Spatial-Spectral Patch-Based CNN for Hyperspectral Image Classification

Abstract

Hyperspectral image classification is critical for remote sensing applications, but it remains challenging due to the high dimensionality and complex spectral–spatial correlations of hyperspectral data. In this paper, we propose a patch-based convolutional neural network (CNN) that effectively leverages both spatial and spectral information for pixel-level classification. The network extracts spatial–spectral patches from hyperspectral images and processes them through a series of convolutional blocks followed by global average pooling and a softmax classifier. We evaluate the proposed method on the Indian Pines dataset, a widely used benchmark in hyperspectral remote sensing. Experimental results demonstrate that the network achieves state-of-the-art performance, with overall accuracy (OA), average accuracy (AA), and Kappa coefficient exceeding 99%. Additionally, per class accuracy, precision, recall, and intersection over union (IoU) metrics indicate robust classification across both large and small classes. Comparative experiments with baseline methods, including U-Net, SegNet, DeepLabV3+, and SVM, highlight the effectiveness, efficiency, and stability of the proposed approach. The proposed framework is computationally efficient, easily extensible, and provides a solid foundation for real-world hyperspectral image classification tasks.