Hyperspectral Image Processing for Automatic Target Detection

Abstract

Hyperspectral sensors measure the radiance of objects among a large amount of continuous wavelength bands and generate hyperspectral image (HSI). Dimensionality reduction (DR) is a major issue as to reduce the data volume of HSI and improve the detection accuracy. While common DR method use linear algebra (e.g. PCA), we adopt another 3-dimensional DR method using multilinear algebra (e.g. Tucker decomposition). Since the objects of HSI are hard to be identified by visual assessment sometimes, many anomaly detectors have been designed based on image™s spectral covariance information. We implement, refine and apply three anomaly detection algorithms, which are, RX detector, uniform target detector and nested spatial window-based detector, to our two data sets. We compare the detection accuracy by using those algorithms and find that RX detector yields better results in general. We further find that anomaly detection on HSI after DR by Tucker decomposition yields highest accuracy.