Detecting anomalous objects in hyperspectral data using segmentation

In this research, our goal is to identify anomalous targets in hyperspectral and multispectral images; our sole starting information is that the targets are larger then a single pixel and their spectral signatures are different than the background. The algorithm is executed as follows: first, we use the Principal Component Analysis (PCA) transformation to find the Projecting the data into this subspace, we create a two-dimensional histogram. From the peaks of this histogram, a set of segments is determined. In comparison, we can alternatively segment our image using the well-known Kmeans approach. We then define the larger clusters to be background segments; each pixel in the image is then given a value based on the minimum "distance" from one of the segment averages. We use three different distance measures: the Mahalanobis distance, the Euclidian distance and the Spectral Angle Mapper (SAM). These dissimilarity measures are used to evaluate the pixels that are extremely different from all of the background clusters. The "anomalies" can be found by thresholding the results. We present the results of a field test using this algorithm; we have succeeded in reaching high detection rates while keeping a very low false alarm rate.