A weight SAE based hyperspectral image anomaly targets detection

Due to less limitation, hyperspectral image (HSI) anomaly detection (AD) is widely applied in agriculture, environment and military applications. However, the assumption requirement of special distribution in background makes traditional HSI AD acts poor when the assumption cannot fit in the real HSI. On the other hand, the model of local anomaly detection methods is susceptible to its neighbor anomaly pixels. In this work, we propose a weight sparse auto-encode (SAE) based anomaly targets detection method which combines the weight of neighbor pixel with distance discriminate algorithm. With the ability of high level feature learning in unsupervised way, a sparse code of HSI can be given by SAE for anomaly detection. This method can improve the accuracy of HSI anomaly detection by reducing the risk of anomaly contaminating through allocating different contribution to local pixels. Experimental results based on San Diego airport HSI dataset show that the performance can be ameliorated by the proposed method.