Deep learning for large-scale mapping of coffee cultivation in Minas Gerais, Brazil

This is the authors' abstract. We don't add key points for this paper.

Despite being the world's largest coffee producer, the Brazilian state of Minas Gerais still lacks up-to-date methodologies for the remote mapping of its coffee fields, largely due to strong regional heterogeneity in biophysical conditions. To help closing this gap, this study produces a state-level map of coffee plantations using deep learning applied to Sentinel-2 imagery. We evaluate three models for semantic segmentation: Single U-Net, Multi-Region U-Net, and GAN-based U-Net. A key innovation is a multitemporal inference strategy using Sentinel-2 images from 2020 to 2024, combined with a temporal filter incorporating agronomic knowledge of coffee cultivation cycles, to improve the identification of recently pruned or newly planted fields. We find that Multi-Region U-Net achieved the highest accuracy and was applied statewide, detecting 915 thousand ha of coffee in 2023. Our results demonstrate the potential of deep learning for scalable agricultural mapping and monitoring of large and diverse geographies.