A Weighted Collaborated Representation Anomaly Detector for Hyperspectral Image

Abstract

Hyperspectral image can provide valuable information for resource development, environment protection, national defense construction, etc. As one key technique for processing hyperspectral image, the anomaly detection can discover the region of interest, which is the direct information for the following analysis. However, how to identify the anomalous target or zone contained in the hyperspectral image is always a difficult problem. Especially, the accuracy of anomaly detection is usually low. Some useful information may be missed, and the following action cannot be implemented in time. To address this issue, this article proposes a novel anomaly detection method for processing hyperspectral image. Firstly, the collaborated representation-based anomaly detector (CRD) for hyperspectral image is studied. The influence factor on the anomaly detection accuracy is determined. Then, a weighted method is proposed to decrease the false alarm of anomaly detection result. To be specific, the weights are generated by the reconstruction errors of the under test hyperspectral image with a stacked autoencoder network. In this way, the performance of the anomaly detection algorithm is enhanced. Two real hyperspectral image datasets are utilized to evaluate the proposed method. Experimental results show that different Lagrange multiplier values have influence on the anomaly detection results. The receiver operating characteristic curve is used as one metric for evaluating anomaly detection results. The proposed method outperforms the collaborated representation AD and the benchmark detector-local Reed Xiaoli AD (RXD). The proposed method provides a novel anomaly detection strategy for the practical application of hyperspectral image.