Hyperspectral Image Classification Using PCA-Based Deep 2D Convolutional Neural Networks with Spatial Regularization

Abstract

Hyperspectral image classification remains a challenging task due to the high dimensionality of spectral data and the limited availability of labeled samples. In this work, we propose an efficient and robust classification framework that combines principal component analysis (PCA) for spectral dimensionality reduction with a deep two-dimensional convolutional neural network (2D CNN) for spatial feature learning. The PCA step effectively compresses the spectral dimension, preserving 98.43% of the spectral variance, allowing the network to focus on spatial pattern extraction with significantly reduced computational complexity. Additionally, a spatial regularization step based on majority filtering is applied to enforce spatial consistency and correct isolated misclassified pixels. Experiments conducted on the widely used Indian Pines dataset demonstrate that the core CNN achieves a strong baseline Overall Accuracy (OA) of 97.79%, which increases to a state-of-the-art 98.93% after spatial regularization. Furthermore, the proposed model achieves a Cohen’s Kappa coefficient of 0.969, outperforming several traditional and advanced deep learning approaches, including complex 3D-CNNs and spectral-spatial residual networks (SSRN). The results confirm that the proposed PCA + CNN framework delivers highly accurate classification while maintaining a relatively simple architecture and low computational cost, making it highly suitable for CPU-based training environments.