Lightweight Transformer Model for Agricultural Land Use and Land Cover Classification

Observing agricultural land use via remote sensing images is essential for ensuring food security, estimating yields and planning efficient exports nonetheless precise classification continues to be difficult because of the varied and evolving characteristics of agricultural environments. This research aims to evaluate and optimize advanced deep learning architectures particularly Vision Transformer (ViT) models for agricultural land-use classification tasks. Specifically, we employed ViTBase-16 and other lightweight models DeiT-Tiny and EfficientNet-B0 applying techniques such as model layer compression and advanced data augmentation CutMix and Cutout to achieve high accuracy while significantly reducing computational complexity. Evaluation was performed using three benchmark remote sensing datasets EuroSAT, NWPU-RESISC45 and SIRI-WHU which include diverse spatial resolutions and agricultural classes relevant for practical monitoring. Findings indicate that the optimized ViT algorithm is highly effective in recognizing global spatial connections, consistently achieving remarkable classification accuracy exceeding 99% on a newly assembled dataset containing around 200 samples of Google Earth imagery. Furthermore, the first time in agricultural image classification compressing the ViTBase model by pruning 50% of its layers significantly reduced complexity maintainingcompetitive accuracy 97.9% on SIRI-WHU. The resulting models are particularly suitable for deployment on devices with limited computational resources supporting real-world operational agricultural monitoring systems. This study emphasizes the revolutionary possibilities and practical use of optimized transformer-based models that offer scalable and efficient solutions specifically designed for precision agriculture applications.