A scalable Earth observation-based machine learning framework for high-resolution mapping and uncertainty assessment...

Bangladesh remains highly vulnerable to climate-induced hazards that disproportionately affect child health, yet spatially explicit assessments of these risks remain limited. This study develops a scalable Earth observation-driven machine learning framework to map climate-sensitive child health vulnerability across Bangladesh at 1-km spatial resolution. Anthropometric and health indicators from the 2022 Bangladesh Demographic and Health Survey (DHS) were integrated with multi-sensor satellite-derived environmental variables to generate high-resolution vulnerability surfaces and quantify prediction uncertainty. Georeferenced data from 674 DHS clusters (n = 8784 children) were combined with environmental covariates, including Land Surface Temperature (MODIS), precipitation (CHIRPS), vegetation indices (NDVI and EVI), and population density. Three supervised learning algorithms-Random Forest, Gradient Boosting Machine, and XGBoost-were trained using 80% of the dataset, with hyperparameters optimized through 5-fold cross-validation. The ensemble framework achieved the strongest predictive performance (R² = 0.683; RMSE = 7.65), outperforming individual models by 3.8%. Thermal stress, rainfall variability, and ecological productivity emerged as the dominant environmental determinants of vulnerability, collectively explaining 62% of model variance. The resulting maps revealed pronounced spatial disparities, with Rangpur exhibiting the highest vulnerability levels (67.3), while comparatively lower scores were observed in parts of Barishal and Sylhet (39.1). Bootstrap-based uncertainty analysis using 1000 iterations produced a mean uncertainty index of 4.8 ± 1.2, with elevated uncertainty concentrated in topographically complex regions. Hotspot analysis identified vulnerable clusters encompassing approximately 6.3 million children under five. The proposed framework provides a transferable approach for precision-oriented climate-health surveillance in data-constrained regions.