Cambodia has grown rapidly yet remains economically vulnerable, with persistent poverty and limited, outdated subnational data. This study combines big-data sources, machine learning, and the Cambodia Socio-Economic Survey to predict and map the Global Multidimensional Poverty Index across 10 indicators in education, health, and living standards at fine spatial scales. By integrating gridded deprivation probabilities with building footprints, we estimate household-level deprivations. A random-forest model attains high accuracy for clean water, sanitation, food consumption, housing materials, cooking fuel, and electricity access. Key predictors include nighttime lights, population density, and road networks. Challenges persistβespecially the need for unbiased training data and limited capacity to capture within-province or within-district disparities. Nevertheless, the approach shows how big data and machine learning can complement traditional surveys to deliver more granular and timely measurements on multidimensional poverty.
Notes: Cambodia has seen strong growth but poverty remains. The study applies a multidimensional approach aligned with the Global MPI to capture deprivations beyond income, focusing on education, health, and living standards.
Notes: The aim is to integrate spatial and survey data using AI/ML to produce detailed poverty maps. This helps policymakers allocate resources efficiently and identify local vulnerabilities.
Notes: Prior research shows satellites and ML can help predict poverty, but integration with socioeconomic surveys for multidimensional poverty is limited. This study fills that gap.
Notes: A wide set of data was used: CSES for household info, EO data for environment and infrastructure, and building footprints to scale down predictions to household level.
Notes: The Random Forest algorithm was selected due to robustness and ability to process mixed data types. Models produce probability maps that can be aggregated at township, district, or province level.
Health (2): Food consumption, access to healthcare
Education (2): Attainment, school attendance
Living Standards (6): Cooking fuel, sanitation, water, electricity, housing, assets
Notes: Ten indicators were chosen following the Global MPI. Equal weights applied across three main dimensions. These indicators reflect SDG priorities like education, health, clean water, and energy.
Notes: Nightlights and population density best explain deprivation. Infrastructure access is also crucial. Indicators with spatial correlation (e.g., utilities) performed better than those tied to household-specific conditions.
Notes: Spatial maps show concentration of deprivation in remote, poorly connected regions. Urban and border areas with infrastructure show lower poverty.
Notes: While promising, ML struggles when data lack spatial correlation. Improved survey design can enhance integration. This hybrid approach shows potential for real-time, fine-grained poverty monitoring.
Notes: This work shows how AI and EO data complement traditional surveys to map multidimensional poverty. Future directions include advanced spatial analysis and deep learning models for better accuracy.