Computational Enhancement in Locating Unknown Sub-Pixel Anomalous Sources in Hyperspectral Images

The outcome of this research developed a new qualitative image processing approach to finding subpixel level anomalies in hyperspectral imagery. There were several steps involved in this process and there were 2 phases. First, anomalies were placed in the input hyperspectral image. In Phase 1, we chose a set of random chips or random pixels which were from the N×M×P hyperspectral image. Here the number of rows, columns, and bands of the hyperspectral image corresponds with N, M, and P respectively. A metric involving edge and turbulence is computed for each 3×3 neighborhood of pixels in the image. From this, a function is constructed where the domain (i.e. horizontal axis) is the band number and the range (i.e., vertical axis) is the corresponding “edge divided by turbulence” value after a first thresholding is applied. Throughout the paper, “edge divided by turbulence” is referred to as E/T. A second threshold is applied to this E/T function to determine the optimal bands to select from the hyperspectral image. In Phase 2, a transition into determining the probability of detection and probability of false alarm by adjusting the incongruence threshold H is performed. For this effort, a suite of Matlab tools was developed detailing information related to probability of detection (POD) and false alarm rate (FAR) for varying levels of contamination, incongruence, and band thresholding. A 224-band hyperspectral image was used as the evaluation image for this effort. The result of this effort was a generic workflow for determination of subpixel anomalies in any hyperspectral image.