世界银行-利用地理空间数据对贫困和财富进行小面积估算：到目前为止我们学到了什么？（英）-2023.6-30页_429kb

Geospatial data, including satellite imagery and derived features, have proven highly effective in predicting wealth and poverty across geographic areas, with correlations often exceeding 0.7 and R² values ranging from 0.37 to 0.95 in multiple studies. Wealth is generally easier to estimate than consumption metrics like per capita expenditure. While various methods (e.g., linear mixed models, CNNs, gradient boosting) have been used, tree-based machine learning models like XGBoost and random forests often outperform traditional approaches in accuracy and uncertainty estimation, especially out-of-sample. Key challenges include accuracy degradation in non-sampled areas due to informative sampling and sensitivity to training data quality. Integration with household surveys enhances predictions, and geospatial data can improve social assistance targeting, particularly when combined with targeting systems. Future research should focus on standardizing evaluations, incorporating time-series predictions, developing uncertainty estimation tools, and making advanced techniques accessible in developing contexts.