TARGET DETECTION USING OPTICAL AND MICROWAVE REMOTE SENSING

Abstract

Remote sensing continues to attract military strategists globally due to its vast operational and intelligence potential. For example, microwave remote sensing has the potential for subsurface landmine detection particularly along western borders of the country which are covered with vast deserts mostly bereft of any vegetation. Similarly, the optical remote sensing using hyperspectral data may be useful in tracking enemy targets along both our eastern and western borders. These two issues have been investigated in the present research. Landmines are regular shaped small explosive devices with little or no metallic content, buried at shallow depths and designed to kill/incapacitate unsuspecting enemy. At shallow depths at which landmines are buried, microwave frequencies in X-band (10 GHz frequency and 3 cm wavelength) may have both adequate penetration and resolution for landmine detection, but the backscattered image from these shallow buried landmines is severely cluttered and mine feature extraction continues to be the main problem in landmine detection. Several signal and image processing models to solve the problem either have functional limitations or result in high false alarms. Similarly, soft computing techniques reported in the literature such as artificial neural networks (ANN), fuzzy algorithms and genetic algorithms (GA) have a requirement for a priori training data, which may be difficult to obtain in the case of actual minefields/landmines. Besides soil clutter, another factor affecting the backscatter is surface roughness. Surface roughness results in loss of backscatter from the mine and needs to be minimized /eliminated to enhance landmine detection. The studies in the past have demonstrated that the data in multiple polarizations (e.g horizontal-horizontal; HH and vertical-vertical; W polarization data) have more discriminating features than that is available in any single polarization data. Algebraic transforms of multipolarisation data may be useful in minimizing surface roughness effects and therefore needs to be investigated. Similarly, independent component analysis (ICA), an emerging signal analysis technique can also be exploited for surface roughness minimization by treating data in each polarization as a mixed signal. The hyperspectral sensors which collect data in hundreds of narrow contiguous spectral bands arecapable of resolving small surface targets in anymixed background but also suffer from data redundancy. Feature reduction using traditional techniques such as principal component analysis (PCA) therefore becomes an essential pre-requisite. PCA however achieves feature reduction byretaining the top few components contributing the maximum variance and rejecting the balance components. Many military targets are very small in size and therefore PCA based feature reduction may miss the target. ICA, which n yields independent components, may therefore serve as an effective feature reduction method for hyperspectral data for target detection applications. The target detection algorithms developed for hyperspectral data can be classified as spectral matching algorithms or anomaly detectors. Spectral matching algorithms have an essential requirement ofapriori availability of target spectra while anomaly detectors exploit the spectral difference between the target and the background. But since most materials in nature exhibit spectral variability (i.e, their spectra is not deterministic), the performance of these algorithms may lead to high false alarms. ICA therefore not only provides effective feature reduction but also an alternative for target detection in hyperspectral data.. Further, target detection algorithms treat full pixel and subpixel targets separately while in practice targets are likely to occupy some of the pixels fully and some others only partially. Therefore, an integrated approach needs to be developed that combines both the full pixel target detection as well as subpixel target detection. The main objective of this research is to investigate the methods and processes for subsurface and surface target detection using microwave and hyperspectral remote sensing data respectively.Experiments have been conducted under smooth surface conditions to investigate landmine detection and estimation of its depth and shape in microwave X-band (10 GHz frequency and 3 cm wavelength) region followed by landmine detection using various transforms (both algebraic and ICA) of multipolarisation X- band dataunderrough surface conditions. A new approach by fusing image analysis, signal analysis and soft computing techniques has been evolved and investigated for landmine detection which does not have requirement of any a priori training or testing data. For surface target detection using hyperspectral data, a in comparative assessment of full pixel target detection has been carried out between ICA and some commonly reported spectral matching algorithms followed by an investigation of subpixel target detection and enhancement using super resolution. For subsurface landmine detection, an experimental setup consisting of a monostatic scatterometer in X-band region was indigenously designed to generate backscatter intensities under controlled conditions in the laboratory using dummy landmines (without explosives) and live landmines with explosives but less fuzes. The data was generated at different depths upto 10 cm under smooth and rough surface conditions and in both HH and VV polarisations. A model with fusion of image processing and electromagnetic techniques was evolved and implemented to estimate depth and shape of the landmine from a suspected region containing the landmine. The main advantage of the proposed model is that it does not have any requirement of separate training and test data set to train the optimizer and validate the results. A new local window basedpreprocessing has been proposed and the results indicate that landmine detection for mines buried in dry sand is possible using the proposed model. A detection figure test has also been proposed for reducing the false alarms. The results indicate that a detection figure in the range of 40-80 reflects correct detections. Analysis of various error plots for estimation of depth indicated that the maximum error in predicted depthup to 7 cm does not exceed 20%of the correct depth andshape may be recovered with 95 %accuracy using waveform based ANN methods. Several algebraic transforms including image differencing, image ratioing, polarization discriminant ratio (PDR) and a new transform proposed as the difference of sum and the difference of data in HH and W polarizations ((HH+VV)- (HH-VV)) have been investigated for surface roughness minimisation. Detection of IV landmines has also been carried out using ICA of microwave data in two polarizations, which can also minimize effect ofsurface roughness. The effect of surface roughness has been assessed quantitatively using Shanon's entropy. For surface target detection, experiments have been conducted using synthetic and hyperspectral data acquired by AVIRIS sensor. Two issues have therefore been investigated. First, whether PCA based feature reduction in hyperspectral data leads to a loss of small targets and second, three metrics, namely skewness, kurtosis, entropy and gradient of entropy, have been investigated for their suitability to indicate ICA components containing the targets. The results show that there is neither any loss of target inPCA based feature reduction nor does there appear any significant advantage of using ICA over PCA in terms of number of components containing the targets. Further, entropy appears to be the most appropriate metric for identification of ICA components containing the targets. Further, a ROC (receiver operating charecteristcis) analysis has been carried out for a comparative assessment between ICA and four spectral matching algorithms (namely spectral correlation mapper (SCM), spectral angle mapper (SAM), orthogonal subspace projection (OSP) and constrained energy minimization (CEM)) and two anomaly detection algorithms (namelyorthogonal subspace projection anomaly detector(OSPAD) and Rx anomaly detector (RXD)). The ROC plot ranks ICA significantly higher than the other methods, thus indicating the usefulness of the ICA for accurate detection of surface targets in hyperspectral data. In order to develop an integrated approach to full pixel and subpixel target detection, the ICA components identified in the previous experiments have been used. A region surrounding the detected full pixels target was segmented and assuming a total of ten endmembers, abundance fraction for all the pixels in this region was estimated using an unsupervised spectral unmixing method. These abundance fractions were then exploited for subpixel target detection and enhancement using super resolution. For super resolution, two algorithms, pixel swapping algorithm (Atkinson, 2005) and a newly proposed inverse euclidean distance method have been explored. Super resolution has been carried out at 3, 5, 1,9, 11 scale factors. The classification accuracy for pixel swapping method was found to be 80% and above but only for targets with simple geometric shapes. The method failed for AVIRIS data containing targets of complex shapes. On the other hand, the classification accuracy achieved for the inverse Euclidean distance method was found to be greater than 70% but the method performed equally well even for theAVIRIS data. Further, the CPUtime for processing was found to be up to ten times less for the inverse euclidean distance method at a scale factor of 11. The research leads to a model by fusing image analysis, signal analysis and soft computing techniques for the buried landmine detection in microwave X-band region, estimation of its depth and shape. A detection figure test has been proposed and a range of detection figures have been proposed for making correct detections. Minimisation of surface roughness effects has been achieved both using a newly proposed algebraic transform of mutipolarisation data (HH and VV polarizations) and ICA. In the case of surface target detection from hyperspectral data, an integrated approach for target detection by combining ICA based feature reduction, ICA based full pixel target detection, estimation of fractions using spectral unmixing and super-resolution mapping using a new inverse euclidean distance algorithm have been proposed. vi