Multi-Scale Hybrid Spatial-Spectral Transformer for Hyperspectral Image Classification

Abstract

Hyperspectral image classification (HSIC) requires the effective integration of spectral and spatial information to achieve accurate land-cover mapping. Conventional convolutional neural network (CNN) based methods are limited in modeling long-range dependencies and multi-scale contextual features inherent in hyperspectral data. In this research paper, a Multi- Scale Hybrid Spatial–Spectral Transformer (MSH-SST) framework for robust HSIC was proposed. The proposed architecture combines convolutional layers with transformer-based self-attention mechanisms to jointly learn local spatial patterns and global spectral-spatial relationships. Multi-scale feature extraction modules are employed to capture contextual information at different spatial resolutions, enhancing the representation of complex structures and boundary regions. A hybrid spatial spectral token embedding strategy is designed to preserve discriminative spectral information while maintaining spatial coherence. The transformer encoder further models long-range dependencies across both spectral and spatial dimensions, improving class separability. Experimental results conducted on well-known Indian Pines, Pavia University, and Salinas hyperspectral datasets validate that MSH-SST consistently outperformed over state-of-the-art CNN and transformer-based approaches in terms of classification accuracy and robustness, particularly under limited training sample conditions. The proposed framework effectively balances fine-grained spatial feature learning and global spectralspatial dependency modeling, making it a powerful solution for high-precision hyperspectral image analysis.