SSEPDP: a spectral super-resolution network with spectral-spatial embedding for physical degradation perception model

Abstract

Dynamic and spectrally varying degradation severely limits the quality of hyperspectral images and poses great challenges for super-resolution, as conventional methods rely on static models that cannot adapt to real-world complexity. To address this, we propose a Spectral-Spatial Embedded Physical Degradation Perception network (SSEPDP). The key idea is to integrate learnable, dynamic degradation parameters into the classical physical imaging model, thereby achieving both high reconstruction accuracy and physical interpretability. Specifically, a dual-channel spectral-spatial feature split module first decouples the degradation into band-wise spectral and spatial latent codes, which are estimated from the radial power spectrum of the low-resolution observation. These codes are then aligned across multiple scales and fed into a spectral-spatial dynamic cross-attention module, which generates adaptive blur kernels that vary with spectral bands. A dedicated neural noise generator further models complex, signal-dependent sensor noise. The entire framework is supervised by a multi-scale spectral-spatial collaborative constraint loss that jointly enforces kernel rationality, spectral continuity, spatial fidelity, and degradation consistency. Extensive experiments on public datasets demonstrate that SSEPDP outperforms state-of-the-art approaches, reducing the spectral angle mapper by up to 22% and the spectral information divergence by up to 38%, while using only about 0.18 M parameters and 3.3 G MACs. Moreover, since the degradation parameters are synthetically generated, we directly compare the predicted and ground truth blur kernels and noise maps and provide statistical-significance and out-of-distribution analyses, giving a complete and reproducible validation of the physical-degradation-perception claim.