Harnessing Community-Informed Geospatial Data to Identify Food Insecure Households

Background & Purpose: Allocating welfare resources effectively remains a significant challenge. This issue is exacerbated in developing countries with inadequate or sparse poverty data. Traditional targeting methods often fail to classify those most at risk, especially in the context of high poverty and limited administrative capacity. While recent advances have employed new data sources and machine learning (ML) techniques, these generally predict asset- or income-based poverty and cover broad administrative regions with a standard machine-driven approach. To bridge this gap, this study develops a method that combines the community’s visual perceptions of poverty with geospatial data to identify food-secure and impoverished households at a granular level.

Methods: In collaboration with Innovations for Poverty Action, a survey of 300 households in Lusaka, the capital city of Zambia, was administered in 2023. We created a composite Wealth Score using asset indicators alongside Food and Nutrient Consumption Scores. The methodology also involves qualitative analysis of open-ended questions regarding visual signs of poverty. Additionally, we analyzed geospatial data from OpenStreetMap (OSM) data, the Normalized Difference Vegetation Index (NDVI), and the Nighttime lights from satellite images. Themes from the community’s visual perceptions of poverty informed the selection of relevant spatial features from OSM and Satellite imagery. We compared the performance of two feature selection approaches: i) a model employing only qualitatively selected features and ii) a data-driven model utilizing Principal Components. Several AI/ML algorithms, including ensemble and neural networks, were implemented to predict wealth, food, and nutrition scores as well as to classify vulnerable households.

Results:

The analysis revealed that food and nutrition scores are less predictable by wealth indicators and demographic data alone, highlighting the need for additional data. Conversely, wealth scores showed moderate predictability from food indicators and demographic data alone (Wealth score R2=0.701). Incorporating spatial data improved accuracy in predicting both food, with a great impact on food scores (13%) than wealth scores (11%). Models leveraging Point-of-Interest IPOI) data correctly classified both food-insecure and poor households 86% of the time (with an AUC of 0.86 and an f1 score of 0.75). Notably, models using community-selected features outperformed those using PCA components in predicting nutritional insecurities, although PCA was more effective in predicting food consumption. Prominent community indicators include road quality and building size. Poor sanitation and high population density are associated with poverty, whereas superior building features and infrastructure are associated with wealth. In contrast to these detailed visual indicators, simplistic subjective ratings of neighborhoods do not have predictive power.

Conclusions & Implications:

To conclude, integrating contextualized spatial features improves the identification of food-insecure and poor households in small communities. The findings suggest that policymakers may reach a larger segment of the food-insecure population by incorporating additional data. Additionally, the viability of community-informed feature selection supports the potential of local insights to refine data-driven models, reduce computational costs, and improve model transparency. This research contributes to the application of computer vision in designing social welfare policy.