Mapping the dimensions of poverty through big data, socioeconomic surveys and machine learning in Cambodia

Khoun Theara, Poortinga Ate, Thwal Soe, Gonzalez de Alba, McMahon Andrea, Carlos Mendez

October 2025

Abstract

Cambodia has grown rapidly yet remains economically vulnerable, with persistent poverty and limited, outdated subnational data.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 persist—especially 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.

Type

Journal article

Publication

Social Indicators Research

🌏 Introduction

Rapid economic growth, yet persistent poverty (17.8% below national line in 2019)
Traditional poverty data: outdated, costly, and coarse
Poverty: not only income but health, education, and living standards (MPI framework)

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.

📊 Research Objectives

Use big earth data + CSES survey + machine learning
Map 10 poverty indicators across 3 MPI dimensions
Generate high-resolution poverty maps
Support targeted, cost-effective policy interventions

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.

📚 Literature & Motivation

Household surveys = costly, infrequent, spatially coarse
Nighttime lights & satellite imagery → proxies for poverty
Machine Learning (RF, XGBoost, CNNs) improve predictions
Gap: Few studies integrate survey + EO data for multidimensional poverty

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.

🗂️ Data Sources

CSES survey (10k households) – health, education, housing, income
Satellite & EO data – nightlights, land cover, population density
Infrastructure data – roads, schools, hospitals, utilities
Building footprints – 3.8M residential/commercial buildings

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.

⚙️ Methodology

Random Forest model for classification
Predicts deprivation probability for each indicator
Training & validation split (90/10)
Outputs: household & regional deprivation maps

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.

📑 MPI Indicators

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.

📈 Results – Variable Importance

Nighttime lights = key predictor across indicators
Population density & road networks also significant
Strongest predictions: cooking fuel, clean water, sanitation, electricity
Weak predictions: school attendance, healthcare, assets

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.

🗺️ Results – Spatial Poverty Patterns

Urban centers: Phnom Penh, Siem Reap, Battambang → low deprivation
Remote provinces: Preah Vihear, Ratanakiri, Mondulkiri → high deprivation
Poverty lower near main roads & borders (trade effects)

Notes: Spatial maps show concentration of deprivation in remote, poorly connected regions. Urban and border areas with infrastructure show lower poverty.

💡 Discussion

Spatial ML useful but limited for indicators with weak spatial signals
Household survey data not designed for ML → location approximation issues
Need for richer survey integration (e.g., accessibility questions)
EO + ML offer granular, dynamic poverty mapping

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.

✅ Conclusion

10 MPI indicators mapped using EO + survey + ML
Best results for infrastructure-related deprivations
Enables household-level poverty estimates
Supports SDGs: No Poverty, Quality Education, Health, Clean Water, Energy
Future research: spatial autocorrelation, inequality decomposition, advanced AI

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.