A Hybrid CNN-Swin Transformer Module for Hyperspectral Image Classification

Hyperspectral image (HSI) classification stands as a pivotal mission in remote sensing applications, demanding intricate fusion of spectral and spatial information for accurate analysis. As a result of substantial variations between the spatial and spectral characteristics in the HSIs, the extraction of local and global spectral-spatial characteristics is still not fully completed. To face this challenge and leverage the advancements in deep learning, this paper presents a novel approach integrating the strengths of both 3D and 2D convolutional neural networks (CNNs) and the Swin Transformer for robust HSI classification, called CNN-Swin. The 3D-CNN leverages the inherent three-dimensional structure of HSIs to capture intricate spectral-spatial features, which are crucial for accurate classification, and the 2D-CNN to know more about spatial representation. The Swin Transformer, with its group attention and sliding window mechanisms, effectively learns long-range spectral dependencies while preserving global information to improve the representation of features. The calculation of sliding window attention can consider the contextual information of various windows, which is to be able to keep the global features of the HSI and enhance the outcomes of classification. Our suggested model is evaluated in the experiments utilizing three benchmark hyperspectral datasets: Indian Pines (IP), Pavia University (PU), and Salinas Scene (SA). Experiments outcomes illustrate that our suggested model can accomplish test accuracies of 99.91%, 99.98%, and 100.0% on the IP, PU, and SA datasets, respectively. The outcomes illustrate that our proposed model outperforms several state-of-the-art classification techniques for HSIs.