Hyperspectral Image Classification Using Custom Spectral Convolutional Neural Networks (CSCNNs)

Custom spectral convolutional neural networks (CSCNNs) combine the strengths of convolutional neural networks with specialized spectral processing, resulting in improved classification accuracy by effectively capturing subtle variations in hyperspectral data. This paper proposes a CSCNNs based approach to classify hyperspectral images. The proposed method leverages the high-dimensional spectral data inherent in hyperspectral images, employing convolutional layers specifically designed to capture spectral-spatial features. By reducing dimensionality through principal component analysis and creating image patches, the model is trained to recognize complex patterns across different spectral bands. In addition, a comprehensive analysis of CSCNN performance is carried out, focusing on its architecture, key features, and benefits in computational efficiency and spectral representation. Experimental results on datasets such as Salinas-A, Pavia University (Pavia-U), and Indian Pines demonstrate that the CSCNN model surpasses traditional methods, achieving higher classification accuracy and more robust performance metrics like overall accuracy (OA), average accuracy (AA), and Kappa coefficient. The proposed model achieved an overall accuracy (OA) of 98.2%, 96.7%, and 92.5% for the Salinas-A, Pavia-U, and Indian Pines datasets, respectively. The model's average accuracy (AA) was 97.5%, 95.4%, and 89.8% for the same datasets. Furthermore, the model attained Kappa coefficients of 97.7%, 95.2%, and 89.1% for Salinas-A, Pavia-U, and Indian Pines, respectively.