Hyperspectral anomaly detection using ensemble and robust collaborative representation

In this paper, we propose a novel ensemble and robust anomaly detection method based on collaborative representation-based detector. The focused pixels used to estimate the background data are randomly sampled from the image. To soften the outliers’ contributions among the selected pixels, we assign low weights to the outliers by adopting a robust norm regression. Consequently, the estimation result is less sensitive to the presence of outliers, as the experiment results attest. However, the algorithm performance is unstable due to the randomness of pixel sampling. To eliminate the instability and boost the detection performance, an ensemble learning method is employed. We repeat modeling background based on random pixel selection, and the detection result is an ensemble of all batches. We show that in most datasets, the proposed method outperforms the traditional algorithms. Moreover, batch processes for detection boosting secure future advances in performance utilization with parallel computing applied.