Abstract. Manual approaches to deforestation monitoring are time-consuming, inconsistent, and difficult to scale. Most existing systems rely on subjective interpretation or semi-automated tools, limiting repeatability and real-time assessment. While Sentinel-2 offers high spatial and temporal resolution, persistent cloud cover, especially in tropical areas like Palawan, hampers consistent observation. Moreover, most deforestation studies focus only on detecting past changes, lacking spatially aware forecasting capabilities. This study presents an automated, regionally trained pipeline for deforestation monitoring and forecasting using cloud-free Sentinel-2 imagery and machine learning. Imagery from 2017 to 2024 was preprocessed using OmniCloudMask for cloud and shadow masking, followed by VPint2 to fill cloud-covered gaps in the imagery. Manual annotations were used to classify land cover into six classes using Random Forest, XGBoost, and LightGBM. Random Forest achieved the best performance (90.59% accuracy). Forest classification in 2019 was validated against NAMRIA’s 2020 Land Cover Map, with an F1-score of 89.5% and IoU of 81% for the forest class, confirming strong agreement with ground truth for forest cover. Forest area declined gradually from 2017 to 2024, especially along edges near expanding croplands. To anticipate future change, a CNN–LSTM model was trained on tree probability maps to forecast forest cover from 2025 to 2029. The model achieved 93.12% accuracy and a forest F1-score of 92.48% when validated on 2024 data. The proposed system provides an objective and repeatable approach for forest monitoring and near-term forecasting.
Laag et al. (Tue,) studied this question.