Anomaly Detection in Hyperspectral Imagery: An Overview

Problems

The magnitude of anomalies motivated researches in anomaly detection and interpretation for hyperspectral images. Since the beginning, researchers had to cope with the problem of the absence of any prior knowledge about the treated data. Therefore, they try to use statistical methods to compare between the Pixel Under Test (PUT) and the background. For statistical methods the background is modeled with a linear distribution of the Probability Density Function (PDF) that supposes its homogeneity. This supposition accentuates the False Alarm Rate (FAR) especially for high resolution images where the supposition of homogeneity seems to be inappropriate since the big diversity of existing materials. To decrease this fact, non linear models of the background have been proposed with the kernel based anomaly detectors. Other researches try to solve the anomaly detection problem with different techniques as feature selection and the segmentation.

Whatsoever the underlying techniques are, there are three principal challenges to overcome. The first challenge concerns increasing the detection rate while decreasing the false alarm rate which is related to the fact that the presence of noise, the contamination of the background statistics with the signature of the anomaly and the supposition of background homogeneity increase considerably the false alarm rate. The second challenge is related to the detection of anomaly with different shapes and sizes. In fact the size of anomalies can range from sub-pixel level to few pixels and the detection of these different sizes anomalies with the same detector steels a big challenge. The third challenge aims to achieve the nearest computational cost to real-time processing to perform anomaly detection on board as pixels are received.