Long-tail distribution based multiscale-multiband autoregressive detection for hyperspectral imagery

摘要

There are often demands for region target detection such as air pollution detection and oil spill monitoring, even though small target detection has gained much attention in the field of hyperspectral detection. In this paper, we present a long-tail distribution based multiscale-multivariate autoregressive hyperspectral detector to handle such region targets. We establish a multiscale-multiband autoregresive model (MMAM) to characterize the interband, the spatial and the band-spatial correlation in hyperspectral data simultaneously and have the corresponding multiscale-multiband likelihood ratio (MMLR) test. Due to the long tail property of MMAM noise, we treat the statistical characteristics of MMAM noise as multivariate t distribution. Then, alternating projection involving fixed-point iteration and gradient based searching (APFPGS) are utilized to fit this statistical distribution. Experimental results on the real hyperspectral imagery recorded with A series of Environmental Probe System (EPS-A) show that our approach has better performance in hyperspectral region target detection than the other four detectors.