Evaluating Afforestation and Deforestation Using Attention-based Semantic Segmentation of Sentinel-2 RGB-NIR Bands

Monitoring deforestation and afforestation is essential for sustainable environmental management. Deep learning methodologies enable the automation of these processes, improving detection efficiency and precision. The paper presents a solution for identifying deforestation within forests by furnishing coordinates of impacted regions to facilitate prompt intervention. The research proposes a customized U-Net model for semantic segmentation to enable change detection. The model performs better by incorporating attention gates, U-Net, Vision Transformers, and heterogeneous kernel convolution. Attention gates improve positional information extraction, while Vision Transformers capture long-range dependencies. Trained on the Sentinel-2 level 2A 4-band dataset, which includes Red-Green-Blue (RGB) and Near-Infrared (NiR) bands, the model excels at handling complex data and generates a binary mask of the input image. Changes in forest cover are identified through pixel-by-pixel comparison of the binary masks of temporal images taken at different times. Utilizing data from the Amazon Forest for training, the model showcases its efficacy in the Arunachal Pradesh forest region and its vicinity, establishing itself as a versatile tool for identifying forest changes. During training, the model demonstrated results as a loss: 0.0707, accuracy: 0.9719, precision: 0.9706, recall: 0.9765; during validation, findings were loss: 0.0582, accuracy: 0.9762, precision: 0.9766, recall: 0.9760, and an F1 score of 0.97 was obtained during training and validation. The proposed model was tested on input images from selected areas with seasonal variations during winter, including forest areas with snow cover. This methodology advances proactive conservation and sustainable land management practices, enabling continuous environmental monitoring.