A Spectral-Spatial Attention Guided Multi-Scale Convolutional Network for Hyperspectral Image Classification

Hyperspectral image classification presents significant challenges due to the high dimensionality of the data and the intricate spatial-spectral relationships inherent within hyperspectral imagery. This proposed work builds on certain well-established techniques, its novelty lies in the integration and adaptation of these components into a unified framework designed to address specific challenges in hyperspectral image classification. Unlike traditional PCA applied globally, this research work performs PCA within graph-based segmented regions. This localized approach preserves the spatial coherence of hyperspectral data while reducing dimensionality efficiently. Next to this step, a novel self-attention mechanism within a hybrid 3D-2D CNN architecture is introduced that allows the model to dynamically prioritize critical spectral bands while extracting comprehensive spatial-spectral features. The combination of graph-based segmentation, localized PCA, and attention-guided CNNs creates a robust and cohesive framework that enhances feature extraction and classification accuracy. The proposed framework is evaluated on four publicly available hyperspectral datasets Indian Pines, Kennedy Space Center, Salinas, and Pavia University and compared against seven state-of-the-art models, including SVM, 2D-CNN, 3D-CNN, AHAN, AF2GNN, DSN, and TDS-BiGRU. The experimental results demonstrate the superiority of the proposed approach, achieving an overall accuracy of 99.28% on Indian Pines, 99.99% on Salinas, 99.97% on Pavia University, and 99.34% on KSC dataset, consistently outperforming the existing methods. This framework effectively leverages segmentation, localized dimensionality reduction, and attention mechanisms, offering a robust and efficient solution for hyperspectral image classification. These results confirm the model's capability to address the complexities of hyperspectral data, providing a significant advancement in the field.