Towards a scalable and transferable approach to map deprived areas using Sentinel-2 images and machine learning

Abstract

African cities are growing rapidly and more than half of their populations live in deprived areas. Local stakeholders urgently need accurate, granular, and routine maps to plan, upgrade, and monitor dynamic neighborhood-level changes. Satellite imagery provides a promising solution for consistent, accurate high-resolution maps globally. However, most studies use very high spatial resolution images, which often cover only small areas and are cost prohibitive. Additionally, model transferability to new cities remains uncertain. This study proposes a scalable and transferable approach to routinely map deprived areas using free, Sentinel-2 images. The models were trained and tested on three cities: Lagos (Nigeria), Accra (Ghana), and Nairobi (Kenya). Contextual features were extracted at 10 m spatial resolution and aggregated to a 100 m grid. Four machine learning algorithms were evaluated, including multi-layer perceptron (MLP), Random Forest, Logistic Regression, and Extreme Gradient Boosting (XGBoost). The scalability of model performance was examined using patches of the different deprived types identified through visual image interpretation. The study also tested the ability of models to map deprived areas of different types across cities. Results indicate that deprived areas have heterogeneous local characteristics that affect large area mapping. The top 25 features for each city show that models are sensitive to the spatial structures of deprived area types. While models performed well on individual cities with XGBoost and MLP achieving an F1 scores of over 80%, the generalized model proves to be more beneficial for modeling multiple cities. This approach offers a promising solution for scaling routine, accurate maps of deprived areas to hundreds of cities that currently lack any such map, supporting local stakeholders to plan, implement, and monitor geotargeted interventions.