Few-shot cross-domain hyperspectral–RGB image fusion via learnable spectral alignment and semi-supervised masking consistency

Most existing hyperspectral-RGB image fusion methods are trained and evaluated within the same dataset or sensor configuration, assuming identical band numbers, wavelength sampling positions, and output spectral spaces. This assumption limits their transferability to practical cross-domain scenarios, where spectral definitions may differ and target-domain high-resolution hyperspectral image (HR-HSI) labels are often scarce. To address this problem, this paper proposes a few-shot cross-domain hyperspectral-RGB image fusion framework with learnable spectral alignment and semi-supervised masking consistency. Starting from a source-domain pretrained fusion backbone, the proposed method freezes the backbone during target-domain adaptation and updates only a small set of adaptation parameters. A wavelength-guided learnable spectral alignment module adaptively maps source-domain spectral representations to the target spectral space, while a residual calibration head compensates for cross-domain spectral bias. Unlabelled target-domain samples are exploited through weak/strong-view consistency learning with RGB masking and spectral consistency constraints. Across six random seeds (2025–2030), the proposed method achieves 4.39 ± 0.04° SAM, 52.90 ± 0.05 dB PSNR, and 0.9945 ± 0.0001 SSIM on Wuhan-to-Xuzhou while updating only 37,831 parameters (approximately 0.11% of the network). Experiments on Wuhan-to-Xuzhou and Wuhan-to-DC Mall, together with an additional DC Mall-to-Xuzhou source-robustness test, demonstrate that learnable spectral alignment, target-specific residual calibration, and semi-supervised consistency jointly improve spectral fidelity, spatial reconstruction quality, and adaptation stability under limited target-domain supervision.