Predicting Deforestation Risk
I began by building a dataset that integrated information from three open-source datasets: Hansen Global Forest Change, SRTM, and CHIRPS. Combining satellite imagery with geospatial data, I constructed a feature table covering more than 6,000 grid cells across two districts in Gia Lai Province.
However, a predictive model should not only perform well on data it has already seen. It must also be tested to determine whether it can generalize to geographically distinct areas that were not used during training.
To address this challenge, I designed a spatial cross-validation framework and trained and benchmarked two machine learning algorithms: Logistic Regression and Random Forest.
The Random Forest model achieved an AUC of 0.942 on internal validation and 0.767 on a spatial transfer test. I further optimized the risk-prioritization mechanism, enabling the model to identify 17.5% of actual deforestation events.
This research is currently being submitted to Remote Sensing.
▶
▶