MAPPING POVERTY IN THE PHILIPPINES USING MACHINE LEARNING, SATELLITE IMAGERY, AND CROWD-SOURCED GEOSPATIAL INFORMATION
Keywords: Poverty estimation, Machine learning, Deep learning, Transfer learning, Remote sensing, GIS, Volunteered geographic information, OpenStreetMap
Abstract. Mapping the distribution of poverty in developing countries is essential for humanitarian organizations and policymakers to formulate targeted programs and aid. However, traditional methods for obtaining socioeconomic data can be time-consuming, expensive, and labor-intensive. Recent studies have demonstrated the effectiveness of combining machine learning and satellite images to estimate wealth in sub-Saharan African countries (Xie et al., 2016, Jean et al., 2016). In this study, we investigate the extent to which this method can be applied in the context of the Philippine archipelago to predict four different socioeconomic indicators: wealth level, years of education, access to electricity, and access to water. We also propose an alternative, cost-effective approach that leverages a combination of volunteered geographic information from OpenStreetMap and nighttime lights satellite imagery for estimating socioeconomic indicators. The best models, which incorporate regional indicators as predictors, explain approximately 63% of the variation in asset-based wealth. Our findings also indicate that models trained on publicly available, volunteer-curated geographic data achieve the same predictive performance as that of models trained using proprietary satellite images.