Hyperspectral anomaly detection via density peak clustering

Abstract

In the last few years, a density peak clustering algorithm (DP) has demonstrated its advantages in hyperspectral data analysis and processing. In this letter, we take the benefits of the DP algorithm to the hyperspectral anomaly detection, to circumvent two negative aspects which affect the detection performance: The untenable supposition of the Gaussian distribution and the contamination of the background statistics caused by anomalies. Specifically, the proposed DP-based hyperspectral anomaly detection method is implemented as follows: A hyperspectral image (HSI) is first divided into local windows to address computationally expensive density computations. In each local window, the DP is performed to calculate the density of each pixel. Finally, we detect anomalies using the obtained density map, based on that anomalies are generally with low probability of existence in the image and thus have low densities. Experimental results obtained on four real hyperspectral datasets demonstrate that the detection performance of the proposed method is superior to some widely used anomaly detection methods.