Multimodal Poverty Mapping and Geographic Transfer Allocation

Governments in developing countries deliver social assistance to millions each year, but often fail to reach the poorest households due to limited data and ineffective geographic targeting. We address this issue by combining multiple data sources at optimal spatial scales to predict poverty and apply these estimates to optimize resource allocation with beneficiary data from Zambia. Comparing Unimodal, Multimodal, and Stacked Multimodal approaches, we achieve an R^2 of 0.801 and generate a high-resolution 1 km by 1 km poverty map to estimate ward-level wealth. Simulating reallocation to the most deprived areas yields a 30% greater poverty severity (P_2) reduction than current practice. Yet, without full household-level data, poverty-focused optimization creates skewed aid distribution under uncertainty. We present egalitarian methods such as ‘‘min–max (water-filling)’’ optimization to balance coverage while targeting severe poverty. Given the weak alignment between current aid distribution and poverty, our study provides a framework for both evaluating and improving targeting.