Agriculture Land Classification Using MobileNetV2 on Euro SAT RGB Dataset

Abstract

LULC classification from the remote sensing imagery is an important and critical task for sustainable agricultural planning, environmental monitoring, and precision farming. In this paper, we propose an efficient deep learning framework for classifying agriculture land from high-resolution satellite using the Euro SATRGB dataset that cover ten land cover classes. We used light weight CNN MobileNetV2 with Transfer learning, Extensive data augmentation, Mixed precision training and cosine decay learning rate scheduler. The proposed model achieved a test accuracy of 98.67%, which is a state-of-the-art among newer models tested on the same dataset and was trained on 19,504 images, validated on 3,446 images and tested on 4,050 images. We have shown that MobileNetV2 is not only the fastest model but also competitive in performance, demonstrating its ability as a light-weight but highly accurate model with potential for large-scale agricultural land classification tasks. This study lays the foundation for lightweight deep learning models for scalable and real-time practical applications in precision agriculture and builds on the rapidly growing body of literature in this area.