Wavelet based statistical signal processing for the detection of subpixel hyperspectral targets

The detection of subpixel target in hyperspectral imagery requires the unmixing of the constituent elements using a linear or non-linear separation model. When the unmixing. approach is not feasible, subpixel targets are detected through feature extraction and classification. A wavelet-based hidden Markov model (HMM) approach for the detection of subpixel hyperspectral targets is presented in this work. A performance comparison with the maximum likelihood (ML) model, for the detection of low amplitude targets or constituent hands in hyperspectral curves, is then made. This study includes 1,000 hyperspectral curves with half of them containing subpixel targets or additive Gaussian absorption bands. Test results show that the ML model outperformed the HMM in terms of overrall accuracy. However,er, when the amplitude of constituent bands is very weak, about 1 % of the background signals, the sensitivity of the system using the HMM is better than the ML model. For the system using the RMM approach, its sensitivity is least affected with changes in the target amplitude, and its overall accuracy is least affected by changes in mother wavelets.