Hyperspectral Anomaly Detection Based on Improved RX with CNN Framework

Hyperspectral anomaly detection aims to separate the abnormal pixels from the background, and plays an important role in remote sensing image processing. Most traditional detectors are based on the RX method but in these detection methods, the detection peak value of outliers is low, which leads to the unobvious detection results. In the small target and sub-pixel target areas, the detection of outliers is difficult because of the complicated background. In this paper, a hyperspectral anomaly detection based on improved RX with CNN framework is proposed. The method firstly estimates the degree of similarity between the pixel to be detected and the target or background by training the CNN, thereby effectively suppressing the background and highlighting the targets. Then the RX algorithm is improved by using higher moments to improve the peak of the targets, and the score is sent to the improved RX algorithm to obtain the detection results. Experimental results on real hyperspectral images show that the method can effectively highlight the target, suppress the background, and have better performance for the detection of small targets.