Answers are grounded in this paper and cite their sources.
Hyperspectral Target Detection (HTD) aims to identify target locations in hyperspectral image (HSI) using limited prior target spectra. Existing methods often use contrastive learning (CL) to construct target and background sample sets from unlabeled HSI and compare their similarity in feature space to enhance background-target separability. However, they often fail to ensure high-purity sample sets, limiting their ability to effectively separate target and background features. Therefore, we propose a spatial-spectral background-target separation network (SBSNet). SBSNet leverages prior target spectra to construct high-purity target and background sets from unlabeled HSI, and integrates them into a multi-scale spatial-spectral feature learning framework to optimize the feature space for more discriminative target detection. Specifically, the primary contributions of this paper are threefold. First, we propose a local spatial-spectral feature fusion module (LSFFM) to extract spatial-spectral feature from the raw HSI, and proposes a spatial-spectral pseudo-label purification strategy (SPPS) to obtain pure target and background pixel sets from unlabeled HSI. Additionally, we introduce the pseudo-label map as prior information to supervise the training process. Secondly, we design a highly robust multi-scale spatial-spectral autoencoder (MSSAE) specifically for HTD, which is used for sample generation during the data preparation and for feature extraction during the training. Thirdly, we propose a clustered adaptive focus training strategy (CAFTS), which synergistically optimizes the feature space through clustered sampling and adaptive exponential weighted loss (AEW-Loss). Finally, experimental results demonstrate that the proposed SBSNet achieves superior detection performance on five public HSI datasets in various scenarios, compared with state-of-the-art HTD methods.
No takes yet. Share an insight, caveat, or question.
Xiang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: