
Dr. Woojin Jung's latest journal article, "Using Artificial Intelligence/machine learning to evaluate the distribution of community development aid across Myanmar," written by Dr. Woojin Jung, Dr. Saeed Ghadimi, Dr. Dimitrios Ntarlagiannis, and Andrew H. Kim is published in Socio-Economic Planning Sciences, volume 98 (paid access to the article can be found here).
Article Highlights
• A single modality of daytime satellite imagery predicts wealth effectively.
• A majority of villages have predicted wealth above the median.
• Gaps exist between the explicit selection criteria and the allocation of aid amount.
• Our best model predicts over 70% of the aid per capita per community.
• Village participation and geo-ethnic considerations are associated with aid size.

Key Findings
The study examines the distribution of community development aid in Myanmar using machine learning and satellite imagery to estimate village-level poverty and predict aid allocation. It compares three poverty estimation methods—nighttime luminosity, RGB pixel intensities, and Convolutional Neural Networks (CNNs) applied to daytime satellite imagery, with CNNs providing the most accurate poverty estimates. While poorer villages receive more aid per capita, wealth alone is a weak predictor. Instead, village capacity (a community’s ability to incorporate participatory projects) and state/ethnicity factors significantly influence aid distribution. Aid is disproportionately allocated to ethnic minority regions such as Chin and Magway, while wealthier states like Mandalay and Nay Pyi Taw receive less. Other important factors include village remoteness and distance from violent conflict, suggesting that logistical access and security concerns play roles in aid allocation.



The results indicate that machine learning models, particularly XGBoost, outperform traditional regression methods in predicting aid distribution, with the best model (XGBoost Infrastructure 50+) explaining 71.3% of the variance in aid per capita allocation. By comparison, SVR achieved R² = 0.613, and ElNet R² = 0.509. Overall, full models perform better than parsimonious models. The most influential predictors were state/ethnicity (6 of the top 10 features), infrastructure access, and community participation, while wealth ranked much lower in importance (58th out of 65 features in XGBoost Infrastructure 50+). The study also finds that aid allocation tends to avoid wealthier areas rather than explicitly target the poorest (as indicated in the Shapley plot figure), highlighting the need for a more transparent and data-driven approach to aid distribution.

Comments