A Lightweight Deep Learning Framework for Land Cover Classification from Sentinel-2 Imagery

Land-cover classification is an important task in monitoring ecological changes, conserving natural resources, and supporting urban planning. With modern advances in remote sensing and machine learning techniques, even highly accurate classifications are now easier and more scalable. This study investigates transfer learning for land cover classification using the EuroSAT dataset composed of Sentinel-2 satellite images belonging to ten distinct land cover classes-using a pre-trained DenseNet121 model. The study has considered a modified DenseNet121 architecture where the last five layers are fine-tuned in order to improve classification. Advanced features such as data augmentation adaptive learning rate scheduling and early stoppages are incorporated to optimize performance for this model. The results of experiments indicate effective performance, with training and testing accuracies of 89.33% and 90.79%, respectively. These findings further validate the effectiveness of transfer learning for improving classification accuracy with less computational power. This methodology provides a scalable and efficient framework for land cover classification which would indeed assist in environmental monitoring and informed decision-making activities.