Clustering based background learning for hyperspectral anomaly detection

Abstract

In real hyperspectral images, which have a complex background with multiple classes, separation of anomalous targets from background is a challenging problem. The Clustering based Background Learning (CBL) method is proposed for hyperspectral anomaly detection in this paper to deal with this difficulty. The proposed CBL method identifies the background classes, and so facilitates detection of the anomalies. CBL is built based on this reality that anomalies comprise a limited number of pixels occupying small areas. To achieve anomalies, the spectral and spatial clustering maps are fused through a new method called Multiplicative Fusion (MF). This fusion method leverages both spectral and spatial information and makes the clustering results to support each other for anomalies detection. The synergy done in the clustering phase followed by applying an adaptive support vector machine leads to identification of various background classes. One important advantage of the CBL method is its generalization ability for anomaly detection in various sub regions of a large scene where the trained model for a sub region can be used for anomaly detection of other sub regions of the same scene. Additionally, the pre-processing and the model training steps can be performed on a common hardware without any requirement to expensive hardware for high processing power. The experimental results demonstrate that the CBL method outperforms the competing methods on various types of hyperspectral data with different spectral and spatial resolutions where a stable behaviour can be seen in detection performance of CBL on different images.