Monthly mapping of deforestation in the Amazon using Sentinel-1 data and a vision foundation model

Large-scale deforestation poses a significant threat to ecosystem stability and climate, leading to increased carbon dioxide emissions, which exacerbates global warming and ecological imbalance. To achieve sustainable forest management, precise large-scale monthly deforestation mapping has become increasingly important. The open access to Sentinel-1 data provides unprecedented opportunities for monthly deforestation mapping. However, previous monthly mapping based on Sentinel-1 and deep learning still needs improvement in accuracy, and the best strategies for large-scale model transfer have not been fully explored. This study proposes a new approach for monthly deforestation mapping based on Sentinel-1 data and an adapted Segment Anything Model (SAM), combined with active learning and transfer learning strategies for large-scale model transfer. The model was tested and evaluated in four different study sites: Rondônia in Brazil, Guangxi in China, California in the USA, and Hainan in China. The results showed the superior performance of our proposed adapted SAM method, with F1 scores ranging from 0.74 to 0.88 and IoU from 0.58 to 0.78. The combined model for the four regions achieved an F1 score of 0.81 and an IoU of 0.68, outperforming the baseline U-net model (combined F1 score of 0.78 and IoU of 0.64). When applied to new sites, the fine tune-based transfer learning significantly improved the model’s spatial generalization capability with the addition of a small number of target domain samples. Moreover, compared with random sampling approach, the active learning technique help reduce the required number of training samples to achieve the same level of accuracy. This study provides a comprehensive workflow for improved monthly deforestation mapping, emphasizing the advantages of combining Sentinel-1 SAR data with advanced models and strategies. Our method offers a reliable and efficient solution for large-scale deforestation monitoring, aiding in the timely detection of deforestation activities and supporting sustainable forest management strategies.