CLASSIFICATION OF SATELLITE IMAGES USING ADVANCED TOKENS-TO-TOKEN TRANSFORMER WITH PSO OPTIMIZATION

Satellite image classification plays a pivotal role in diverse applications, including land use monitoring, urban planning, and environmental analysis. This paper explores the comprehensive classification of satellite images into five classes: desert, forest, green fields, oceans, and urban areas. Initial preprocessing techniques such as resizing, histogram equalization, noise reduction, rotation, cropping, color jittering, and random erasing were applied to enhance data quality. Four Transformer Neural Network (TNN) models i.e., Vision Transformer (ViT), Class Attention Image Transformer (CAiT), Pyramid Vision Transformer (PVT), and Tokens-to-Token Vision Transformer (T2T-ViT) were analyzed. Among these, T2T-ViT demonstrated the best accuracy at 73.21%. Further optimization of T2T-ViT using machine learning techniques, including ensemble methods, feature scaling, and stratified k-fold cross-validation, achieved an accuracy of 84.09%. Subsequently, Particle Swarm Optimization (PSO) was employed for hyperparameter tuning, boosting the model accuracy to 98.75%. This research highlights the efficacy of combining advanced TNN architectures with optimization strategies for robust satellite image classification.