Hyperspectral image super-resolution via joint network with spectral-spatial strategy

Hyperspectral image (HSI) super-resolution (SR) faces significant challenges due to the inherent difficulty in acquiring large-scale training data and the complex spectral-spatial relationships in HSIs that conventional deep-learning-based methods often fail to fully exploit. While existing approaches typically stack convolutional layers to increase network depth, they frequently overlook the structured continuity of spectral bands and non-local spatial similarities, resulting in limited performance and overfitting risks. To address these limitations, we propose SRLSGAT, a novel joint spectral-spatial network that combines a vertical-horizontal bi-directional LSTM (VH-BiLSTM) for modeling multi-directional spectral correlations and a multi-adjacent weight matrix graph attention network (MAW-GAT) for capturing non-local patch relationships. Besides, we design a spectral attention mechanism (SpeAM), which dynamically weights remote dependencies through bidirectional spectral sequence analysis, while the graph-based spatial processing adaptively learns patch similarities through optimized edge connections. Extensive experiments on three HSI datasets show that SRLSGAT has better performance relative to SOTA SR methods. The code is accessible at https://github.com/Dongyx1128/SRLSGAT.