Mapping land uses following tropical deforestation with location-aware deep learning

The rates of tropical deforestation remain alarmingly high. To enable effective, targeted policy responses, detailed data on its driving forces is needed-each deforestation event needs to be attributed to an agricultural commodity or another land use. Remote sensing allows us to monitor land use conversion following deforestation, providing a proxy of drivers. However, recognizing individual commodities is challenging due to spectral similarities, the limited spatial resolution of free satellite imagery, and limited labeled data. To tackle these challenges, we propose a deep learning, multi-modal approach for the recognition of post-deforestation land uses from a time series of Sentinel-2 images, geographic coordinates, and country-level statistics of deforestation drivers. To integrate the modalities, we design a Transformer-based model with modality-specific encoders. The approach reaches 87% accuracy, an improvement of 10% over the image-only baseline, with little increase in data volume, computations, and model size. It works well in low-data regimes, and can be easily extended to include other modalities. Overall, this work contributes towards detailed, repeatable, and scalable mapping of deforestation landscapes, providing necessary data for the design and implementation of targeted interventions to protect tropical forests.