TARGET DETECTION AND ENHANCEMENT IN HYPERSPECTRAL DATA USING SUPER RESOLUTION MAPPING

Abstract

Hyperspectral data, with its enhanced capabilities in terms of data capture in greater number of distinctive bands, has attracted the researchers over years to utilize and further analyze it for achieving the objectives of varied application areas. One such application area is target detection and target identification. The utilization of hyperspectral data in this field of application is still under active investigation. Several issues need to be addressed while performing detection of targets to achieve accurate results. It has been reported in literature that owing to the huge amount of data associated with the hyperspectral data, it needs to be reduced to a subset of useful bands corresponding the application under study. Dimensionality reduction has been one such method for the selection of the appropriate number of bands, however this needs to be carefully performed to ensure that there is no loss of target information. Also, due to the low spatial resolution of many sensors, there is another issue of mixed pixels, that needs to be addressed. Mixed pixels arise when more than one component jointly occupy a single pixel thereby complicating the detection of such mixed pixel targets. The problem aggravates due to the increased spectral variability which may arise due to a variety of reasons. Several techniques to handle the mixed pixels and the difficulties that arise in detection of such mixed pixel targets have been reported in literature. Spectral unmixing is one such method which aids the extraction of pure end members in the image and their corresponding abundance fractions in each pixel. However, this process needs careful handling as errors may occur in the generation of such end members and their abundance fractions. Besides, some literature is available to suggest that spectral index-based approach may also be considered for extraction of end members. The spectral unmixing methods provide the abundance fractions, but the exact spatial location of end members within the pixels remains unknown. In case of applications involving detection of targets such as tanks, aircrafts for vi supporting the military applications, there is a requirement of proper identification of the target shape along with the detection of target location. Super resolution techniques have been proposed to optimize the abundance fractions to generate a subpixel map of the target but its implementation on hyperspectral data for recovery and identification of target is still limited. Moreover, the techniques are mostly based on random allocation and recursive optimizations that further complicate the process and add to the inaccuracy in detection and target identification. Therefore, in view of these limitations, the objectives of this research which have been explored are as follows: a) To review various existing Dimension reduction algorithms with a view to achieve maximum possible Dimensionality reduction while ensuring minimal/no loss of the target data. b) To efficiently characterize the target and background spectral signatures for subpixel detection. c) To study various end member extraction techniques in available datasets using spectral unmixing and spectral indices-based approaches. d) Comparative assessment of random recursive technique with non-random nonrecursive super resolution mapping technique, for subpixel target detection and enhancement. In the first objective, a study of various combinations of dimensionality reduction (DR) techniques combined with full pixel and subpixel target detection (TD) algorithms has been performed to analyze the loss of target pixels in each case. Therefore, the tasks have been subdivided into a) Full pixel Target Detection with and without Dimensionality Reduction vii b) Full pixel Target Detection with and without Dimensionality Reduction, along with the study of impact of Background characterization and, c) Subpixel Target Detection with and without Dimensionality Reduction, along with the study of impact of Background characterization For achieving the tasks mentioned above, two different sets of hyperspectral datasets have been explored. From the experiments it has been observed that in the case of full pixel targets, both dimensionality reduction and target detection result in the loss of target information, however, there is a greater loss of target information in the case when dimensionality reduction precedes target detection in comparison to a case where target detection is applied without dimensionality reduction. Background characterization appears to aid in improvement of full pixel target detection, and K-means is seen to provide better results of detection. In the case of subpixel target detection, however, there appears to be loss of subpixel target information in the case where detection alone is performed in comparison to a case where dimensionality reduction precedes target detection. In the second objective, the focus is on target and background subspaces, and how these subspaces aid or inhibit the process of detection of full pixel and mixed pixel targets has been discussed. The tasks outlined in this objective are: a) Detection of low probability full pixel / subpixel targets with known spectral signatures, b) Detection of targets using background and target subspaces, c) Analysis of the impact of various combinations of background subspaces on full pixel and subpixel target detection, and, d) Analysis of the impact of illumination conditions on the targets. Various conclusions have been drawn from the results obtained. In case of detection using target and background subspaces, it may be concluded that for any given algorithm, if viii the algorithm is performing well for detection of target pixels, the different background subspaces appear to have only marginal impact. Also, in case of analysis of target detection algorithms, Matched Filter is observed to perform better than other algorithms considered. It has been observed that illumination affects the detection of targets immensely, where target pixels of targets in full illumination placed over Gravel roads and Grass are detected well in comparison to those under trees. Also, the targets in partial shade are detected whereas the targets in full shade are not detected by the above discussed algorithms for any combination of background subspace. After an analysis of the spectral profiles, it has been concluded that the surrounding vegetation (trees) has a greater impact on the spectral behaviour of pixel containing blue felt target. The gravel roads have minimum impact on spectral variation of this target out of the three background types. This appears to indicate that spectral contrast / similarity between the target and the background has a significant role in its detection. In the third objective, the end member extraction in mixed pixels using Spectral Unmixing and Spectral Indices based approaches has been performed. Accordingly, there are three major tasks that have been explored / studied in this chapter. a) End member extraction and abundance estimation using spectral unmixing, b) Evaluating spectral indices for end member identification, and In case of spectral umixing, it has been observed that ICA-EEA and Nfindr perform best while recovering the image end members including the target spectra, followed by ATGP, while PPI performed poorly in this case. Also Spectral indices provide a suitable way to extract / generate end members. The fourth objective further extends the study of extraction of full pixel and subpixel targets in hyperspectral data by optimizing the spatial distribution of subpixels inside any given pixel based on the available abundance fractions. The super resolution has been studied as a means to suggest the most optimized spatial distribution/ arrangement of these subpixels ix belonging to different end members/ components inside the mixed pixel. Random recursive pixel swapping method and the non – random non - recursive inverse Euclidean distance method have been studied on synthetic data sets and analyzed in terms of computation time and accuracy of detections. It has been observed that inverse Euclidean distance-based method (IED) (for binary class problems) certainly performs better than the PS (Pixel swap) in terms of accuracy for both the synthetic and hyperspectral data sets. In this research, certain major issues, problems and gaps in target detection and enhancement such as size of targets, spectral variability, problem of mixed pixels, advantage of non-random non-recursive super resolution algorithms etc. have been explored. Some of the major findings / contributions are briefly mentioned here. Both dimensionality reduction (DR) and target detection (TD) lead to a loss of target information, however there is a greater loss if information when DR precedes TD. However, in the case of subpixel targets, there appears to be a loss of subpixel target information when TD alone is performed. Similarly, it has been found that non-random non-recursive algorithms definitely perform better than random recursive algorithms both in synthetic as well as hyperspectral data. While several minor / major issues in target detection have been explored in this research, there still remain many more issues in target detection and enhancement that need to be explored in future particularly using hyperspectral data.