Anomaly detection method for hyperspectral imagery based on locally linear fitting

In this paper, a locally linear fitting based method is proposed for anomaly detection in hyperspectral imagery. The procedure contains four steps. Firstly, locally linear fitting is used to distinguish the isolated anomaly by the residue. To mark the background pixels in the scene, pixels are clustered in the second step. Due to different spectral signatures, the anomalous pixels and mixed background pixels will fail to cluster. Then, the envelop curve is obtained for marking the mixed background pixels. Anomaly targets lie in the remaining pixels. By these steps, the pixels are divided into the potential anomaly part and the background part. At last, fitted by the surrounding background pixels, potential anomaly pixels with residue larger than the given threshold are confirmed as the final anomaly. Experiment results indicate that the proposed method is efficient and outperforms Reed-Xiaoli detector and collaborative representation based detector.