Cross-Domain Hyperspectral Target Detection Based on Physical Constraints and Spectral Generation

Abstract

To address the problems of significant spectral scale differences between multispectral images (MSIs) and hyperspectral images (HSIs), strong spectral variability of targets, and insufficient collaborative utilization of spatial–spectral features in cross-domain hyperspectral target detection (HTD), this article proposes a physically constrained spectral generation (SG) and dual-stream spatial–spectral fusion network for HTD (PSG-HTD). PSG-HTD comprises three parts. First, a cross-domain spectral self-attention (CSSA) mechanism is introduced to model long-range cross-band dependencies between MSIs and HSIs. Second, PSG-HTD incorporates a standard spectral library as a physical prior to improve the semantic rationality and physical credibility of generated spectra. Third, an adaptive dual-stream spatial–spectral extraction (ADSSE) module jointly exploits spectral discriminative and spatial structural features for cross-domain HTD. Experimental results demonstrate strong performance in both SG and cross-domain HTD tasks: 1) for SG, PSG-HTD reduces the root-mean-square error (RMSE) by up to 71.0% and improves the peak signal-to-noise ratio (PSNR) by about 37.4% over comparison methods, indicating that embedding physical constraints effectively alleviates the spectral distortion problem and ensures physical credibility and 2) for cross-domain HTD, it achieves the highest mAP@0.5 of 0.917 and 0.714 on RarePlanes and RSOD-VHR10, respectively, indicating that high-fidelity SG provides robust, domain-invariant features that significantly enhance downstream cross-domain target identification capability. This study provides a new idea for SG and cross-domain HTD, and has important application potential in the fields of fine recognition and accurate classification of remote sensing targets.