the ionospheric E and F regions.

The pre-seismic phenomena around Sumatra island are de-scribed in the following. Anomalous TEC changes possi-bly preceding large earthquakes around Taiwan have beendescribed above, and similar results have been reported byHattori et al. (2009), who statistically examined the relation-ship between TEC by global ionosphere map (GIM) and 184M ≥ 5.0 earthquakes around Sumatra island that occurredduring the period from 1998 to 2008. They found a signifi-cant decrease in TEC at 3 to 17 days before the earthquakesof interest. They also reported that significant decreases inTEC appeared five days prior to the 2004 Sumatra-Andamanearthquake. Furthermore, the anomalous propagation of VLFradio waves has been investigated. For example, Horie etal. (2007) investigated data for subionospheric VLF wavesfrom the Australian VLF transmitter NWC (19.8 kHz) atthe Japanese receiving stations at Chofu, Chiba, and Kochi.The VLF amplitude data at Japanese stations indicated a re-duction in amplitude and an enhancement of the nighttimeamplitude fluctuation before the earthquake. Molchanov etal. (2006) investigated the reception on board the DEME-TER satellite of the VLF signals transmitted by the NWC.They found a drop in the signals (scattering spot) connected

with the occurrence of the 2004 Sumatra-Andaman Earth-quake. The extension of the “scattering spots” zone was sig-nificantly large (1000–5000 km).

As mentioned above, although the 2004 Sumatra-Andaman Earthquake is a very interesting target, the num-ber of useful GPS data is insufficient for the present study.Therefore, we consider herein the 2007 Southern Sumatraearthquake (M = 8.5) because of the availability of GPS net-work data and because significant decreases in GIM-TEChave been detected (Hattori et al., 2009).

In the present paper, we demonstrate the performanceof the developed RMTNN method and investigate the 3Dstructure of possible earthquake-related ionospheric anoma-lies associated with the 2007 Southern Sumatra earthquake(M = 8.5) by GPS-TEC. In order to verify the reliability ofresults obtained by the developed RMTNN method, we com-pared estimated electron densities with occultation data ob-served by FORMOSAT-3/C.

2 Ionospheric tomography using a neural network

The residual minimization training neural network(RMTNN) is a multilayer neural network that is trained byminimizing an object function composed of an appropriatelyprepared residual of equations (Ma et al., 2000; Liaqat etal., 2003). In this method, we reconstruct three-dimensionalionospheric electron density distributions as a computertomographic image by using the ability of a multilayerneural network system to approximate an arbitrary function.

2.1 Basic function

A slant TEC (Mannucci et al., 1998) along a ray path be-tween a GPS satellite and a ground receiver is defined as theintegrated value of ionospheric and plasmaspheric electrondensity, including instrumental bias, as follows:

Iji =

∫ rj

r i

N(r)ds +Bi +Bj (1)

whereIji is the slant TEC,N(r) is the electron density, and

Bi andBj are the instrumental bias of thei-th ground re-ceiver and thej -th satellite bias, respectively. In order todetermineN(r), which represents the electron density at aposition r, a neural network is constructed. The input pa-rameter of the proposed neural network isr, and the outputis N(r).

In order to evaluate the residuals of the integral equation,the equation is discretized as

Iji ≈

Q∑q=1

αqN(r)+Bi +Bj+P

ji (2)

whereq andα denote a sampling point and the correspond-ing weight for the numerical integration, respectively,Q is

Nat. Hazards Earth Syst. Sci., 11, 2341–2353, 2011 www.nat-hazards-earth-syst-sci.net/11/2341/2011/

S. Hirooka et al.: Neural network based tomographic approach 2343

Fig. 1. Schematic diagram of the data flow for ionospheric tomog-raphy using RMTNN. In the figure,1w is the update value for themain neural network, and1Bi and1Bj are the instrumental biasof theith ground receiver and thej -th satellite bias, respectively.

the total number of sampling points on a ray path, andP isthe contribution of the plasmaspheric electron density to theslant TECI . We herein define altitudes of 100 to 700 kmas the ionosphere and altitudes above 700 km as the plasma-sphere modeled by the simple diffusive equilibrium model(Angerami and Thomas, 1964). In order to estimate the elec-tron densityN(r), we take the squares of the residuals ofthe above integral equations as the objective function of theneural network. The objective functionE1 is given as

E1=

Q∑q=1

(αqN(r)+Bi +Bj+P

ji −I

ji )2. (3)

2.2 Assimilation of ionosonde data

The disadvantage of ionospheric tomography using theground GPS receivers is the difficulty in obtaining sufficientvertical resolution due to a lack of paths. In order to solvethis problem, we use information on the peak electron den-sity (NmF2) and the corresponding height (hmF2) observedby an ionosonde station for restriction. The neural networkwas trained using these data through the conventional super-vised back propagation algorithm (Rumelhart et al., 1986).The object functionE2 for the ionosonde data is given by

E2=

S∑s=1

(Ns(rs)−N ions )2 (4)

whereS is the number of ionosondes,Ns is the output ofthe neural network for the corresponding position that gives

Fig. 2. Ionospheric electron density distribution at longitude100◦ E. (a) Original electron density distribution using the NeQuickmodel.(b) Reconstructed electron density distribution.

hmF2, andN ions is the observed NmF2 value. Therefore, the

overall objective functionE is derived as

E = gE1+E2 (5)

where g is a balance parameter between the GPS andionosonde data.

2.3 Procedure

Figure 1 shows a schematic diagram of the data flow forionospheric tomography by RMTNN. The RMTNN systemis composed of the main neural network for reconstructionand additional neural networks consisting of single neuronsfor instrumental bias (Bi , Bj ). The receiver biasBi and thesatellite biasBj are estimated at the time of the reconstruc-tion. The procedure used to obtain the electron density is asfollows:

1. Provide initial weights with random values.

2. Provide the input data (coordinates) and the desired out-put data (observed STEC values) for the neural network.

3. Compute the output value of the main neural networkand evaluate the objective functionE1.

4. Update the weight values by the back propagationmethod (for RMTNN) as follows.

For the objective functionE1, the weight updating processof the main neural network is derived in clusters for then-thpath as

1w = −η ∂E1∂w

≈ −ηQ∑

q=1(INN

n +Bi +Bj+P

ji −In)

∂INNn

∂Nq∂Nq

∂w

= −ηQ∑

q=1(INN

n +Bi +Bj+P

ji −In)αq

∂Nq

∂w

(6)

wheren represents the ray path (i, j), INNn is a line integral

along the numerically obtainedn-th path using the outputdata of the neural network,w is the weight of the main neuralnetwork, andη is the learning rate. Return to Step 2 until theall ray path is processed.

www.nat-hazards-earth-syst-sci.net/11/2341/2011/ Nat. Hazards Earth Syst. Sci., 11, 2341–2353, 2011

2344 S. Hirooka et al.: Neural network based tomographic approach

Fig. 3. Differences in the integrated electron content (IEC) betweenthe original and the reconstructed data at 150 km intervals from al-titudes of 100 to 700 km.(a) IEC at altitudes of 100 to 250 km,(b) IEC at altitudes of 250 to 400 km,(c) IEC at altitudes of 400 to550 km altitudes, and(d) IEC at altitudes of 550 to 700 km.

5. Provide the input data (a coordinate of hmF2 above theionosonde station) and the desired output data (an ob-served NmF2 value).

6. Compute the output values and evaluate the objectivefunctionE2.

7. Update the weight values by the general back propaga-tion method (Rumelhart et al., 1986) for objective func-tion E2.

8. Return to Step 2 until a criterion is satisfied. Inthe present paper, the number of iterations is fixed as10 000. Regarding the errorE, after 10 000 iterations,E < 10−2. The criterion ofE is empirically determinedby numerical simulations. If the iteration process is notconvergent, change the initial weight values and retrythe iteration process.

In the present paper, we used data for a period of 15 min forthe reconstruction. Therefore, we assume that the electrondensity distribution is constant during this period. We alsoassume that the electron density is constant within an area of0.5 degrees latitude by 0.5 degrees longitude and 30 km inaltitude.

3 Simulation

Ma et al. (2005b) demonstrated the effectiveness of recon-structing the electron density using the RMTNN algorithm

Fig. 4. Map of the epicenter of the 2007 Southern Sumatra earth-quake, GPS ground receivers, and Kototabang ionosonde station.

for the GEONET in Japan for quiet periods. The targetof the present study is the Sumatra region, Indonesia, andthe reconstruction is difficult because of the small numberof GPS receivers. Therefore, we investigated the numeri-cal simulation. For the simulations, the actual positions ofthe GPS receivers and the ionosonde station around Suma-tra and model STEC values are used. The model STECvalues are generated by the NeQuick model (Radicella etal., 2001; Nava et al., 2008). NeQuick model is a three-dimensional, time-dependent ionospheric electron densitymodel developed by the Aeronomy and RadiopropagationLaboratory of the Abdus Salam International Centre for The-oretical Physics (ICTP) and the Institute for Geophysics andMeteorology of the University of Graz. We compared thereconstructed electron density distribution with that derivedfrom the NeQuick model. In this simulation, the datasetof 18:00–18:15 UT, 9 September 2007 is prepared. For theRMTNN, 22 GPS receivers from SuGAr, two GPS receiversfrom IGS, and the observation from the ionosonde station lo-cated in Kototabang are used. The reconstructed region is10◦ S–10◦ N and 90◦ E–110◦ E in latitude and longitude and100 to 700 km in altitude. The model and reconstructed dis-tributions are shown in Fig. 2a and b, respectively. Thesefigures are drawn for a vertical cross section at a fixed longi-tude of 100◦ E. The reconstruction data agrees well with themodel data. Figure 3 shows the integrated electron content(IEC) at altitude intervals of 150 km from 100 to 700 km fora detailed comparison of the model and the reconstruction.The IECs at altitude ranges of 250 to 400 km, 400 to 550 km,and 550 to 700 km exhibit only slight errors, except at the

Nat. Hazards Earth Syst. Sci., 11, 2341–2353, 2011 www.nat-hazards-earth-syst-sci.net/11/2341/2011/

S. Hirooka et al.: Neural network based tomographic approach 2345

Fig. 5. Variations of GIM-TEC*,Kp index, and Dst index from 1 September to 13 September 2007 for various longitudes. The solid greenline shows the time at which the earthquake occurred. The blue areas indicate periods during which a significant decrease in GIM-TEC*over−2σ occurred at multiple stations. Red frames mark significant decreases in GIM-TEC* around the epicenter.

edge of the reconstruction area. The IEC at altitudes in therange of 100 to 250 km is underestimated compared to otheraltitudes and appears to be latitude dependent. This seems tobe due to a lack of restrictions on electron density at loweraltitudes. However, the difference is within±1 TECU at allIECs. Therefore, it is safe to say that the RMTNN recon-struction is applicable for sparse conditions.

4 Application to the 2007 Southern Sumatraearthquake

The Southern Sumatra earthquake (M = 8.5 and depth34 km) occurred on 12 September 2007. The epicenter waslocated at 4.520◦ S, 101.374◦ E, off the south coast of Suma-tra, Indonesia. A map of the epicenter, the selected GPS

receivers, and the ionosonde is shown in Fig. 4. Before ap-plying the proposed RMTNN method, we confirm the exis-tence of possible TEC anomalies before the earthquake. Af-ter reconstruction of the electron density, we compare theresults with GPS occultation data obtained by satellite.

4.1 TEC anomalies before the 2007 Southern Sumatraearthquake

In order to estimate the TEC anomaly associated with the2007 Southern Sumatra earthquake, we used TEC data setsfrom the GIM published by the Center for Orbit Determina-tion in Europe (CODE). We hereinafter refer to GIM-TECasTEC from GIM data and GPS-TEC as TEC from GPS data.The computation of the GIM-TEC is performed as follows.The original spatial resolution of GIM is 2.5◦ in latitude and

www.nat-hazards-earth-syst-sci.net/11/2341/2011/ Nat. Hazards Earth Syst. Sci., 11, 2341–2353, 2011

2346 S. Hirooka et al.: Neural network based tomographic approach

Fig. 6. Variations of GIM-TEC*,Kp index, and Dst index from 1 September to 13 September 2007 for various latitudes. The solid greenline shows the time at which the earthquake occurred. The blue areas indicate periods during which a significant decrease in GIM-TEC*over−2σ occurred at multiple stations. Red frames mark significant decreases in GIM-TEC* around the epicenter.

5.0◦ in longitude, and the temporal resolution is 2 h. We per-formed linear interpolation to obtain a resolution of 1 hourat a certain resolution. We computed the mean GIM-TECvalues for the previous 15 days and the associated standarddeviation as a reference at specific times. Then, the normal-ized GIM-TEC (GIM-TEC*) was defined as follows:

GIM TEC∗(t) =GIM TEC(t)−GIM TECmean(t)

σ (t)(7)

where GIMTEC(t) is the extracted GIM data at timet ,GIM TECmean(t) is the mean value for the previous 15 days,andσ(t) is the associated standard deviation. Figures 5 and 6show the variations in GIMTEC* in terms of the longitudeand latitude, respectively. The corresponding variations intheKp and Dst indices are also shown in both figures. Geo-magnetic conditions were found to be relatively quiet duringthe analyzed period. The GIMTEC* around the epicenterexceeded the lower threshold of−2σ on 4 and 9 September,which was eight and three days before the earthquake. Inparticular, the decrease at 07:00–08:00 UT (14:00–15:00 LT)on 9 September was significant. Figure 7 indicates the dis-turbed area derived from GIM-TEC and GPS-TEC investi-gations, and the disturbed area is found to extend 10 degrees

in latitude and 40 degrees in longitude. Based on the aboveresults, there was a possible TEC anomaly before the earth-quake. Therefore, the proposed RMTNN method was ap-plied to the earthquake in order to investigate the structure ofpre-seismic ionospheric anomaly.

4.2 Results of RMTNN tomography

The 24 selected GPS receivers, including 22 stations fromSuGAr and two stations from IGS, and the ionosonde sta-tion (Kototabang) for the restriction, are shown in Fig. 4. Inthe present paper, we focus on the period of 07:30–07:45 UT(14:30–14:45 LT) on 9 September 2007, during which an ab-normal TEC decrease was detected by GIM-TEC*. In orderto clarify the anomalous changes in three-dimensional elec-tron density distributions, we computed the difference be-tween the reconstructed data on 9 September and the modeldata that is constructed by the median value through 15-daybackward computations. The reconstructed region is 10◦ S–10◦ N in latitude and 90◦ E–110◦ E in longitude and 100 to700 km in altitude, which is the same area used in the simu-lation.

Nat. Hazards Earth Syst. Sci., 11, 2341–2353, 2011 www.nat-hazards-earth-syst-sci.net/11/2341/2011/

S. Hirooka et al.: Neural network based tomographic approach 2347

Fig. 7. Spatial distribution of anomalous GIM-TEC* and GPS-TEC* on 9 September 2007. The red star and the solid red line in-dicate the epicenter and magnetic equator, respectively. Pink circlesand orange triangles indicate points/stations at which GIM-TEC*and GPS-TEC* values, respectively, produced anomalies. Purplecircles and blue triangles indicate positions at which GIM-TEC*and GPS-TEC* values, respectively, did not produce anomalies.Note that GPS-TEC* was calculated directly by data observed fromthe GPS receiver as well as GIM-TEC*.

Figure 8 shows the difference in the IEC at altitudes of100 to 700 km between the observed data and the model data.The IEC (100–700 km) is found to decrease significantly inthe Southern Hemisphere, including over the epicenter. Thisregion corresponds to the regions of Figs. 5 through 7 derivedby GIM-TEC*. Furthermore, a region in which the IEC in-creased exists in the northwest region.

Figure 9 shows the IEC at altitude intervals of 150 km from100 to 700 km. The IEC over altitudes of 100 to 250 km in-dicates that the density decreases in the southwest region,including the epicenter. On the other hand, an increase oc-curs over the Sumatra-Java islands. The IECs at altitudesof 250 to 400 km and 400 to 550 km decrease significantlyaround the epicenter. The IEC at altitudes of 550 to 700 kmdecreases slightly, except in the southwest and in regions eastof the epicenter.

The ionospheric electron density profile above the epicen-ter obtained from the tomographic data is shown in Fig. 10.The solid line indicates the profile on 9 September, and thedashed line represents the median profile computed from thereconstructed data during 25 August and 8 September. Theobserved electron density on September 9 is 40 % of the me-dian value at the peak electron density. And, hmF2 is similarin both cases at an altitude of 330 km. Moreover, the electrondensity profile observed on 9 September is lower than firstquartile value except below 230 km altitude. Figure 11 showsthe longitude-altitude and latitude-altitude cross sections,

Fig. 8. Difference in the integrated electron content (altitudes of100 to 700 km) between the observed data and the reference datawith the 15-day backward median.

which include the epicenter. In addition, Fig. 12 shows onlythe region in which the difference exceeded the Inter Quar-tile Range (IQR) in Fig. 11. The results also show a sig-nificant decrease in the electron density at approximately330 km. The area of this decrease expands eastward withaltitude and is concentrated in the Southern Hemisphere overthe epicenter. Moreover, in order to determine whether theoutliers arise from incidental errors, we performed tomogra-phy around the period of 07:30–07:45 UT. Figures 13 (1) and13 (2) show the anomalous features (i.q. Fig. 11) for the pe-riods of 07:15–07:30 UT and 07:45–08:00 UT, respectively.These distributions of anomalies are similar to those shownin Fig. 11, which suggests that the obtained results (Figs. 9through 12) are reasonable.

4.3 Evaluation of the electron density estimated byFORMOSAT-3 data

In order to validate the electron density distribution ob-tained by the RMTNN algorithm, we compare the verti-cal electron density profiles derived by tomography withFORMOSAT-3/C radio occultation measurements (Lei et al.,2007). FORMOSAT-3/C is composed of six microsatel-lites, each of which houses a GPS occultation experiment.The accuracy of the ionospheric electron density profile ob-served by FORMOSAT-3/C is extremely high (Lei et al.,2007). However, the observable time and space depends onthe positional relationship of the satellites. In the presentstudy, we analyze the occultation data for the afternoon pe-riod of 06:00–09:00 UT (13:00–16:00 LT) during 25 Augustand 9 September. These data provide the electron densityprofile for the 9 September and the 15-day backward median

www.nat-hazards-earth-syst-sci.net/11/2341/2011/ Nat. Hazards Earth Syst. Sci., 11, 2341–2353, 2011

2348 S. Hirooka et al.: Neural network based tomographic approach

Fig. 9. Differences in the integrated electron content (IEC) between the observed and the reference data at 150 km intervals from altitudes of100 to 700 km.(a) IEC at altitudes of 100 to 250 km,(b) IEC at altitudes of 250 to 400 km,(c) IEC at altitudes of 400 to 550 km altitudes,and(d) IEC at altitudes of 550 to 700 km.

Fig. 10. Ionospheric electron density profile obtained using theRMTNN method. The solid line indicates the profile from thereconstructed data at 07:30–07:45 UT on 9 September 2007. Thedashed and dotted lines indicate the median and the first and thirdquartile values, respectively, of the reconstructed data at 07:30–07:45 UT from 25 August to 8 September 2007.

for the day. The investigated orbits of the satellite are in theregion of 5◦ S–15◦ N and 80◦ E–120◦ E, as shown in Fig. 14.In this region, a significant decrease in GIM-TEC* was de-tected, as shown in Fig. 15, which shows the electron densityprofiles obtained by FORMOSAT-3/C. The F2 peak densityon 9 September was found to be decreased significantly (byapproximately 25 %) compared to the median density. Thehmf2 is approximately the same in both cases and at 340 km.The tendency of this profile is consistent with that derivedusing the RMTNN method (Fig. 10). However, the estima-tions by the RMTNN tomography and by FORMOSAT-3/Cbelow 200 km are different, and the RMTNN estimation pro-vides lower values. This tendency is also in agreement withthe results of previous studies (Hsiao et al., 2009, 2010).

5 Conclusion and discussion

In the present paper, we applied the RMTNN method to es-timate the electron density distribution of the ionosphere us-ing STEC and ionosonde data. The analyzed day is threedays before the 2007 Southern Sumatra earthquake. Regionsof significantly reduced 15-day backward median values ofelectron density at the same local time are recognized in the

Nat. Hazards Earth Syst. Sci., 11, 2341–2353, 2011 www.nat-hazards-earth-syst-sci.net/11/2341/2011/

S. Hirooka et al.: Neural network based tomographic approach 2349

Fig. 11. Example of electron density distributions (07:30–07:45 UT). The red dots indicate the projection of the epicenter. Left panelindicates the IECs.(a) Longitudinal difference in electron density between the observed data and the reference data at the same latitude ofepicenter (4.520◦ S). (b) Latitudinal difference in electron density between the observed data and the reference data at the same longitude ofthe epicenter (101.374◦ E).

Fig. 12. Example of electron density distribution in which the electron density exceeds the IQR (07:30–07:45 UT). Black areas indicateregion in which the electron density is smaller than the IQR. Left panel indicates the IECs.(a) Longitudinal difference in electron densitybetween the observed data and the reference data at the same latitude of epicenter (4.520◦ S). (b) Latitudinal difference in electron densitybetween the observed data and the reference data at the same longitude of the epicenter (101.374◦ E).

TEC, IEC, and profile analyses using reconstructed tomo-graphic data. The characteristic of the reduction for the TECanalysis is consistent with the characteristics of the GIM-TEC and GPS-TEC. The results for the IEC indicate thatsignificant decreases occur at altitudes of 250 to 400 km,especially at 330 km. However, the height that yields themaximum electron density remains the same. The obtainedstructure is such that a region of significantly reduced 15-day backward median values of electron density exists in thesouthwest and southeast side of the IEC (altitude: 400 to550 km) and in the southern side of the IEC (altitude: 250to 400 km). The global tendency is that this region expandsto the east with increasing altitude and is concentrated in theSouthern Hemisphere over the epicenter.

Anomalous pre-seismic features obtained by GIM/GPS-TEC and tomographic TEC/ IEC exhibited good agreementwith the results of previous studies. Hattori et al. (2009) re-ported GPS/GIM-TEC anomalies during the 2004 Sumatra-Andaman Earthquake. They found an anomalous reductionin GPS/GIM-TEC during the afternoon five days before theearthquake, and reported that the region of decreased elec-tron density expanded 30 degrees in latitude and 40 degreesin longitude from the epicenter. Meanwhile, Liu et al. (2009)reported a reduction in GIM-TEC during the afternoon of6–4 and three days before the 2008 Wenchuan earthquake(Mw = 7.9) and that the region spread 10–15 degrees in lati-tude and 15–30 degrees in longitude. Similarly, we observedanomalous decreases in GIM/GPS-TEC around the epicen-ter during the afternoon 3 days before the earthquake and

www.nat-hazards-earth-syst-sci.net/11/2341/2011/ Nat. Hazards Earth Syst. Sci., 11, 2341–2353, 2011

2350 S. Hirooka et al.: Neural network based tomographic approach

Fig. 13. Example of electron density distributions (1) 07:15–07:30 UT, (2) 07:45–08:00 UT. The red dots indicate the projection of theepicenter. Left panel indicates the IECs.(a) Longitudinal difference in electron density between the observed data and the reference data atthe same latitude of epicenter (4.520◦ S). (b) Latitudinal difference in electron density between the observed data and the reference data atthe same longitude of the epicenter (101.374◦ E).

that the anomalous area spread 10 degrees in latitude and40 degrees in longitude. The tendency for disturbed areas intime and space is found to be common. Although the recon-struction area obtained by tomography is not sufficient forthe estimation of the spatial extent of the anomalous area, theobtained results are consistent with the result of GPS/GIM-TEC. In addition, the electron density at hmF2 exhibits asignificant decrease, which agrees with previous results ob-tained by ionosonde (Liu et al., 2006a).

The profile obtained from the tomographic reconstructionis compared with that obtained from FORMOSAT-3/C in or-der to evaluate the performance of the proposed RMTNNmethod. The results indicate good agreement, with the ex-ception of altitudes in the range of 100 to 200 km. The ten-dency of the decrease in electron density at approximately300 km before the occurrence of large earthquakes has beenreported (Hsiao et al., 2009, 2010), which suggests the highapplicability of the RMTNN method to the estimation ofionospheric electron density.

Differences between the results of the present study andthe results of previous studies were also found. Hsiao etal. (2009) reported the three-dimensional ionospheric struc-tures obtained prior to the 2006 Pingtung earthquake us-ing FORMOSAT-3/C. They found decreases in the F2 peakheight, significant decreases in the electron density at alti-tudes of from 300 to 350 km, and increases in the lower al-titudes within five days prior to the earthquakes. Similar re-sults have been reported for the 2008 Wenchuan earthquake(Liu et al., 2009; Hsiao et al., 2010). On the other hand, theRMTNN results of the present study indicate no enhance-ment at lower altitudes. The simulation results also revealedthat the IEC at altitudes of 100 to 250 km are slightly under-estimated compared to the model. With respect to the peakheight, no obvious reduction was found from the profiles de-rived from the RMTNN and FORMOSAT-3/C data. This in-dicates that no remarkable reduction in hmF2 occurred forthe 2007 Southern Sumatra earthquake.

One problem associated with the proposed method is theunderestimation of the electron density at lower altitudes.

Nat. Hazards Earth Syst. Sci., 11, 2341–2353, 2011 www.nat-hazards-earth-syst-sci.net/11/2341/2011/

S. Hirooka et al.: Neural network based tomographic approach 2351

Fig. 14. Locations of points at which profiles were computed fromFORMOSAT-3/C occultation data during 06:00–09:00 on 25 Au-gust 2007 and 9 September 2007. The red star indicates the epi-center. The blue line and the black lines represent the trajectorieson 9 September 2007 and for the reference data (15-day backwardmedian).

This might arise from the lack of information on the lowerionosphere. Additional information on the bottom soundingand/or lower height data using satellite occultation may berequired. These problems must be solved in order to improvethe RMTNN method. If these problems are solved, we willbe able to estimate the electron distribution at the lower iono-sphere and discover the sources of nighttime fluctuations andtermination time shifts in subionospheric VLF waves (Horieat al., 2007).

Even though these results are preliminary, the RMTNNmethod reveals an interesting three-dimensional structurethat may be associated with the 2007 Southern Sumatraearthquake. Such a structure might provide a key to a possi-ble solution to pre-seismic ionospheric anomalies. Recently,several mechanisms for lithosphere-atmosphere-ionospherecoupling, such as the atmospheric electric field, the atmo-spheric gravity wave (AGW), and gas emanation, have beenproposed (Pulinets and Boyarchuk, 2004; Freund, 2000;Molchanov et al., 2006). If positive holes excited around theepicenter diffuse and outflow from the epicenter, an upwardelectric field is generated on the surface of the earth (Freund,2000). The upward electric field and the magnetic field ofthe earth generate a westward E× B drift of plasma (Kelley,1989). Then, the reduction of ionospheric electron densityoccurs east of the epicenter. We observed a significant de-crease in electron density east of the epicenter (include overthe epicenter) at IEC (250 to 400 km) obtained by tomog-raphy. However, the obtained region of decreased electrondensity did not extend northward from the epicenter and wasconcentrated at 300 km. These results are inconsistent with

Fig. 15. Ionospheric electron density profiles obtained fromFORMOSAT-3/C. The solid line indicates the profile observed on9 September. The dashed and dotted lines indicate the medianand first and third quartile values calculated during the period from25 August to 8 September, respectively.

the theory proposed by Pulinets and Legan’ka (2003). Theyreported that the electric field is transferred into the iono-sphere along the magnetic field lines, so that, in most cases,the disturbed region is shifted to the north from the epicen-ter projection in the southern hemisphere. Therefore, in thiscase, other factors might need to be considered regarding therelationship between the electric field and ionospheric distur-bances. Further investigations of the eastward expansion ofregions of anomalous decreases with increasing altitude willbe required.

Continuous tomographic images can provide dynamicvariations of the electron density, which reveal the trans-port of energy from the lithosphere. In other words, the to-mographic approach has the advantage of providing a finespatial distribution and is important for clarifying the cou-pling mechanism between earthquake preparation processesand ionospheric anomalies. In the present paper, we consid-ered only three periods (07:15–07:30 UT, 07:30–07:45 UT,and 07:45–08:00 UT) of tomographic results. In order to dis-cuss the development of pre-ionospheric anomalies, longer-term continuous analysis is required. Moreover, in order toenhance the reliability of anomalous structures obtained bytomography, some stochastic test (Huang, 2006) and/or con-tinuous analysis are important. Therefore, improvement ofprocessing efficiency will be required.

The proposed RMTNN tomography is useful for the studyof not only pre-seismic ionospheric disturbances but alsoco/post-seismic disturbances and general upper atmosphericphenomena. If RMTNN tomography is used to examineco/post-seismic ionospheric disturbances, clarification of thepropagation process of the AGW generated by earthquakes

www.nat-hazards-earth-syst-sci.net/11/2341/2011/ Nat. Hazards Earth Syst. Sci., 11, 2341–2353, 2011

2352 S. Hirooka et al.: Neural network based tomographic approach

and/or tsunamis in the ionosphere may become possible. Theresults of the present study indicate the applicability of theRMTNN method to the investigation of ionospheric electrondensity.

Acknowledgements.The authors would like to thank the ScrippsOrbit and Permanent Array Center (SOPAC) for GPS data of the In-ternational GNSS Service (IGS) and Sumatran GPS Array (SuGAr)stations, the National Institute of Information and CommunicationsTechnology, Japan for the ionosonde data at Kototabang, theCenter for Orbit Determination in Europe (CODE) for the GIMdata, and the World Data Center (WDC) for Geomagnetism,Kyoto University for the Dst and Kp index. We would also liketo thank X. F. Ma for his helpful suggestion concerning softwaredevelopment. The present research is supported in part by aGrant-in-Aid for Scientific Research from the Japan Society forPromotion of Science (No. 19403002) and National Institute ofInformation and Communication Technology (R & D promotionfunding international joint research).

Edited by: K. EftaxiasReviewed by: two anonymous referees

References

Angerami, J. J. and Thomas, J. O.: Studies of planetary atmo-spheres, The distribution of electrons and ions in the Earth’s ex-osphere, J. Geophys. Res., 69, 4537–4560, 1964.

Balasis, G. and Mandea, M.: Can electromagnetic disturbancesrelated to the recent great earthquakes be detected by satellitemagnetometers?, special issue “Mechanical and ElectromagneticPhenomena Accompanying Preseismic Deformation from Lab-oratory to Geophysical Scale”, Tectonophysics, 431, 173–195,doi:10.1016/j.tecto.2006.05.038, 2007.

Freund, F.: Time-resolved study of charge generation and propa-gation in igneous rocks, J. Geophys. Res., 105, 11001–11019,2000.

Garcia-Fernandez, M., Hernandez-Pajares, M., Juan, M., Sanz, J.,Orus, R., Coisson, P., Nava, B., and Radicella, M.: Combiningionosonde with ground GPS data for electron density estimation,J. Atmos. Sol. Terr. Phys., 65, 683–691, 2003.

Hasbi, A. M., Momoni, M. A., Ali, M. A. M., Misran, N., Sh-iokawa, K., Otsuka, Y., and Yumoto, K.: Ionospheric and geo-magnetic disturbances during the 2005 Sumatra earthquakes, J.Atmos. Sol. Terr. Phys., 71, 1992–2005, 2009.

Hattori, K.: Electromagnetic Phenomena Possibly Associated withthe 2004 Sumatra-Andaman Earthquake, The Institute of Elec-trical Engineers of Japan, Trans. Fundament Materials, 129A,345–351, 2009 (in Japanese).

Hattori, K., Nishihashi, M., and Liu, J. Y.: Statistical and case stud-ies of ionospheric TEC anomalies related to large earthquakes inIndonesia, Geophys. Res. Abstr., Vol. 11, EGU2009–8376, 2009.

Heki, K., Otsuka, Y., Choosakul, N., Hemmakorn, N., Komolmis,T., and Maruyama, T.: Detecttion of ruptures of Andamanfault segment in the 2004 great Sumatra earthquake with coseis-mic ionospheric disturbances, J. Geophys., Res., 111, B09313,doi:10.1029/2005JB004202, 2006.

Horie, T., Yamauchi, T., Yoshida, M., and Hayakawa, M.: Thewave-like structures of ionospheric perturbation associated with

Sumatra earthquake of 26 December 2004, as revealed from VLFobservation in Japan of NWC signals, J. Atmos. Sol. Terr. Phys.,69, 1021–1028, 2007.

Hsiao, C. C., Liu, J. Y., Oyama, K. I., Yen, N. L., Wang, Y. H., andMiau, J. J.: Ionospheric electron density anomaly prior to theDecenber 26, 2006M7.0 Pingtung earthquake doublet observedby FORMOSAT-3/COSMIC, Phys. Chem. Earth, Parts A/B/C,34, 474–478, 2009.

Hsiao, C. C., Liu, J. Y., Oyama, K. I., Yen, N. L., Liou, Y. A.,Chen, S. S., and Miau, J. J.: Seismo-ionospheric precursor of the2008 Mw7.9 Wenchuan earthquake observed by FORMOSAT-3/COSMIC, GPS Solut., 14(1), 83–89,doi:10.1007/s10291-009-0129-0, 2010.

Huang, Q.: Search for reliable precursors: A case studyof the seismic quiescence of the 2000 western Tottoriprefecture earthquake, J. Geophys. Res., 111, B04301,doi:10.1029/2005JB003982, 2006.

Iyemori, T., Nose, M., Han, D., Gao, Y., Hashizume, M.,Choosakul, N., Shinagawa, H., Tanaka, Y., Utsugi, M., Saito,A., McCreadie, H., Odagi, Y., and Yang, F.: Geomagnetic pulsa-tions caused by the Sumatra earthquake on December 26, 2004,Geophys. Res. Lett., 32, L20807,doi:10.1029/2005GL024083,2005.

Kelley, M. C.: The Earth’s Ionosphere, Elsevier, New York, p. 487,1989.

Lei, J., Syndergaard, S., Burns, A. G., Solomon, S. C., Wang, W.,Zeng, Z., Roble, R. G., Wu, Q., Kuo, Y-H, Holt, J. M., Zhang,S-R., Hysell, D. L., Rodrigues, F. S., and Lin, C. H.: Compar-ison of COSMIC ionospheric measurements with ground-basedobservations and model predictions: Preliminary results, J. Geo-phys. Res., 112, A07308,doi:10.1029/2006JA012240, 2007.

Liaqat, A., Fukuhara, M., and Takeda, T.: Optimal estimation ofparameter of dynamical system by neural network collocationmethod, Comput. Phys. Commun., 150, 215–234, 2003.

Liu, J. Y., Chuo, Y. J., Shan, S. J., Tsai, Y. B., Chen, Y. I., Pulinets,S. A., and Yu, S. B.: Pre-earthquake ionospheric anomalies reg-istered by continuous GPS TEC measurements, Ann. Geophys.,22, 1585–1593,doi:10.5194/angeo-22-1585-2004, 2004.

Liu, J. Y., Chen, Y. I., Chuo, Y. J., and Chen, C. S.: A statistical in-vestigation of pre-earthquake ionospheric anomaly, J. Geophys.Res., 111, A05304,doi:10.1029/2005JA011333, 2006a.

Liu, J. Y., Tsai, Y. B., Chen, S. W., Lee, C. P., Chen, Y. C., Yen,H. Y., Chang, W. Y., and Liu, C.: Giant ionospheric disturbancesexcited by the M9.3 Sumatra earthquake of 26 December 2004,Geophys. Res. Lett., 33, L0213,doi:10.1029/2005GL023963,2006b.

Liu, J. Y., Chen, Y. I., Chen, C. H., Liu, C. Y., Chen, C. Y., Nishi-hashi, M., Li, J. Z., Xia, Y. Q., Oyama, K. I., Hattori, K., andLin, C. H.: Seismo-ionospheric GPS TEC Anomalies Observedbefore the 12 May 2008 Mw7.9 Wenchuan Earthquake, J. Geo-phys. Res., 114, A04320,doi:10.1029/2008JA013698, 2009.

Ma, X. F., Fukuhara, M., and Takeda, T.: Neural network CT imagereconstruction method for small amount of projection data, Nucl.Instrum. Method Phys. Res., A, 449, 366–377, 2000.

Ma, X. F., Maruyama, T., Ma, G., and Takeda, T.: Deter-mination of GPS receiver differential biases by neural net-work parameter estimation method, Radio Sci., 40, RS1002,doi:10.1029/2004RS003072, 2005a.

Ma, X. F., Maruyama, T., Ma, G., and Takeda, T.: Three-

Nat. Hazards Earth Syst. Sci., 11, 2341–2353, 2011 www.nat-hazards-earth-syst-sci.net/11/2341/2011/

S. Hirooka et al.: Neural network based tomographic approach 2353

dimensional ionospheric tomography using observation data ofGPS ground receivers and ionosonde by neural network, J. Geo-phys. Res., 110, A05308,doi:10.1029/2004JA010797, 2005b.

Mannucci, A. J., Wilson, B. D., Yuan, D. N., Ho, C. H., Lindqwis-ter, U. J., and Runge, T. F.: A global mapping technique for GPSderived ionospheric electron content measurements, Radio Sci.,33, 562–582, 1998.

Molchanov, O., Rozhnoi, A., Solovieva, M., Akentieva, O., Berthe-lier, J. J., Parrot, M., Lefeuvre, F., Biagi, P. F., Castellana, L., andHayakawa, M.: Global diagnostics of the ionospheric perturba-tions related to the seismic activity using the VLF radio signalscollected on the DEMETER satellite, Nat. Hazards Earth Syst.Sci., 6, 745–753,doi:10.5194/nhess-6-745-2006, 2006.

Nava, B., Coisson, P., and Radicella, S. M.: A new version of theNeQuick ionosphere electron density model, J. Atmos. Sol. Terr.Phys., 70, p. 15,doi:10.1016/j.jastp.2008.01.015, 2008.

Otsuka, Y., Kotake, N., Tsugawa, T., Shiokawa, K., Ogawa, T., Ef-fendy, Saito, S., Kawamura, M., Maruyama, T., Hemmakorn, N.,and Komolmis, T.: GPS detection of total electron content vari-ations over Indonesia and Thailand following the 26 December2004 earthquake, Earth Planet. Space, 58, 159–165, 2006.

Pulinets, S. A. and Boyarchuk, K. A.: Precursors of Earthquakes,Springer, Berlin, Germany, 2004.

Pulinets, S. A. and Legan’ka, A. D.: Spatial-Temporal Character-istics of Large Scale Disturbances of Electron Density Observedin the Ionospheric F-Region before Strong Earthquakes, CosmicRes., 41, 221–229, 2003.

Radicella, S. M., and Leitinger, R.: The evolution of the DGR ap-proach to model electron density profiles, Adv. Space Res., 27,35–40, 2001.

Rumelhart, D. E., Hinton, G. E., and Williams, R. J.: Learning inter-nal representations by error propagation, in: Parallel DistributedProcessing, MIT Press, 1, 318–362, 1986.

Saito, A., Teraishi, S., Ueno, G., Fujita, N., and Tsugawa, T.: GPSIonospheric Tomography over Japan with Constrained Least-squares Method, American Geophysical Union, Fall Meeting,abstract #SA13A-1061, 2007.

Takeda, T. and Ma, X. F.: Ionospheric Tomography by NeuralNetwork Collocation Method, Plasma Fusion Res., 2, S1015,doi:10.1585/pfr.2.S1015, 2007.

www.nat-hazards-earth-syst-sci.net/11/2341/2011/ Nat. Hazards Earth Syst. Sci., 11, 2341–2353, 2011