Vision Transformer-Based LULC Classification Using Remotely Sensed Hyperspectral Image

Abstract

Deep-learning-based techniques have played a vast role in hyperspectral image classification. Nowadays, transformer-based deep-learning techniques have become widespread for image analysis. However, such techniques evolved basically in the natural language processing stream. These techniques involve a self-attention mechanism which is more insightful than well-known traditional convolution techniques to extract the features. We have investigated the vision transformer (ViT) for the best fit on reduced dimensionality hyperspectral image in the context of land use land cover (LULC) classification. In the related experiment, we have used a publicly available hyperspectral image, i.e., Indian Pines (IP), to demonstrate the results. The principal component analysis (PCA) is a popular technique for dimensionality reduction. The number of principal components (PCs) ranging from 10 to 30 is chosen based on the explained variance conveyed by them collectively for high-performance classification. However, significantly high (99.22–99.64%) performance of LULC classification can be achieved using ViT model merely on 1–3 PCs. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024.