Beyond Specific Degradations: Prompt-Driven Spatial–Spectral Modeling for Hyperspectral Image Enhancement

Abstract

Hyperspectral images (HSIs) are often degraded by complex and coupled corruptions such as noise, blur, and haze. Most existing methods are designed for specific degradation types, lacking the flexibility for unified enhancement. In addition, jointly modeling spatial structures and spectral correlations remains challenging. In this article, we propose a prompt-driven spatial–spectral framework for HSI enhancement, which improves adaptability across diverse degradations via semantic–visual–spectral (SVS) multiprompt learning. The proposed framework introduces cross-modal prompts to provide adaptive guidance, enabling the model to handle diverse, mixed, and shifted degradation patterns within the considered HSI degradation space. To extract richer spatial representations, we design a direction-aware rectangle self-attention (DRSA) mechanism. It effectively captures horizontal and vertical dependencies to better restore anisotropic structures while reducing computational cost compared with dense global self-attention. To preserve spectral fidelity, we introduce a frequency-aware prompt-guided spectral modulation (FPSM) mechanism, which dynamically modulates spectral features using a learnable spectral prototype bank. We further devise a cross-modal semantic–visual dynamic prompting (SVDP) mechanism that integrates textual degradation descriptions and visual prototypes to generate degradation-aware prompts, enabling adaptive perception of diverse corruption patterns. Extensive experiments on remote sensing HSI datasets show that the proposed method achieves performance superior to state-of-the-art methods. It consistently improves both visual quality and spectral accuracy under various degradation conditions. These results validate the potential of multiprompt learning for unified HSI enhancement tasks.