Leveraging Multimodal Deep Learning Models for Poverty Mapping and Geographic Transfer Allocation

Background and purpose: Data scarcity and limited administrative records hinder efforts to map poverty and target aid in developing countries. The integration of multimodal data can improve data sparsity. Key questions remain about how to select and integrate features to best predict outcomes. Prediction is also not the end goal but real-world policy impacts should be measured. This study develops multimodal approaches to create granular national poverty maps and optimize the regional allocation of a social safety net program. We draw on agricultural input transfer data from Zambia. This is the first study to reallocate agricultural input transfers using real-world data at the smallest administrative level. The study offers a novel contribution by presenting a prediction-to-allocation pipeline rather than focusing solely on diagnosing poor targeting.

Methods: We predict granular poverty using multimodal data, allocate budgets based on these predictions, and evaluate the resulting impacts on poverty reduction and geographic coverage. Three models, Unimodal, Multimodal, and Stacked Multimodal, are trained to estimate the wealth index of 535 georeferenced village clusters, using features from satellite imagery, social media, and crowd-sourced geographic information. Features are extracted through deep learning at varying buffer sizes (1–10 km). After comparing three models, we adapt the Stacked Multimodal approach, which outperforms the alternatives. Next, we estimate the number of poor households within the ward. To link predicted wealth indices to consumption, we linearly map wealth index data to the log of household consumption data using available samples from ground truth surveys. Building on the first step, we use the Stacked Multimodal model to estimate the mean wealth index in a ward. We then estimate the mean and variance of wealth to specify the probability distribution of household consumption, and estimate the number of poor households below the poverty line. Based on this poverty rate, we reallocate the Food Security Pack (FSP) across three optimization strategies, i) poverty-focused, ii) Boltzmann, and iii) min-max. Optimization performance is then evaluated using two criteria: the percentage change in 𝑃2, following FSP transfer and the number of recipient wards.

Results: The best prediction models explain 76.0%–80.1% of wealth variance, with the Stacked Multimodal achieving an R² of 80.1%. Infrastructure features demonstrate the strongest predictive performance. Relative to the status quo, optimization methods reduce P₂ by approximately 20–30%. The poverty-focused approach achieves the greatest reduction in poverty severity (7.34%) but covers only 9.5% of wards. We therefore present more egalitarian methods, such as “min-max (water-filling)” optimization, to balance coverage while targeting severe poverty. The min-max approach achieves a 6.78% reduction across four times more wards.

Conclusion and Implications: This study demonstrates that multimodal inputs improve poverty prediction and that current allocation methods often fail to distribute aid equitably. Our approach supports pro-poor geographic targeting through high-resolution poverty maps and concrete FSP caseload estimates per ward. Limitations include the absence of direct consumption measures, incomplete FSP ward coverage, and the assumption of stable wealth over time, pointing to future research combining geographic and household-level targeting.