MSSARN - multi-layer spectral-spatial attention based residual network for HSI classification

Abstract Hyperspectral imaging (HSI) presents several challenges in remote sensing, particularly in classification tasks. One significant challenge is the high dimensionality of HSI data, which includes hundreds of spectral bands. Another challenge is the spectral-spaatial complexity of HSI, which requires algorithms to effectively integrate both spectral and spatial information for accurate classification. This paper introduces the Multi-Layer Spectral-Spatial Attention Residual Network (MSSARN), deep-learning model designed to address high dimensionality and spectral-spatial complexity challenges. Our MSSARN model combines Spectral Attention (for spectral feature extraction), Neighborhood Convolution (for spatial feature extraction), Follow Patch techniques (features matching), and Residual Connections, creating a synergistic architecture that enhances feature extraction and representation from HSI. The network also incorporates residual connections to mitigate the vanishing gradient problem, enhancing the robustness and efficiency of the training process, particularly for high-dimensional, complex datasets. Furthermore, the model achieves a balance between computational efficiency and performance, which is validated through extensive experiments across six hyperspectral datasets. Compared to eight different HSI classification algorithms, the MSSARN consistently outperforms in terms of classification accuracy (more than 98% for all datasets), establishing its effectiveness and potential as a solution for HSI classification challenges. By leveraging the integrated strengths of various deep learning components, the MSSARN presents a significant advancement in the field, offering new insights and methodologies for the intelligent extraction of information from hyperspectral remote sensing images.