STATIONARY COVARIANCE MATRICES FOR HYPERSPECTRAL POINT TARGET DETECTION

Segmentation appears to be an attractive preprocessing procedure when performing point target detection in hyperspectral data. Unfortunately, the computation cost is not always worth the improvement. Recent work proposed guidelines to decide if segmentation is worthwhile, but the results were only known after segmentation was performed. In this paper, we study the statistical basis of the covariance matrix used in hyperspectral subpixel target detection and propose a metric applied directly to the original matrix. Based on the degree of non-stationarity of the original matrix, one can predict how worthwhile it is to compute segmentation before analyzing on the image.