Deep Spectral-Spatial Residual Networks for High-Precision Hyperspectral Image Classification

Hyperspectral imaging (HSI) provides rich spectral and spatial information for material and land-cover classification, but high dimensionality and limited labeled data present significant challenges. In this work, we propose a Spectral-Spatial Residual Network (SSRN) that efficiently integrates spectral and spatial features using residual convolutional blocks. PCA-based dimensionality reduction and patch-based processing reduce computational complexity while preserving critical information. The proposed SSRN is evaluated on the benchmark Indian Pines dataset and achieves an overall accuracy of 99.14%, average accuracy of 95.42%, and a Kappa coefficient of 0.9902, outperforming traditional methods such as SVM and Random Forest, as well as 1D and 2D CNNs. Visualization of predicted classification maps and confusion matrix analysis confirm that SSRN accurately captures spatial structures and discriminates spectrally similar classes. These results demonstrate the effectiveness of combining spectral-spatial feature learning with residual learning for high-performance hyperspectral image classification.