PCA-ViT: Hyperspectral Image Classification using Principal Component Analysis and Vision Transformer

Hyperspectral image classification is essential in remote sensing applications, aiming to accurately categorize land cover or materials depicted in hyperspectral data. This paper introduces PCA-ViT, a novel approach for hyperspectral image classification that integrates Principal Component Analysis (PCA) with the Vision Transformer (ViT) architecture. PCA serves as a spectral dimension reduction technique to alleviate the curse of dimensionality inherent in hyperspectral data, transforming the original data into a more manageable format. The transformed data are then input into the ViT model, which utilizes self-attention mechanisms to capture spatial dependencies among image patches, avoiding traditional convolutional layers. Extensive experiments on benchmark hyperspectral datasets, including Indian Pines, University of Pavia, and Salinas Scene, demonstrate PCA-ViT's superior performance. It achieves 99.95% accuracy on Indian Pines, 100% on University of Pavia, and 100% on Salinas Scene, showcasing the effectiveness of transformer-based architectures in hyperspectral image classification tasks.