Weakly Supervised Hyperspectral Target Detection Based on Contrastive Learning and Graph-Structural Position

Hyperspectral target detection is significant in remote sensing, yet existing methods heavily rely on strong supervision. Target prior spectra are sensitive to illumination, atmospheric, and sensor variations, causing inevitable deviations that degrade performance. Pixel-level annotations are prohibitively costly due to high spatial resolution and subpixel localization difficulties. Multiple instance learning enables weakly supervised detection using region-level labels, but most existing methods neglect spatial dependencies among pixels and suffer from limited localization capability. To address these issues, this article proposes a graph-structured positional transformer for multiple-instance learning (GSPT-MIL), a graph-structured and position-aware weakly supervised multiple-instance learning (MIL) method that achieves accurate pixel-level detection using only bag-level labels, without requiring target priors or pixel-level annotations. The method performs iterative optimization under an expectation-maximization (EM) framework. The expectation step (E-step) infers target locations via a local similarity graph convolution module (LS-GCM) that models spatial dependencies, while the maximization step (M-step) fuses absolute position information via a multiscale contextual position encoder and enhances feature discriminability via a spectral–spatial feature alignment network (SS-FAN). Extensive experiments demonstrate that GSPT-MIL consistently outperforms existing weakly supervised and conventional detection methods, validating its effectiveness and superiority.