Neural Network Modelling of Tehran Land Subsidence Measured by Persistent Scatterer … · 2017-03-08 · Measured by Persistent Scatterer Interferometry MARYAM DEHGHANI, Shiraz,

PFG 2013 / 1, 005–0017 ArticleStuttgart, February 2013

© 2013 E. Schweizerbart'sche Verlagsbuchhandlung, Stuttgart, Germany www.schweizerbart.deDOI: 10.1127/1432-8364/2013/0154 1432-8364/13/0154 $ 3.25

Neural Network Modelling of Tehran Land SubsidenceMeasured by Persistent Scatterer Interferometry

MARYAM DEHGHANI, Shiraz, Iran, MOHAMMAD JAVAD VALADAN ZOEJ & IMAN ENTEZAM,Tehran, Iran

Keywords: persistent scatterer, neural network, modelling, sensitivity analysis

Summary: A large area located in the southwest ofTehran is subject to land subsidence induced byover-exploitation of groundwater. Since the use ofconventional SAR Interferometry was not possibledue to the large spatial baseline and rapid temporaldecorrelation, persistent scatterer interferometry(PS-InSAR) using two different ascending and de-scending datasets of ENVISAT ASAR was appliedin order to monitor the deformation. PS-InSAR is arecently developed technique used to address thedecorrelation problem by identifying scatterers,called persistent scatterers (PS), the echo of whichvaries little in time. The estimation of the deforma-tion rate was thus possible only in the PS pixels. Inorder to retrieve the spatial pattern of the subsid-ence, deformation rate at non-PS pixels were esti-mated by a proposed neural network modellingmethod separately applied on both descending andascending datasets. Input variables of the networkare geology and hydrogeology parameters of theaquifer system, while the network output is the sub-sidence rate. The performance of the neural net-work trained by the PS pixels was tested on a sepa-rate validation data. It was found that the trainednetwork is able to predict the subsidence rate withthe accuracy of less than 5 mm/a. The results werethen compared to the levelling measurements ac-quired over a different time interval. The root-mean-square error (RMSE) between the measure-ments and the modelled deformation rate across theleveling tracks is 19.8 mm/a. The different defor-mation rates of both datasets in some points weremost likely due to the different time intervals cov-ered by the radar and levelling data. Neural net-work-based sensitivity analysis was finally per-formed to evaluate the influences of different geol-ogy and hydrogeology factors on the subsidence.The sensitivity analysis results that were interest-ingly similar for both radar datasets showed thatthe hydraulic conductivity, the thickness of fine-grained sediments and the water level decline arethe first three most effective factors on the subsid-ence occurrence in Tehran basin.

Zusammenfassung: Neuronales Netz zur Model-lierung von Bodensenkungen bei Teheran gemes-sen durch Persistent Scatterer Interferometry. Eingroßes Gebiet südwestlich von Teheran ist von Bo-densenkungen wegen übermäßiger Grundwasser-entnahme betroffen. Die vorliegende Untersuchungbaut auf zwei Datensätzen von ENVISAT ASARauf, einem in Nord- und einem in Südrichtung ge-flogenen. Früher war der Einsatz von SAR Interfe-rometrie wegen langer Basen und zu schneller zeit-licher Veränderungen am Objekt nicht möglich.Die neue Persistent Scatterer Methode (PS-InSAR)nutzt Objekte, deren Reflexionscharakteristik nureiner geringen zeitlichen Variabilität unterliegt. Al-lerdings kann die Deformation nur an ausgewähl-ten, den so genannten PS-Pixeln bestimmt werden.Die dazwischenliegenden Räume wurden in dieserUntersuchung durch neuronale Netze überbrückt,die geologische und hydrogeologische Parameterberücksichtigen. Die innere Genauigkeit der Me-thode wurde durch unterschiedliche Trainingsda-tensätze überprüft und ergab eine Genauigkeit derPrognose des Senkungsbetrages von weniger als5 mm/a. Die äußere Genauigkeit wurde durch Ni-vellement bestimmt und ergab einen RMS-Fehlerentlang der Nivellementslinien von 19,8 mm/a. Diegroße Abweichung dürfte auf unterschiedlicheZeitintervalle zwischen der SAR-Datenerfassungund den Nivellements beruhen. Die neuronalenNetze wurden außerdem zur Bewertung des Ein-flusses von geologischen und hydrogeologischenParametern auf die Senkungsraten eingesetzt. Alsdie drei wichtigsten Einflussfaktoren haben sich diehydraulische Leitfähigkeit des Untergrundes, dieMächtigkeit der feinkörnigen Sedimente und derRückgang des Grundwasserspiegels gezeigt. BeideSAR-Datensätze führten zum gleichen Ergebnis.

6 Photogrammetrie • Fernerkundung • Geoinformation 1/2013

The southern part of the area primarily con-tains Quaternary (Qt1 and Qt2) units as in-dicated in Fig. 1b. The Qt1, or Kahrizak for-mation, features different properties in thenorthern and southern parts of the area, con-sists of silt and cream clay in the subsidencearea, and is exposed by different faults such asthe Kahrizak, south Rey and North Rey faults(Fig. 1a). The deposit of fine-grained sedi-ments due to flooding generates the clay silt ofKahrizak. The Qt2, or Tehran formation, con-tains young alluvial fans mainly covering thesouthern part of the Tehran Basin. This forma-tion has an unsorted alluvial and a flood de-posit with an average thickness of 60 m, pri-marily composed of gravel, pebble and sandin a sandy silt matrix (SHEMSHAKI et al. 2005).The extent and thickness of the fine-grained

sediments serves as one of the most importantfactors for land subsidence caused by ground-water extraction. The greater extent and thick-ness an aquifer system has, the more compac-tion occurs. According to the studies per-formed by the Engineering Geology group ofthe Geological Survey of Iran (GSI), the thick-est part of the alluvial layers that constitute theaquifer system belongs to the area with a high-er subsidence rate. The contours indicatingthe alluvial thickness are illustrated in Fig. 4e.Moreover, the central part of the subsidence

1 Introduction

Excessive groundwater pumping causes an in-creasing effective stress in aquifer systems.The changes in the effective stress cause somedegree of compaction in fine-grained sedi-ments within the aquifer system resulting inland subsidence. The land subsidence inducedby overexploitation of groundwater is one ofthe major problems with its environmentalconsequences in most parts of Iran. One ofthese areas is located in the southwest of Teh-ran, the capital city of Iran. The Tehran Basinwith an area of 2250 km2 is located betweenthe Alborz Mountains to the north and Aradand Fashapouye Mountains to the south.The National Cartographic Center (NCC)

of Iran firstly measured the land subsidence byprecision levelling during 1995–2002. Later,the Geological Survey of Iran (GSI) in cooper-ation with other organizations including NCCand the Water Management Organizationlaunched a comprehensive study on subsid-ence in and around Tehran. The area was stud-ied from different points of view including ge-ology, geophysics, tectonic and geotechnics.According to the geological map (1:100,000scale), the Tehran Basin consists of differentgeological units including lower-middle Eo-cene (Karaj formation) and Quaternary units.

Fig. 1: (a) Location of different faults surrounding the study area in Tehran province. The blackrectangle depicts the area subject to subsidence. (b) Geological units mainly consisting of Quater-nary sediments. The blue circles represent piezometric wells while the black triangles indicate thelevelling stations.

Maryam Dehghani et al., Network Modelling 7

between 1968 and 2003 shows that the abil-ity of the aquifer system to yield water hassignificantly decreased due to insufficient re-charge. The water level changes between 1968and 2003 are shown as contour lines in Fig. 4a.Among the various ground- and space-

based techniques available, interferometricSynthetic Aperture Radar (InSAR) providesprecise measurements of land surface defor-mation over large areas and at high spatialresolution (GALLOWAY et al. 1998, AMELUNG etal. 1999, PELTZER et al. 1998, FRUNEAU & SAR-TI 2000, TESAURO et al. 2000, CROSETTO et al.2002, MOTAGH et al. 2006). However, conven-tional interferometry fails when the groundsurface is covered by significant amounts ofvegetation, due to a loss of correlation (GAL-LOWAY & HOFFMANN 2007) as in Tehran ba-sin. Persistent scatterer InSAR (PS-InSAR)is a recently developed technique used to ad-dress the decorrelation problem by identifyingscatterers, called persistent scatterers (PS),the echo of which varies little in time (FER-RETTI et al. 2000, HOOPER et al. 2004, KAMPES2005). PS-InSAR gives measurements at afew sparse locations, i.e. PS points. However,despite of the improvement due to PS-InSARapproaches it should be noted that, if the spa-tial density of the detected persistent scatter-ers is low, the deformation pattern cannot bereliably retrieved. A sufficient number of PSpixels is required to map local variation of thedeformation for deeper understanding of thesubsidence extent and spatial distribution thatenables us to better manage the water resourc-es and construction tasks. It is however impor-

area contains higher amounts of fine-grainedsediments, including clay, at up to 100% asobserved in Fig. 4f. Moreover, geotechni-cal tests indicate that the storage coefficienthave smaller values in the southern part of theaquifer system compared to the northern part(Fig. 4c). The decrease of the storage coeffi-cient is due to the existence of higher amountsof fine-grained sediments in the aquifer sys-tem. The average value of the storage coef-ficient of Tehran aquifer system is estimatedas 5% (SHEMSHAKI et al. 2005). Moreover, asshown in Fig. 4d, the hydraulic conductivity ofthe aquifer system decreases generally fromnorth to south of the area and in the northeastof the area. The decrease of the hydraulic con-ductivity in the northeast is mostly due to thedecrease of thickness of the saturated zone.The main part of the Tehran basin dedicated

to the agricultural activities is subject to theland subsidence due to the over-exploitationof groundwater. According to the latest stud-ies, groundwater level depth exceeds 100 min the north of the aquifer system (Fig. 4b). Aunit hydrograph of the Tehran basin (Fig. 2)indicating the overall changes of water levelwas extracted from the piezometric informa-tion (SHEMSHAKI et al. 2005). The locations ofpiezometric wells are shown in Fig. 1b. As ob-served in Fig. 2, the water level dramaticallydeclined by about 9 m in 20 years. The volumechange of the aquifer system during 10 yearsis estimated as 6.6 km3. The number of pump-ing wells has increased from 3906 in 1969 to26076 in 2003. A comparison of the numberof wells and the volume of the extracted water

Fig. 2: Unit hydrograph for the Tehran aquifer system extracted from the groundwater information(Shemshaki et al. 2005).


tion 3, the information layers fed to the neu-ral network and the obtained results will bepresented. Moreover, the sensitivity analysisresults are given in this section. Finally, con-cluding remarks are given in section 4.

2 Subsidence Modelling basedon Neural Network

Neural networks are based on the structureand functioning of the human brain and con-sist of a large number of simple processingunits known as neurons. The neural networkfunction is determined largely by the connec-tions between neurons. It is trained by adjust-ing the values of the connections (weights) be-tween neurons so that a particular input leadsto a specific target output. The network is ad-justed based on a comparison of the outputand the target until the network output match-es the target.The output signal of neuron k in layer l is

calculated as follows:( ) ( ) ( 1) ( 1)0 1 1

( 1) ( 1) ( 1)1 1 0

( ...

)1

l l l lk k

l lkK K ll l k

l

y s w y

w y wk K

− −

− − −− −

= ⋅ +

+ ⋅ +≤ ≤ (1)

Where Kl is the number of neurons in lay-er l, s(l)0 (.) is the activation function, w(l−1)ki isthe weight of the input y(l−1)ki to the neuron kin which 1 ≤ i ≤Kl−1 and w(0)k0 is the bias val-ue corresponding to the neuron k. It shouldbe noted that y(l−1)ki is actually the output ofith neuron in the previous layer. If l= 1, y(l−1)kiis the input variables to the neural network.Weights are updated using the mean-squarederror (MSE) as cost or error function throughthe training process. Readers are referred toe.g. HAYKIN et al. (1995) for further informa-tion about the training process and updatingthe weights.One of the important factors that affect the

success of modelling procedure is the abilityto extract information about the relationshipsbetween the model structure’s inputs and out-puts from the trained neural network (HASHEM1992). This information can be used as a basisfor the analysis of the model and determina-tion of the most significant factors that affect

tant to have the subsidence spatial extent andpattern for land managements (GONZÁLEZ &FERNÁNDEZ 2011). As in the Tehran basin manyagricultural fields exist, the persistent scatter-er density is low due to the lack of coherentscatterers over the time interval of consider-ation. There are various algorithms to retrievethe spatial pattern of the subsidence using thevalues at PS pixels. Classic interpolation tech-niques such as kriging can be generally usedto estimate the deformation at a non-persistentscatterer based on the known deformation ofsurrounding persistent scatterers. However,these methods are based only on the deforma-tion values of the neighbouring pixels. Thereare some analytical methods that try to pre-dict the aquifer compaction using the geologyand hydrogeology information (HOFFMANN etal. 2003). However, we are left with great er-rors using these methods when we are facingthe lack of accurate and detailed informationof the aquifer system.In such cases, the methods based on the

intelligence systems can be effectively usedinstead of modelling the subsidence. In thispaper we present a method based on a back-propagation neural network in order to mod-el the subsidence signal. The deformation atnon-persistent scatterer is estimated by neuralnetwork interpolation using not only the PS-InSAR derived deformation of the persistentscatterers but also their hydrogeology prop-erties that highly affects the subsidence. Theconstructed neural network is trained usingthe hydrogeology parameters of the aquifersystem as input variables of the network andPS-InSAR derived subsidence rate at PS pix-els as the network output. The PS pixels whosedeformation could be measured by PS-InSARare used to train the network. The trained neu-ral network is then used: (i) to retrieve the spa-tial pattern of the subsidence or to estimatethe subsidence rate at non-PS pixels and to (ii)perform sensitivity analysis to identify a hy-drogeology factor with the most influence onthe subsidence. In other words, we can recog-nize to which hydrogeology factor the subsid-ence in Tehran basin is most sensitive.This paper is structured as follows: Section

2 presents the proposed modelling methodbased on neural network and the mathematicalframework of the sensitivity analysis. In sec-


respectively. xi is its ith input. The updatedweights from training step are used to calcu-late (2). The more sensitive the output w.r.t. to

xi is, the higher thei

yx

∂∂

is. A little change to xiwill then cause a great change to the output ŷ.Subsidence due to the over-exploitation of

groundwater is a complex procedure influ-enced by different geology and hydrogeologyfactors. Groundwater information collected atdifferent piezometric wells in the Tehran basinshowed that the water level decline is not theonly important cause for the subsidence oc-currence in this area (DEHGHANI et al. 2010).Instead, some other geology and hydrogeolo-gy factors such as the fine-grained sedimentsthickness in the aquifer system are controllingthe subsidence rate. In this paper, a methodbased on neural network is presented to mod-el the subsidence for areas in which we couldnot measure the deformation rate using PS-In-SAR. Moreover, in order to identify the main

it. HASHEM (1992) presented a closed-form ex-pression for first order as well as higher orderoutput sensitivities w.r.t. variations in the in-put variables for multilayer feedforward neu-ral networks. These sensitivities can be usedas a basis for inference about input-output re-lationships. In order to estimate the sensitiv-ity of the network output w.r.t. to its inputs,

the partial derivative,i

yx

∂∂

is estimated where

y is the network output and xi is its ith input.Consider a neural network with 2 hidden lay-

ers consisting of K1 and K2 neurons.i

yx

∂∂

isthen calculated as follows:

(2)(1)2 2

(2) (1)212

ˆ ˆ. . 1,...,

k

kki i

Ky y y y i nx y xy=

∂ ∂ ∂ ∂= =∂ ∂ ∂∂

∑ (2)

Where n is the number of input variables, ŷis the network output, and y(1) and y(2) are theoutputs of the first and second hidden layer,

Fig. 3: Framework of the presented method for modelling the subsidence and sensitivity analysis.


Fig. 4: Network inputs and output: (a) water level decline between 1969 and 2003, (b) groundwaterdepth, (c) storage coefficient, (d) hydraulic conductivity, (e) alluvial thickness, (f) frequency of fine-grained sediments in terms of percentage.


basin subsidence was monitored using PS-In-SAR technique by which the annual deforma-tion rate was measured at persistent scattererlocations. The available radar data consistedof 22 descending ENVISAT ASAR imagesof track 149 acquired between 2003 and 2008and 19 ascending ENVISAT ASAR images oftrack 414 spanning between 2004 and 2009.PS-InSAR method was separately applied onboth ascending and descending radar data.The results obtained from PS-InSAR indicatethat the deformation time series contained asignificant linear component on which the in-significant seasonal fluctuations are superim-posed (DEHGHANI et al. 2010, DEHGHANI et al.2009). The PS-InSAR results were then usedto extract the annual deformation rate that isassumed to be constant from 2003 to 2009 asdeformation time series suggested. The es-timated subsidence rates obtained from de-scending and ascending data are slightly dif-ferent due to the different imaging geometriesof both datasets and non pure vertical com-ponents of the deformation system (SAMIEIE-ESFAHANY et al. 2009).The number of PS pixels detected from

ascending and descending data is around400,000 and 800,000, respectively. Due to thelow spatial density of PS pixels, the deforma-tion pattern cannot be easily recognized in thearea. Hence, the presented modelling methodis used to estimate the deformation rate at non-persistent scatterers in order to retrieve the de-formation pattern. About 60% of the detect-ed persistent scatterers are used to train thenetwork while the rest is applied for networkvalidation. The introduction of the inputs andoutput of the network as well as the obtainedresults from the modelling are presented in thenext section.

3 Results

The available hydrogeology information ofthe Tehran basin includes hydraulic con-ductivity, storage coefficient, frequency andthickness of fine-grained sediments, ground-water depth, amount of water level decline be-tween 1969 and 2003 extracted from piezo-metric measurements and alluvial thicknessof the Tehran aquifer system. These proper-

controlling factors in the Tehran basin subsid-ence, the sensitivity analysis was performedusing the trained network. The framework ofthe presented method for subsidence model-ling and sensitivity analysis is illustrated inFig. 3.The first step in the network construction

is to determine the number of layers and theirneurons as well as the activation functions.The most proper network architecture herewas identified in a trial-and-error manner. Asthe network initialization for training is ran-domly done, the obtained results includingthe updated weights and the sensitivity anal-ysis result are slightly different in each run.It should be noted that each run includes acomplete training procedure. The criterion forstopping the training procedure in each run isthe number of iterations. According to Fig. 3 aneural network is fully trained in N indepen-dent runs where N can be arbitrarily set. Thetrained network is tested at the end of each runusing the validation data. Moreover, the sensi-tivity analysis result obtained in each run willbe saved for further integration. The updatedweights corresponding to the minimum over-all network error are used for the final simula-tion step.The network input variables are hydroge-

ology factors controlling the subsidence ratewhile the output is the subsidence rate at PSpixels derived from PS-InSAR. The Tehran

Fig. 4: Network inputs and output: (g) thick-ness of fine-grained sediments.


moid activation function was constructed.The hydrogeology information at PS pixelsand their derived deformation rate were usedas the network input and output, respectively.The constructed network was trained using60% of the detected PS pixels. As already not-ed, the network was constructed in N runs (Nis here set to 50) for each of which the trainednetwork was tested using the validation data,i.e. 40% of the PS pixels. After comparing theresults of different runs, the weights corre-sponding to the minimum overall network er-ror are used for final simulation step.The correlation between the trained net-

work output and the observed subsidence ratefor training and validation pixels in both as-cending and descending datasets is illustratedin Fig. 5.

ties were measured at different measurementsites across the study area using various meth-ods including geotechnical testing and geo-physics approaches. The point measurementswere then converted to contour lines using in-terpolation. All the information layers super-imposed as contours on the deformation ratemap of PS pixels are shown in Fig. 4 as alreadyexplained.In order to make the hydrogeology data

compatible with the PS-InSAR-derived sub-sidence data in the neural network, all datawere interpolated into grids with the pixel sizeof 90 m. Therefore, the total number of 320 ×210 pixels for which there are 7 features, i.e.hydrogeology information, covered the studyarea. A network consisting of two hidden lay-ers of 20 and 5 neurons with the tangent sig-

Fig. 5: Correlation between the subsidence rate simulated from the trained network (y axis) andthe observed one derived from the PS-InSAR (x axis) for different datasets: (a-b) training andvalidation data of the descending mode, and (c-d) training and validation data of the ascendingmode.


better validate the performance of the model,the subsidence rate at all PS pixels (trainingand validation pixels) for which the observedsubsidence rate is available from PS-InSAR issimulated by the trained network as shown inFig. 6. In this figure, the observed and simulat-ed subsidence rates for both datasets are illus-trated. Furthermore, the residual map as the

As can be observed in Fig. 5, a high cor-relation exits between the simulated subsid-ence data and the observed one extractedfrom PS-InSAR. The linear regression equa-tion for each dataset presented below its cor-responding plot is significantly close to A=Twhere A and T is the simulated and observeddeformation values, respectively. In order to

Fig. 6: Observed deformation rate obtained from the (a) descending and (b) ascending data; sim-ulated deformation rate using the (c) descending and (d) ascending data; residual map as thedifference between the observed deformation rate and the simulated one using the (e) descendingand (f) ascending data.


is related to their different imaging geometry.The maximum deformation rates estimatedfrom descending and ascending datasets are241 mm/a and 203 mm/a, respectively. Thecoincidence of the spatial pattern of the sub-sidence area with the cultivated area indicatesthat the subsidence in the Tehran basin is dueto groundwater exploitation. The spatial pro-files across the deformation show that the sub-sidence follows a “v” type pattern.The results obtained from the proposed

method were compared to the levelling datacollected by NCC in two periods, 2004 and2005 (ARABI et al. 2008). The available lev-elling data consists of five levelling tracks,SB1, SB2, SB3, SB4 and HQH as illustratedin Fig. 1 in different colours. The deformationrates at levelling stations are estimated fromthe trained neural network and then compared

difference between the observed and modelledsubsidence rate is presented in Figs. 6c and 6f.Accordingly, the most amount of subsidencecould be simulated by the proposed model andthe residual map has a noisy behaviour ratherthan a systematic one. The root-mean-squareerror (RMSE) of the residual map is estimatedas 3.53 mm and 4.58 mm for the descendingand ascending datasets, respectively. The lowvalues of RMSE are an indication of the highperformance for the subsidence modelling.The subsidence rate for all pixels (PS and

non-PS) in the study area was then simulatedby the trained network (Fig. 7).As shown in Fig. 7, the spatial subsidence

pattern including 3 main lobes has been re-trieved. The slight difference observed be-tween the spatial subsidence patterns simu-lated from descending and ascending modes

Fig. 7: Subsidence rate simulated from the proposed method applied on (a) descending datasetand (b) ascending dataset.

Fig. 8: Comparison of subsidence rate inferred from the proposed method (red lines) and levelling(blue lines) across different levelling tracks SB1, SB2, SB3, SB4 and HQH.


pixels for which one factor is indicated as themost effective one (Fig. 9). For example, ac-cording to Fig. 9 in about 3500 pixels the mosteffective factor on the subsidence occurrenceis the hydraulic conductivity (#1). In fact, theeffectiveness of a factor is determined by thenumber of pixels in which that factor is themost effective one. According to Figs. 9a and9b, the subsidence has the most sensitivity tothe hydraulic conductivity, thickness of fine-grained sediments and water level decline.A considerable fact is that the sensitivity

analysis results separately obtained from thedescending and ascending data are highlycomparable. The significance orders extract-ed from both datasets are exactly similar. Thismay indicate the correctness of the generatedmodel.

4 Conclusions

Studying the land subsidence first requiresmeasuring the magnitude of the deformationcaused by this phenomenon. Analysis of thetemporal and spatial behaviour of the defor-mation caused by subsidence allows us to mit-igate its devastating effects and manage wa-ter resources. PS-InSAR is a method recent-ly developed to measure the deformation inan area suffering from temporal and spatial

to the values measured by levelling. The com-parison between levelling measurements andthe vertically-converted subsidence rate esti-mated from the proposed method is shown inFig. 8. The overall RMSE between levellingmeasurements and the vertically-convertedsubsidence rate extracted from the proposedmethod is estimated as 19.8 mm/a.The different deformation rate of both da-

tasets is most likely due to the different timeintervals covered by the radar data (2004 and2008) and levelling measurements (2004 and2005), though the spatial distribution of therates from levelling is consistent with that de-rived from the proposed method.According to the diagram of Fig. 3, in order

to study the effect of each input variables onthe subsidence the sensitivity analysis is per-formed at the end of each run for all pixels.In each run, the sensitivity of the subsidenceto different input variables (7 factors in here)is calculated for each pixel using the updat-ed weights (2). The calculated sensitivities arethen sorted based on their values indicatingthe significance order of the input variables af-fected by the subsidence rate. The maximumvalue is corresponding to the most effectivefactor. Finally, the sensitivity analysis out-comes of all 50 runs are integrated to obtainmore reliable results. The final result is pre-sented as a histogram showing the number of

Fig. 9: Histograms of the sensitivity analysis results obtained from (a) descending data and (b)ascending data. The X-axis indicates the input variables (#1: hydraulic conductivity, #2: storagecoefficient, #3: frequency of fine-grained sediments, #4: thickness of fine-grained sediments, #5:groundwater depth, #6: amount of water level decline and #7: alluvial thickness). Y-axis representsfor the number of pixels.


Acknowledgements

We wish to thank the European Space Agency(ESA) for providing ENVISAT ASAR data.We convey our sincere gratitude to the Geo-logical Survey of Iran for providing the hydro-geological information. The authors wish tothank the Delft University of Technology forall their kind scientific supports for persistentscatterer interferometry.

References

AMELUNG, F., GALLOWAY, D.L., BELL, J.W., ZEBKER,H.A. & LACZNIAK, R.L., 1999: Sensing the upsand downs of Las Vegas – InSAR reveals struc-tural control of land subsidence and aquifer-sys-tem deformation. – Geology 27 (6): 483–486.

ARABI, S., MALEKI, E. & TALEBI, A. 2008: Report ofland subsidence in south-west of Tehran. – Na-tional Cartographic Center of Iran, national sci-entific report.

CROSETTO, M., TSCHERNING, C.C., CRIPPA, B. & CAS-TILLO, M., 2002: Subsidence monitoring usingSAR interferometry: reduction of the atmo-spheric effects using stochastic filtering. – Geo-physical Research Letters 29 (9): 26.1–26.4.

DEHGHANI, M., HOOPER, A., HANSSEN, R., VALADANZOEJ, M.J., SAATCHI, S. & ENTEZAM, I., 2009: Hy-brid conventional and Persistent Scatterers SARInterferometry for Land Subsidence Monitoringin southwest Tehran, Iran. – Fringe Workshop2009, Italy.

DEHGHANI, M., VALADAN ZOUJ, M.J., ENTEZAM, I.,SAATCHI, S. & SHEMSHAKI, A., 2010: Interfero-metric Measurements of Ground Surface Sub-sidence induced by Overexploitation of Ground-water. – Journal of Applied Remote Sensing 37(1): 147–156.

FERRETTI, A., PRATI, C. & ROCCA, F., 2000: Nonlin-ear subsidence rate estimation using PermanentScatterers in Differential SAR Interferometry. –IEEE Transactions on Geoscience & RemoteSensing 38 (5): 2202–2212.

FRUNEAU, B. & SARTI, F., 2000: Detection of groundsubsidence in the city of Paris using radar inter-ferometry: isolation from atmospheric artifactsusing correlation. – Geophysical Research Let-ters 27 (24): 3981–3984.

GALLOWAY, D.L., HUDNUT, K.W., INGEBRITSEN, S.E.,PHILLIPS, S.P., PELTZER, G., ROGEZ, F. & ROSEN,P.A., 1998: Detection of aquifer system compac-tion and land subsidence using interferometricsynthetic aperture radar, Antelope Valley, Mo-

decorrelation. However, this technique is ableto measure the deformation at only PS pix-els, the backscattering of which varies little intime. The subsidence estimation from PS-In-SAR does not show local variations when thespatial density measurement points decreaseswhich is mostly the case due to the occurrenceof the subsidence in the rural areas. It is, onthe other hand, critical to know about the lo-cal variations of the deformation particularlyfor the construction purposes. In this paperwe presented an algorithm based on a neuralnetwork to estimate the subsidence rate in thenon-PS pixels in which we were not able to di-rectly measure the deformation by PS-InSARmethod. Hydrogeology properties of the aqui-fer system were used as input variables of thenetwork while the subsidence rate is taken asthe network output. The PS pixels for whichthe subsidence rate is known from the PS-InSAR measurements were used to train andvalidate the network. The trained network wasthen applied to simulate the subsidence rate atall pixels (PS and non-PS pixels). This leads toretrieve the subsidence spatial pattern in theTehran basin. The coincidence of the spatialpattern of the subsidence area with the cul-tivated area indicates that the subsidence inTehran basin is due to groundwater exploita-tion. The maximum deformation rates esti-mated from descending and ascending data-sets are 241 mm/a and 203 mm/a, respective-ly. The spatial profiles across the deformationarea indicate that the subsidence follows a “v”type pattern. The results were then comparedto the levelling measurements. The compari-son showed that the deformation rates pre-dicted at the levelling stations broadly agreeswith the levelling measurements. The RMSEof 19.8 mm/a is most likely due to the differenttime intervals covered by the radar data (2004and 2008) and levelling measurements (2004and 2005), though the spatial distribution ofthe rates from levelling is consistent with thatderived from the proposed method.The sensitivity of the subsidence to differ-

ent input parameters was finally investigat-ed using the updated weights obtained in thetraining process. We found that the Tehransubsidence is mostly sensitive to the hydraulicconductivity, thickness of fine-grained sedi-ments and water level decline.


SAMIEIE-ESFAHANY, S., HANSSEN, R.F., THIENEN-VIS-SER, K.V. & MUNTENDAM-BOS, A., 2009: On theEffect of horizontal Deformation on InSAR Sub-sidence Estimates. – Fringe Workshop 2009, Ita-ly.

SHEMSHAKI, A., BLOURCHI, M.J. & ANSARI, F.M.,2005: Earth subsidence review at Tehran plain-Shahriar (first report). – Report No:Engeo84-06/02, http://gsi.ir/General/Lang_en/Page_27/GroupId_01-01/TypeId_All/Start_20/Action_ListView/WebsiteId_13/3.html(30.10.2012).

TESAURO, M., BERADINO, P., LANARI, R., SANSOSTI,E., FORNARO, G. & FRANCESCHETTI, G., 2000: Ur-ban subsidence inside the City of Napoli (Italy)observed with synthetic aperture radar interfer-ometry at Campi Flegrei caldera. – Journal ofGeophysical Research 27: 1961–1964.

Addresses of the authors:

Dr. MARYAM DEHGHANI, Assistant Professor, Dept.of Civil and Environmental Engineering, School ofEngineering, Shiraz University, Zand St., Shiraz,Iran, Tel.: +98-711-6133162, Fax: +98-711-6473161,e-mail: [email protected]

Dr. MOHAMMAD JAVAD VALADAN ZOEJ, AssociateProfessor, Faculty of Geodesy and Geomatics Engi-neering, K. N. Toosi University of Technology,Tehran, Iran, Tel.: +98-21-88770218, e-mail: [email protected].

Dr. IMAN ENTEZAM, Engineering Geology Group,Geological Survey of Iran (GSI), Tehran, Iran, Tel.:+98-21-64592228, e-mail: [email protected]

Manuskript eingereicht: Mai 2012Angenommen: September 2012

jave Desert, California. – Water Resources Re-search 34 (10): 2573–2585.

GALLOWAY, D.L. & HOFFMANN, J., 2007: The appli-cation of satellite differential SAR interferome-try-derived ground displacements in hydrogeol-ogy. – Hydrogeology Journal 15: 133–154.

GONZÁLEZ, P.J. & FERNÁNDEZ, I., 2011: Drought-driven transient aquifer compaction imaged us-ing multitemporal satellite radar interferometry.– Geology 39: 551–554.

HASHEM, S., 1992: Sensitivity analysis for feedfor-ward Artificial Neural Networks with differen-tiable activation functions. – 1992 InternationalJoint Conferences on Neural Networks, IEEEPress 1: 419–424, Baltimore, MD, USA.

HAYKIN, S.S., HAYKIN, S. & VAN VEEN, B., 1995:Neural Network: a Comprehensive Foundation,IEEE Version.

HOFFMANN, J., LEAKE, S.A., GALLOWAY, D.L. &WIL-SON, A.M., 2003: MODFLOW-2000 ground-water model – user guide to the subsidence andaquifer-system compaction (SUB) package. –U.S. Geological Survey Open-File Report 03-233.

HOOPER, A., ZEBKER, H., SEGALL, P. & KAMPES, B.,2004: A new method for measuring deformationon volcanoes and other natural terrains using In-SAR persistent scatterers. – Geophysical Re-search Letters 31: 611–615.

KAMPES, B.M., 2005: Displacements Parameter Es-timation using Permanent Scatterer Interferom-etry. – Ph.D. thesis, Delft University of Technol-ogy.

MOTAGH, M., DJAMOUR, Y., WALTER, T.R., WETZEL,H.U., ZSCHAU, J.& ARABI, S., 2006: Land subsid-ence in Mashhad Valley, northeast Iran: resultsfrom InSAR, levelling and GPS. – GeophysicalJournal International 168, doi: 10.1111/j.1365–246X.2006.03246.x.

PELTZER, G., ROSEN, P., ROGEZ, F. & HUDNUT, K.,1998: Poro-elastic rebound along the Landers1992 earthquake surface rupture. – Journal ofGeophysical Research 103: 30.131–30.145.

Documents

Neural Network Modelling of Tehran Land Subsidence Measured by Persistent Scatterer … · 2017-03-08 · Measured by Persistent Scatterer Interferometry MARYAM DEHGHANI, Shiraz,