Hybrid 2D–3D Convolutional Neural Network for Hyperspectral Land-Cover Classification

Abstract

Hyperspectral image (HSI) classification faces multiple challenges because it deals with extensive spectral data which creates intricate patterns between different bands and complex spatial differences in scenes that shallow classifiers struggle to handle. The technology of convolutional neural networks (CNNs) offers an effective solution because these models learn hierarchical spectral and spatial features through data analysis which eliminates the need for human–designed feature extraction. The research introduces a hybrid CNN model which combines 2D and 3D convolutional operations to process spatial information together with spectral–spatial structural data. The 2D branch extracts spatial features from image patches yet the 3D branch analyzes spectral information and spatial relationships through spectral-spatial cubes. The system generates a combined representation which unites the advantages of both streams through its late fusion approach. We evaluate the model on the Salinas HSI benchmark and compare it against conventional machine– learning baselines (e.g., KNN, SVM, RF). The hybrid network reaches better overall accuracy together with enhanced stability when handling classes that show small spectral differences and complicated texture patterns.