Target Detection Improvements in Hyperspectral Images by Adjusting Band Weights and Identifying end-members in Feature Space Clusters

          Spectral target detection could be regarded as one of the strategic applications of hyperspectral data analysis. The presence of targets in an area smaller than a pixel’s ground coverage has led to the development of spectral un-mixing methods to detect these types of targets. Usually, in the spectral un-mixing algorithms, the similar weights have been assumed for spectral bands. However, the various uncertainties such as the different effects of the atmospheric conditions and the relative radiometric calibration of the sensor lead to differentiations data recorded in each band. So, the Modification of the weights of the spectral bands is the first objective of this paper in order to improve the accuracy of target detection in the spectral un-mixing process. Considering the complexities of direct estimation of the band weights, an algorithm based on the Variance Component Estimation (VCE) is proposed to optimize the weights of the spectral bands. On the other hand, in addition to the availability of target spectrums, the spectral response of the backgrounds is a necessity to perform reliable target detection. The unsupervised detection of the background endmembers is known as the popular way of doing that. The second contribution of this paper is the proposal of cluster-based background detection to be used in the target detection process. It prevents the presence of the unrelated endmembers in each cluster which has improved the spectral un-mixing for target detection. The proposed methods have been implemented in the target detectors of Unconstrained Linear Spectral Un-mixing (UCLSU), Sum to one Constrained Linear Spectral Un-mixing (SCLSU), Non-negativity Constrained Linear Spectral Un-mixing (NCLSU), and Fully Constrained Linear Spectral Un-mixing (FCLSU). The results indicate their success in the improvement of the target detection accuracies. Considering the best choice on the number of spectral clusters and the number of background endmembers, accuracy improvement of up to 17 percent in the target detection has occurred.