Prior-guided unmixing transformer for underwater hyperspectral target detection

Target detection in nearshore underwater hyperspectral imagery is challenging due to water-induced degradations and extremely sparse targets. In turbid nearshore environments, existing detectors often fail to separate faint target signatures from dominant background clutter. Hyperspectral unmixing provides an appealing paradigm by modeling each pixel spectrum as a mixture and directly estimating target abundances. In this article, we propose the guided spectral unmixing transformer autoencoder (GSUTA), an end-to-end autoencoder-based unmixing network that outputs abundance maps for detection. To mitigate extreme target sparsity and severe class imbalance, we develop a guided training strategy that synthesizes target-enriched pseudo-pixels via Dirichlet sampling for cycle-consistent reconstruction and introduces NDWI-based regularization to constrain predictions to aquatic regions. To address high spectral dimensionality and strong inter-band correlations, we design a hierarchical transformer encoder with efficient tokenization and a lightweight refinement stage for discriminative feature extraction. Moreover, a stacked cross-attention decoder is employed to capture nonlinear mixing effects by adaptively fusing abundance and endmember representations. Extensive experiments on the three real-world River Scene datasets achieve AUCs of 0.9743, 0.9920, and 0.9982, respectively, demonstrating that the proposed method consistently outperforms competing approaches for nearshore underwater hyperspectral target detection.