Ensemble and Random Collaborative Representation-Based Anomaly Detector for Hyperspectral Imagery

TL;DR

This paper introduces ERCRD, a novel ensemble and random sampling approach for hyperspectral anomaly detection that improves accuracy and reduces computational costs compared to existing methods.

Contribution

The paper proposes ERCRD, combining random sub-sampling and ensemble learning to enhance detection accuracy, stability, and efficiency in hyperspectral anomaly detection.

Findings

ERCRD outperforms ten state-of-the-art methods in accuracy.

ERCRD significantly reduces computational complexity.

ERCRD demonstrates strong generalization across datasets.

Abstract

In recent years, hyperspectral anomaly detection (HAD) has become an active topic and plays a significant role in military and civilian fields. As a classic HAD method, the collaboration representation-based detector (CRD) has attracted extensive attention and in-depth research. Despite the good performance of the CRD method, its computational cost mainly arising from the sliding dual window strategy is too high for wide applications. Moreover, it takes multiple repeated tests to determine the size of the dual window, which needs to be reset once the dataset changes and cannot be identified in advance with prior knowledge. To alleviate this problem, we proposed a novel ensemble and random collaborative representation-based detector (ERCRD) for HAD, which comprises two closely related stages. Firstly, we process the random sub-sampling on CRD (RCRD) to gain several detection results instead of the sliding dual window strategy, which significantly reduces the computational complexity and makes it more feasible in practical applications. Secondly, ensemble learning is employed to refine the multiple results of RCRD, which act as various "experts" providing abundant complementary information to better target different anomalies. Such two stages form an organic and theoretical detector, which can not only improve the accuracy and stability of HAD methods but also enhance its generalization ability. Experiments on four real hyperspectral datasets exhibit the accuracy and efficiency of this proposed ERCRD method compared with ten state-of-the-art HAD methods.