Predicting forest loss risk for deforestation regulation using Convolutional Neural Networks

The EU Deforestation Regulation (EUDR) aims to reduce embedded deforestation in certain commodities imported into the European Union by requiring companies to prove that products are deforestation-free. Here, the level of due diligence obligations required is based on the overall risk score assigned to a specific country of origin. The first version of these risk scores, published last year, aims to reflect past deforestation rates and governance risks. However, the scores have been widely criticized by political and environmental advocacy groups for being politically motivated rather than representative of real deforestation risks, and for being too coarse in their national scale and commodity-invariant design. Hence, we here provide an additional, high-resolution, spatially explicit perspective on deforestation risk for the upcoming year. Using Convolutional Neural Networks (CNNs) and spatiotemporal data on past forest losses, landscape characteristics, and human development, we compute global risk maps for different drivers of forest loss, including deforestation for different commodities. With this analysis, we aim to complement the existing EUDR risk scores by highlighting sub-national variation and driver-specific risk patterns. We aim to contribute a transparent, data-driven perspective to ongoing discussions on deforestation risk in international policy processes.