Position-Aware Dual-Domain Hybrid Transformer for Single Hyperspectral Image Super-Resolution

Abstract

Hyperspectral image super-resolution (HSI-SR) is a technique that increases spatial resolution while preserving spectral accuracy. This capability is important for remote sensing interpretation and for many downstream applications. Most existing deep learning solutions are predominantly built upon convolutional neural networks (CNNs). Nevertheless, these approaches often struggle to simultaneously capture long-range spectral correlations, local spatial patterns, and explicit positional cues.To address these issues, we propose a novel Position-Aware Dual-Domain Hybrid Transformer (PADHT). Concretely, we devise a Position-Aware Attention Module (PAM) that employs MLP-based positional encoding to inject spatial structure information, thereby strengthening positional representation. Furthermore, we construct a Dual-Domain Hybrid Transformer Module (DHTM) that integrates Local Window Self-Attention (LWSA) and Deformable Multi-Head Spectral Attention (DMSA) in parallel. LWSA focuses on spatial neighboring correlations within windows. It introduces Exponential Spatial Relative Position Bias (ES-RPB) to finely distinguish weights of different relative positions. DMSA adaptively adjusts the receptive field through deformable convolution to capture dependencies across distant spectral bands. Moreover, we propose a Query Interaction Mechanism (QIM) to enable complementary fusion of cross-domain features, where outputs of spatial and spectral branches serve as cross-domain queries for each other. Comprehensive evaluations conducted across four hyperspectral datasets confirm that PADHT delivers better results than representative comparison methods on multiple evaluation metrics.