High-Resolution Estimation of Rural Poverty: Insights from Ethiopia

High-resolution satellite imagery and machine learning techniques enable precise mapping of rural poverty, as demonstrated in this paper for the case of Ethiopia. Using household survey data with precise GPS coordinates, combined with multiple geospatial datasets, we develop a novel two-stage Random Forest model that first predicts roof materials from satellite imagery and then uses this prediction alongside other geospatial features to estimate household consumption levels. The model achieves high predictive accuracy across different spatial scales, with Spearman correlations between 0.81 and 0.87 when using regional characteristics and distance measures. A key finding is that broader regional information (10km radius) predicts household welfare more accurately than local characteristics (100m radius), while a parsimonious model using only characteristics of the roof of the building nearest to the interview location performs surprisingly well (0.74 or 0.83 with distances). This approach represents a cost-effective method for better understanding rural poverty and monitoring socioeconomic changes in rural areas, with potential applications across sub-Saharan Africa.