Spectral–Spatial Land Cover Mapping Using a Deep 3D Convolutional Neural Network on Hyperspectral Data

Abstract

Recent studies have demonstrated that incorporating both spectral and spatial information significantlyenhances the accuracy of hyperspectral image (HSI) classification. Since HSI data is commonly structuredas 3D cubes, where two dimensions represent spatial information and the third encodes spectral bands,applying 3D spatial filtering becomes an intuitive and effective way to extract spectral–spatial features simultaneously.In this work, we propose a deep 3D Convolutional Neural Network (3D–CNN) architecturefor accurate and efficient HSI classification. Unlike many previous methods, the proposed model processesthe hyperspectral cube directly as a volumetric input, eliminating the need for manual feature extractionor dimensionality reduction. This design allows the network to jointly learn spatial and spectral dependencieswhile maintaining a compact architecture with fewer trainable parameters. Experimental evaluationson two benchmark datasets—Pavia University and Salinas Valley—demonstrate that the proposedmodel achieves 95% overall classification accuracy, compared to 88% obtained with the conventionalK-Nearest Neighbors (KNN) baseline. The 3D–CNN consistently outperforms traditional methods acrossmost land-cover classes, particularly in spectrally similar or complex regions, confirming its effectivenessfor spectral–spatial land cover mapping. These results highlight the practicality and robustness of deepvolumetric learning for real-world hyperspectral image analysis