Machine Learning-Based Approach for Fine-Grid Population Mapping for Disaster Risk Assessment and Mitigation: Projecting Present and Future Population Distributions in the Colombo Metropolitan Region, Sri Lanka

Accurate and fine-scale population distribution data are essential for effective disaster risk assessment and mitigation, particularly in rapidly urbanizing regions where exposure to natural hazards is intensifying. This study presents a framework for population mapping under present and future conditions by integrating conventional approaches with machine learning. The framework combines scientific observations, such as satellite imagery, with policy-driven information, including zoning plans, thereby reflecting local social, political, and regulatory contexts. Specifically, this study focuses on creating high-resolution (40 m) gridded population maps for two densely populated watersheds, Kalu Oya and Mudun Ela, in Sri Lanka. The population grids were generated via a dasymetric approach, which disaggregated census data from administrative boundaries with building footprint datasets for the current condition. Future population distributions for 2030, 2050, and 2070 were projected via random forest (RF) algorithms under shared socioeconomic pathways (SSP1–SSP5). The RF model, developed using 25 features, showed the highest sensitivity to the normalized difference vegetation index (NDVI), land cover associated with urbanization, and distances from roads, healthcare facilities, and rivers. Current population grids were used to train the model, and accuracy was evaluated via statistical indicators, yielding an R² of 0.64, an MAE of 0.80 per/grid, and an RMSE of 1.06. The model was further enhanced by incorporating future development data extracted from official urban development plans for 2030. The outcome of this study provides a comprehensive framework for data-scarce region to generate high-resolution population maps useful for disaster risk management during flood and other natural disaster using publicly available dataset.