Anomaly detection using background prediction in hyperspectral images

Anomaly detections in hyperspectral images are of great importance in a variety of applications such as military reconnaissance, scene classification, and disaster evaluation, etc. Most interferences in anomaly detection are generated by inhomogeneous background in practical applications. Therefore, many researchers resort to Gaussian mixture model by classifying inhomogeneous background into numbers of homogenous regions. However, the model might not perform well as anomalies presented locally. Thus, we propose a new anomaly detection approach by introducing background prediction to suppress the interferences of background. Firstly, a conventional background prediction model named two-dimensional least mean square (TDLMS) filter is applied on each spectral image. Since anomalies have distinct spectral signatures from their surrounding background pixels, they will be suppressed in the background prediction process. Next, a residual image is obtained by subtracting the predicted background from the original hyperspectral image. And the background in original image will be extremely eliminated while the anomalies are well preserved in the residual image. Finally, the conventional RX algorithm is used to detect the anomalies in the residual image. We test the new algorithm via two real hyperspectral subimages. The experimental results show that background prediction can suppress the interferences of background effectively and improve the detection performance of the RX algorithm.