The Amazon plays a crucial role in mitigating global warming and preserving biodiversity, which are vital for the Earth's environment. However, deforestation has been ongoing for an extended period. Particularly, the vastscale of the region poses challengesin accurately identifying deforested areas, highlighting the growing importance of leveraging satellite information to prevent further damage and develop effective restoration plans. This study proposes a deep learning-based method to automatically generate high-resolution deforestation labels using low-resolution satellite imagery and existing deforestation label data. The proposed method initially performs primary training using low-resolution images and labeled data. Through this process, pseudo-label data are generated and used for iterative learning, ultimately improving the accuracy of deforestation area labeling on high-resolution satellite images. The output of this research can contribute to generating highly accurate high-resolution labeling data, even for satellite images without prior deforestation labels. This data can be utilized for detailed analysis of deforested areas and the development of precise restoration strategies.



Keywords: Amazon deforestation, Pseudo-labeling, High-resolution label, Semantic segmentation