ForestCast: Forecasting Deforestation Risk at Scale with Deep Learning

Matthew Overlan, Charlotte Stanton, Maxim Neumann, Arianna Manzini, Julia Haas, Michelangelo Conserva, Melanie Rey, Kira Prabhu, Youngin Shin, Kuan Lu, Drew Purves

Deforestation is a major threat to biodiversity and the stability of the climate. Current monitoring solutions provide reactive alerts only after deforestation has occurred, rendering them largely insufficient for prevention. Proactive deforestation prevention necessitates forecasting at-risk areas, however, previous forecasting efforts have been constrained by their reliance on simple statistical models and limited feature sets. This paper introduces the ForestCast Southeast Asia dataset, the first publicly available dataset dedicated to training deep learning models for the task for deforestation risk forecasting. We benchmark several deep learning models from the literature, as well as a Random Forest Decision Tree model, and find that the deep learning models perform the best. Furthermore, we test the relative importance of three classes of input data: satellite imagery, derived feature layers such as slope and distance to roads, and change history, an image summarizing past deforestation. Surprisingly, we find that our best models can achieve equal results using only the change history as input. Lastly, we present a preliminary assessment of the ethics of using deforestation risk models in practice.