Typology Based Deep Learning for Borderless Humanitarian Aid in the Venezuelan [Refugee] Context

As part of a larger collaboration between the MIT Leventhal Center for Advanced Urbanism (LCAU) and the World Food Programme (WFP) this project sought to fulfill the WFP’s need of obtaining close to real-time data of the fast evolving improvized structures in La Pista, Maicao, one of Colombia’s largest Venezuelan refugee communities.

Through the use of deep learning, my collaborator (MIT undergraduate student Manolo) and I trained a deep learning model by building typology rather than regional borders. This training method applies the best practices employed in environmental sciences, which train models based on ecosystem type, rather than geopolitical locations and enable existing to trained models to be used globally without steep decline in performance. Training in this method, allows us to avoid regional biases that persist in existing pre-trained models, and provides the WFP a pre-trained model and training modelling approach that can allow for wider applicability that crosses geopolitical borders.