Deep Learning Maps Oil Palm Plantations in Malaysia and Indonesia Without Manual Annotation

A deep learning model utilizing Sentinel-2 satellite imagery produces oil palm plantation maps at a resolution of 10 meters for Indonesia and Malaysia, covering the years 2020 to 2024, without requiring new manual annotations. The U-Net framework, enhanced with Determinant-based Mutual Information (DMI), effectively mitigates label noise stemming from the less precise 100-meter historical maps. When validated against 2,058 points that were manually confirmed, the technique demonstrates overall accuracies of 70.64% for 2020, 63.53% for 2022, and 60.06% for 2024. This innovative method facilitates precise and current monitoring of swift land-use transformations, addressing both economic and environmental issues in Southeast Asia.