depth2

Algorithm Overview Based On Image Processing with Electromagnetic(EM) Techniques in X Band and GA Approach for Depth Estimation of

Shallow Burried Dummy MinesKC Tiwari*, D. Singhb@ and M. Arora*

*Department of Civil Engineering ( Geomatics), IIT Roorkee, Roorkee - 247667, India@Department of Electronics and Computer Engineering, IIT Roorkee, Roorkee - 247667, India

Email: chtphd mail.com, armfcajitr. rnetinandmanocenaiitrernetin

Abstract Radar remote sensing is a very encouraging fieldof research with many advantages over various otherDetetio ofurid lndmies nd stimtio of techniques. The conditions existing in the western borders

depth by modeling layered media is a complex and of India match that of sandy7 deserts where surfacecomputationally intensive task. Microwave remote ..sensing withthecapbiltyopnetatesubroughness conditions are smooth and the area is alsosensing with the capability to penetrate subsurface and itS brf fayvgtto raymjrln etrs hs

ability to resolve landmines as well as non-lethal targets conditiofn make them eremey sable foraplcTioncan therefore be used for subsurface landmine detection. cofdar rmoesensingrfordetrectio of minefields.A model based on electromagnetic scattering and image Various signal deimg ocesin lgrhanalysis techniques at X-band frequency for sub-surface are under development world over to accurately solve thedetection ofdmm dandmines bed uroaosad layer mine detection problem. Literature reports neural network

paper. An extensive set of lab experiments have been [3], SVM [4] and Markov random models [5] basedpaprier Ant extsivesetmoflabdminesexperim tent s ivesb techniques applied for solving the detection problem. In acarried out using dummy landmines (without explosives) cmaaieeprmn utatv uz ehiu aand backscatter observed at different depths. The raw data cmaaieeprmn utatv uz ehiu a

andbackscatterobservedatderiesoffer gendepthces.iThe ra a found to perform better when the training and testing dataiS processed through a series of image processing steps, esaesprt 4.Sm inlpoesn ehdthe detection is carried out using Otsu's thresholding and . .. . . ~~~~~search for the hyperbola for charecterising the landminethe depth estimated by optimizing through a GA (genetic s

algorithm) based electromagnetic model developed. The response to predict landmine detection, its geometry anddepth [5]. Most works have however ignoredmethod does not have any requirement of separate dph [] otwrshv oee goe

mthodnn andotesnt hatavsetany trequire oftimszeparat electromagnetic interactions at the air-ground-landminetraining and test data set to train the optimizer and inefc. Dail eta aeeautd hoeia oevalidte te reults Theresuls uner lboraory nterface. Daniels et al have evaluated a theoretical modelvalidate the results. The results under laboratory using microwave remote sensing in P- band (441 MHz,

conditionusindicate thatsholdetion fchnidummy lanmins ia 68 cm) for buried reflectors at depth H and the resultspossible using thresholding techniques with data hv en wl upre yteeprmnahave been well supported by the experimentalgenerated in X band and the proposed model is capable of investigations conducted in Negev Desert [6]. Modelsestimating depth of the buried landmine to a significant analyzing electromagnetic interactions at layered roughaccuracy. surfaces are suitable for estimation of depth of landmine.Keywords Microwave X band, landmines, Ulaby et al have also proposed a relationship betweenelectromagnetic scattering, detection, feature extraction, radar observation depth as a function of observationentropy, thresholding, image analysis. genetic algorithm. frequency and soil moisture content by considering the

power of an electromagnetic wave incident upon a soilsurface [7]. Despite extensive research however, eitherthe models developed are very complex and

I. INTRODUCTION computationally intensive or there is a requirement ofsome a priori information.

Thousands of innocent civilians are killed/ Lower frequencies penetrate higher, while highermaimed annually due to buried mines scattered all over frequencies resolve better. Microwave X-band at 10 GHZ,the world in which the conflict has long ceased [1,2]. 3 cm with significan subsurface penetration capabilityDetection and removal of landmines however is provides an optimal choice as it also has sufficientcomplicated by variety in types of mines, soil types, bandwidth to permit resolution of mine targets as well asscattering from layered media, vegetation etc. Landmines non-lethal targets such as rocks etc [7]. Besides, at thisare often laid flush with the ground or at shallow depths, wavelength volume scattering from inhomogenities of thehence their responses to any emitted pulse overlaps with layered media ( small rough surfaces such as in sandsoil clutter. Landmine detection poses two main layer) can be ignored because of the dimensions ofchallenges in detecting small shallow buried landmines inhomogeneous particles and the distance between themwhich contain little or no metal i.e. reduction of soil effectand mine feature extraction.

1-4244-0284-0/07/$20.OO ©2007 IEEE 331

is of the order of 2.5x10 -3 , both of which are much The experimental setup consists of a monostaticsmaller than the wavelength [6]. scatterometer with one pyramidal horn antenna connected

The problem of landmine detection and through a circulator to microwave transmitter on one sideestimation of depth using layered media analysis presents and power meter on the other side. A wooden boxa non-linear problem and hence optimizers such as measuring 120 cm by 120 cm has been created and filledgradient descent or direct search methods are not likely to with sand for mines to be buried. The system hasyield good results because standard optimizers require provision to move the antennae in both X and Y planes.objective function to be continuous, differentiable, non- The horizontal bars on the two sides (called X-Ystochastic and linear. A variety of nonparametric direction) were marked serially from 1 to 24 to make aoptimization techniques such as neural networks, support grid. The circulator and the horn were moved laterally (Y-vector machines and fuzzy methods etc have been put to direction) from 1-24 at each of the horizontal (X-use [3, 4, 8] but these non-parametric methods often direction) positions from 1 to 24 thus making a total ofrequire different training and testing data sets to train the 24x 24 positions. Backscatter readings were recorded atoptimizer and validate the results. Thus the results are all the 24x 24 positions.often representative of the test data and not necessarily Backscatter reading has been taken for eachthe ground truth. mine at depths of 0.5 cm, 1.0 cm, 1.5cm, 2.0 cm, 2.5cm,

A new approach to the problem of detection of 3.0 cm, 4.0 cm, 5.0cm, 7.0cm and 10 cm. All thelandmines and estimation of the depth of landmine by experiments were conducted in microwave X band at 10developing a theoretical model by combining image GHz.processing and signal processing approaches has been The dielectric constant of dummy landminesdiscussed in this paper. A theoretical model based on without explosives was assumed in the range of 4-10 andelectromagnetic interactions at the air-sand-landmine that of smooth dry sand in the range of 3-5. The sandinterface has been used for depth analysis and optimized surface was assumed smooth because at this wavelength (using a genetic algorithm based fitness function. The X band, wavelength - 3 cm), the volume inhomogenietiesmain advantage of this method is that it does not require are of the order of 2.5x1 0 -3 cm which are much smallerseparate training and test data set and therefore the results than the wavelength [6].represent the ground truth more accurately.

II. EXPERIMENTAL SETUP III. THEORETICAL MODELING ANDIMPLEMENTATION

A set of experiments have been conducted with Clutter removal and mine feature extraction isdummy mines (without explosives) in dry sand. In each the basic process for detection of landmines. However,category dummy, anti tank, fragmentation, and influence any radar image is literally swamped with clutter. Cluttermines have been used. The schematic layout of the detected by radar includes initial ground reflection andexperimental setup is shown in Fig-I and the dimensions background resulting from several scatterers within theof the landmines used are given in Fig-4. soil. All forms of undesired signals require estimation and

subsequent removal in order to enhance the target signal.Tranrtikr7 A common method for clutter reduction is to simply

compute the mean vector and subtract this value fromindividual pixel reading. This method however fails incase the contour of the ground surface is not smooth. The

li T

|11

4 | | |aim of mine feature extraction is to classify a signal intomine or non-mine features and to make a decisionbetween the two during post processing and extract thefeatures containing the landmine. A series of steps havebeen formulated for extraction of mine features and

L Sa 1/> estimation of depth which are given in Figure -2.Raw data was generated in a grid of 24x24 array

which was calibrated using an aluminium sheet whichhaving conductivity of 3.5 x107 Seimens per meter and

Fig- 1: Schematic Diagram of Monostatic reflectivity coefficient -1 (perfect reflector). TheScatterometer System calibrated data was normalized to put all the data in one

range. The illumination area of the antenna system is notlimited to the pixel size of the image, so there is a

332

significant contribution from the neighboring pixels. specular reflection and diffuse scattering by the air-sandTherefore, a 5x5 convolution filter [11 11 1; 1 1 2 1 1; 1 interface, reflection from the mine buried at depth H,2 3 2 1; 1 1 2 1 1; 1 1 1 1 1 1] was applied and the diffraction and scattering by surface irregularities of theconvoluted data was then assumed to be pure reading per reflector and volume scattering from the sand layer [6].pixel. The total returned field ER ( Figure-3) at the radar

To obtain the suspected area where an object receiver due to the mine buried at depth H is given by -may be buried, it is very important to select a properthreshold value, say 't'. Otsu's thresholding was used for ER:= ES + EC1 + EC2 + EC3 . Ecndetection of landmine [9].

Ei ES ECi EC2 EC3Raiw image Calibrationt

Convolution Filter Normalization Air - Medium 1

z

Thresholding(Otsu's I Maximum Entropy)

Sand -Medium 2Cdentification of T

Suspected area and Detection Figure Test y, Mine - Medium 3

Mlalugkiof erge backsctterISuspected Rgion inormalized intesgity Fig 3: Radar wave propagation in an air-

o|f maskekd aresa soil-mine interface

Electromagnetic Mode ing for Dep0th Daniels et al [6] have derived the followingOptimisation Using GA model for estimation of backscatter in a layered media

which can be used for estimation of depth.

Estimated Detth ER =s\/4k cos1 exp(-±sin 1)* ~~~~~~~~~~~~~R+R exp (-2 ;1HAFig. 2: Flow Chart of Algorithm X 1-2 23 P (-2 H)(

For reduction of false points, a quantity detection + R23exp(-2y2H)figure (D) as under was calculated where, A(FG) is whereaverage reading for foreground pixels, A(BG) is averagereading for background pixels and A(FG+BG) is average k -2 n/k= wave No, c7 roughness parametersreading for all pixels. It was found that if detection figureobtained is less than 40, it could be claimed that there is 01 incident angle, H depthno object buried in sand. R I- R23 2=3

1-2 ~~~~2-3 -_A(FG)-A(BG)x1OO 44 + F2 + _

Detection figure =A(FG + BG) ( , 2 & 3 refer to first, second & third medium i.e. air,

sand and mine)The interaction of an electromagnetic wave with

soil is a complex phenomenon. Figure-3 gives a 2icschematic diagram of electrical field propagation for 72 X= _2subsurface penetration as in the present case. (propagation constant in second

The total returned electrical field received at the medium i.e. sand)scatterometer end ER is the result of simultaneous

333

Genetic algorithms (GA) are robust optimizers, genetic algorithm. The GA tool was run several timesstochastic search methods modeled on the principles and with different initial population and the optimized resultsconcepts of natural selection and evolution. GA is for estimation of the depth are given in Table 1. A grapheffective in solving complex problems particularly when showing variation of calculated/predicted backscatter andthe goal is to find global maxima in a high dimension, the observed backscatter with increasing depth is shownmulti model function domain in a near optimal manner. in Figure-5. The variation is as expected and showsThe optimization in genetic algorithms is done through a degradation in performance with the increase in depth.fitness function (also called cost function or objective The oscillatory behaviour of backscatter with the depth isfunction) which is used to assign a fitness value to each of also noticed. An error plot of the calculated/predictedthe individuals in the GA population. The theoretical depth is given in Figure-6 which confirms optimal results.model based on electromagnetic theory for calculation ofbackscatter (eqn- 1) was used to obtain EC(predicted/calculated backscatter) and the proposed V. CONCLUSIONfitness function optimized using genetic algorithm. Theproposed fitness function is :- Subsurface landmine detection and estimation of

FitnessFunction = min Eo - Ec 2 depth has been carried out successfully in X- band (10GHz, 3 cm) upto a depth of 10 cm. Experimental data was

where Eo is observed backscatter and EC is generated with different dummy mines without explosivespredicted/calculated backscatter .

buried in dry sand at different depths in a laboratoryenvironment. Detection was carried out using Otsu'sthresholding. The results were validated with known

IV. RESULTS AND DISCUSSIONS location of the mines used in the experiments. A newmodel has also been developed and implemented for

The backscatter data collected for different estimation of depth of landmine which was optimizeddepths of the experiments was processed through various using genetic algorithm. However, it was found thatsteps as discussed and plots generated at each stage to optimisation using genetic algorithm was highlyanalyze and detect the presence of landmine features. dependent on choice of initial populations and it affectedPlots generated for raw data, calibrated data and the results significantly. The model does not have anyconvoluted data and mine feature extracted after requirement of training and test data sets. A Detectionthresholding are given in Figure-4. It is noticed that the Figure test was carried out before processing the data forraw data contains severe clutter from various sources depth analysis to assess the presence of any object inparticularly from the corner of boxes and the plots for the thresholded image.raw data is highly random in nature. Convolution with a5x5 kernel filter after calibration results in a smoothened ACKNOWLEDGEMENTimage with noise around the corners severely restricted.The convolution plot for each mine, which in these The authors are thankful to the Defenceexperiments was kept at the centre of the box Research and Development Organisation, Ministry ofsignificantly highlights the likely area containing the Defence, India for providing financial support for themines. The mine features were segmented using Otsu's project.thresholding method. It is noticed that although the REFERENCESthresholded values accurately indicate the location of themine but there apparently is enough clutter/noise in the [1]. Bureau of Political & Military Affairs, "Hiddensame range. This highlights the need of preprocessing the Killers", US Department of StatePublicationlO575,data so as to cause statistical variation between the clutter http:/www.state.gov/www/global/arms/rpt/-9809\-and the mine like feature so that detection becomes demine\-loc.html, September 1998,correctly segments the mine features and eliminates the [2]. D Potin, P Vanheeghe, "An Abrupt Changeclutter/noise. A Detection Figure test was carried out and Detection Algorithm For Buried Landmine Localization",the values obtained in all the cases are found to be more IEEE Transactions on Geoscience And Remote Sensing,than 40 thus indicating presence of an object. Vol 44, No 2. Feb 2006

After masking and extracting the mine features [3]. Carosi S and Cevini G, "An Electromagneticin the region of interest, depth was estimated using the Approach Based On Neural Networks For the GPRmodel proposed. Fitness function formulated as difference Investigation of Buried Cylinders", IEEE Geoscience andof normalised observed backscatter and RemoteSensing Letters, Vol 2, No 1, Jan 2005.calculated/predicted backscatter was minimized using' '

334

Mine / Depth Anti Tank Mine Fragmentation Influence MineDepth 1.5 cm Mine, Depth 1.5 Depth 1.5 cm

cm

Image

Dimensions L a Ln -2

W1 7¢2a_ 20j

Raw Plot

CalibratedPlot

ConvolutedPlot

Otsu'sThresholding

Fig-4 Mine detection using threshholding

[4]. Collins L et al, " A Comparision of the [7]. Ulaby F T, Moore R K and Fung A K, "RadarPerformance of Statistical and Fuzzy Algorithms for Remote Sensing and Surface Scattering EmissionUnexploded Ordnance Detection", IEEE Transactions Theory, Vol II & III", Addison Wesley Publishingon Fuzzy Systems Vol 9 No 1, Feb 2004. Company, 1982.[5]. Gader et al, " Landmine detection with ground [8]. Bermani B et al, " An Innovative Real Timepenetrating radar using hidden Markov models", IEEE Technique for Buried Object Detection", IEEETransactions Geoscience and Remote Sensing", Vol 41, Transactions on Geoscience and Remote Sensing, VolNo 4, Jun 2001. 41 No 4, April 2003[6]. Daniels, et al, Microwave Remote Sensing of [9]. Tian H et al, "Implementing Otsu'sPhysically Buried Objects in Negev Desert Thresholding Process Using Area-Time EfficientImplications for Subsurface Martian Exploration, Logarithmic Approximation Unit", 0-7803-776 1-Journal of Geophysical research, Vol. 108 No.48033, 3/03/$ 17.00 C 2003 IEEE, Mar 2003.2003

_ ~~33

Normalised Observed E Fieldn3..,,~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~.............................0-9 Normalised Predicted E Field

01

0.3'i.. .~~~~~~~~~~~~~~~~~~~~~~~~~~~...

ffi~ ~ ~~~~~~~~~~~~~~~~~~~~~........ ...

01

0 0.01 0.02 003 004 0.05 006 0 0 .06 0.09 0.1

Figure - 5 Depth vs Predicted & Observed Backscatter

TABLE -IActual and Predicted Depth

Dielctric constant of dry Sand = 5 150(From Ulabys graph) and DielectricConstant of Mine = 4 ______

Actual GADepth(H) Optimised % Error

epth

0.005 0.00605 21.00000.01 0.0096 -3.6000

g~~~~~~~ 00.015 10.01668 111.20000.02 0.02361 18.05000.025 -0.02583 3.32000.03 0.0316 5.33330.04 0.03438 -14.0500 0 mu10 1500.05 0.04496 -10.0800 Actual Depth (cmn)0.07 0.0523 -25.28570.1 0.13529 35.2900 Figure -6 Error plot - Actual vs predicted depth

TABLE-I ~ ~ ~ ~ 33