Evaluation of Deep Learning-Based Segmentation Models for Forest Cover Detection in Iran: A Study on Sentinel-2 Imagery

Forests, as critical pillars of ecosystems, play an essential role in stabilizing the climate and preserving biodiversity. Industrial development, urban sprawl, and unsustainable exploitation have led to a reduction in forest cover and an increase in wildfire occurrences, particularly in regions of Iran where the invaluable Hyrcanian and Alborz forests are under threat. In this research, Sentinel-2 satellite images with a 10-meter spatial resolution were meticulously annotated and partitioned into training, validation, and test sets to assess deep learning-based image segmentation techniques for detecting forested areas. For this purpose, three deep learning models—Nested U-Net, SegNet, and FCN—were compared in terms of their ability to accurately differentiate forest cover from other land cover types. The results indicate that the Nested U-Net model, achieving an IoU of 81.36% and a Recall of 99.54%, outperforms the other models, demonstrating a high capacity for multi-scale feature extraction and precise delineation of forest boundaries. These findings underscore the significance of integrating remote sensing technologies with advanced segmentation algorithms in the management and conservation of natural resources.