Multiscale Wavelet and Graph Network With Spectral Self-Attention for Hyperspectral Image Classification

• Published in: IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
• Main Authors: Anyembe C. Shibwabo, Zou Bin, Tahir Arshad, Jorge Abraham Rios Suarez
• Format: Article
• Language: English
• Published: IEEE 2025-01-01
• Subjects: Attention mechanism; convolutional neural network (CNN); feature integration; graph convolution network (GCN); hyperspectral image (HSI) classification
• Online Access: https://ieeexplore.ieee.org/document/11018235/
• ISSN: 1939-1404; 2151-1535

Hyperspectral image (HSI) classification has gained increasing attention in remote sensing due to its finegrained spectral information. However, existing methods still face significant challenges in preserving high-frequency details, modeling long-range dependencies, and integrating spectral, spatial, and frequency-domain features. In this work, we propose MWGN-SSA, a powerful network designed to enhance HSI classification by fusing multidomain features. MWGN-SSA consists of three core modules: a multiscale learnable wavelet network (MLWN), a window-based spectral self-attention (WSSA) mechanism, and a deep-hop graph convolutional network (DH-GCN). First, MLWN adaptively decomposes HSIs into frequency subbands, retaining critical high-frequency textures for small or spectrally subtle targets. Second, WSSA captures both local and global spectral correlations using a windowed self-attention scheme. Third, DH-GCN constructs a deep graph structure to model spatial topology and overcome oversmoothing. A feature integration module combines outputs from all branches for final prediction. Extensive experiments on four benchmark datasets demonstrate that MWGN-SSA achieves superior accuracy and robustness, particularly in complex and imbalanced HSI scenes.