Segmentation Approach and Comparison to Hyperspectral Object Detection Algorithms

Abstract

This study applies a technique from multi-spectral image classification to object detection in hyperspectral imagery. Reducing the decision surface around the object spectral signature helps extract objects from backgrounds. The object search is achieved through computation of the Mahalanobis distance between the average object spectral signature and the test pixel spectrum, a whitened Euclidean distance (WED). This restricted object search (WED), the Adaptive Cosine Estimator (ACE), and the matched filter (MF) were applied to independent data sets, specifically to visible/near IR data collected from Aberdeen, MD and Yuma, Arizona. The robustness of this approach to object detection was tested by inserting object signatures taken directly from the scene and from statistically transformed object signatures from one time to another. This study found a substantial reduction in the number of false alarms (1 to 2 orders of magnitude) using WED and ACE relative to MF for the two independent data collects. No additional parameters are needed for WED. No spatial filtering is used in this study. No degradation in object detection is observed upon inserting the covariance matrix for the entire image into the Mahalanobis metric relative to using covariance matrix taken from the object.