Spectral-Spatial Classification of Hyperspectral Imagery Using CNNs and GANs: A Study on Pavia University and Salinas Datasets

Abstract

Hyperspectral imaging (HSI) has emerged as a cornerstone of remote sensing, capturing detailed spectral information across hundreds of bands. This study rigorously evaluates the efficacy of deep learning models specifically Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs) for HSI classification, using the benchmark Pavia University and Salinas datasets. We address the high dimensionality and limit labelled data challenges inherent in HSI through advanced preprocessing (e.g., noise reduction, normalization) and data augmentation. Our CNN architecture leverages spectral- spatial feature extraction, achieving overall accuracies of 90.25% (Pavia University) and 93.75% (Salinas), surpassing traditional methods. GANs, employed for synthetic data generation, enhance robustness but yield slightly lower accuracies (88.15% and 91.55%, respectively). This work provides a mathematical and empirical foundation for deep learning in HSI, offering insights into model optimization and future research directions.