Hyperspectral Target Detection Based on a Background-Aware Sparse Transformer Network

Hyperspectral target detection (HTD) relies on prior target spectra to locate the targets of interest within hyperspectral images (HSIs). Recently, deep learning methods have shown their potential in hyperspectral feature extraction and multiscale feature fusion. In this article, we propose a background-aware sparse transformer network (BASTNet) for HTD, to solve the problems of target sample imbalance and underutilization of global information. First, the proposed method utilizes random masking and target spectra generation strategies to establish an image-level training paradigm, constructing sufficient and balanced training samples to prompt the network to learn spatial-contextual features between the target and the background. We then introduce a Siamese sparse transformer network (S²TNet) with an encoder–decoder structure to achieve fast inference for large-scene hyperspectral imagery. Specifically, S²TNet consists of pyramid feature extraction, multiscale feature fusion, and a target detector, with a sparse self-attention mechanism enhancing the focus on target regions and improving the separability between target and background. Furthermore, a background-aware learning mechanism is introduced that uses a foreground and background guidance loss that attenuates the interference of background noise on the target detection. Experiments on five benchmark datasets demonstrate the superiority and applicability of the proposed BASTNet method, showing that it outperforms the current state-of-the-art (SOTA) HTD methods.