Deep Learning for Satellite Image Segmentation: Deforestation Detection and Reforestation Zone Identification

Abstract

Rapid changes in land use and land cover, particularly due to deforestation, present significant challenges for sustainable environmental management. Accurate monitoring of these changes is essential for effective conservation and restoration initiatives. This study explores advanced deep learning and machine learning methodologies to enhance land use and land cover (LULC) classification accuracy. Utilizing the robust Sent2 LULC dataset, we employ four distinct models: U-Net with a ResNet50 backbone, MobileNet, DenseNet integrated with custom shallow features, and Random Forest. Each model is selected for its specific strengths in image processing and classification tasks. U-Net excels in semantic segmentation, making it ideal for detailed mapping of deforested and reforested areas. DenseNet's densely connected architecture maximizes feature extraction, while MobileNet offers a lightweight, efficient alternative suitable for computationally limited scenarios. The Random Forest model serves as a robust benchmark, valued for its straightforward statistical analysis. This research conducts a comprehensive comparison of these models in terms of accuracy, efficiency, and practical applicability in detecting deforested and reforested zones. By leveraging the strengths of both traditional machine learning and advanced deep learning techniques, this project aims to enhance our understanding of LULC dynamics, contributing to informed strategies for environmental conservation and sustainable land use planning. This initiative not only advances scientific understanding but also fosters greater collaboration between remote sensing and computer vision fields, driving innovations in environmental monitoring.