• Proposing a regional-distinctive geographic factors (RDGF)-incorporated approach for fine-grained population prediction. • After incorporating RDGF, the model’s R2 increased by 6%, and the relative error was markedly reduced in high-density population areas. • The contribution of RDGF to the model exhibits a more pronounced geographically decaying pattern in space and can significantly reduce prediction errors. Predicting fine-grained population distribution is crucial for effective urban planning. However, existing models widely ignore Region-Distinctive Geographical Factors (RDGF) in regional population modeling. This omission may compromise prediction accuracy, particularly in coastal zones where over 50% of the global population. To address this gap, we proposed an RDGF-incorporated approach for fine-grained population prediction, using the coastal Pearl River Delta as a case study. Leveraging multi-source geospatial data, based on generalized geographical factors (GGF) (e.g., topography, POI density, nighttime light intensity, etc.), we supplemented multi-dimensional RDGF including ecology, agriculture and transportation, etc. derived from unique regional environments (e.g., distance to shoreline, aquaculture, ports, etc.). We employed an interpretable machine learning framework (Random Forest + SHAP) to model and explain factor contribution. Results demonstrate: (1) incorporating RDGF substantially improves prediction accuracy in both model performance (with the average R 2 increasing by 6% under spatial cross-validation) and output (The relative error in densely populated areas can be reduced by up to 40%), thereby providing opportunity for more effective infrastructure planning and disaster risk management. (2) GGF still make the primary contribution to the model; however, RDGF are able to reveal local spatial heterogeneity and geographic decay patterns in population distribution, demonstrating greater potential for reducing prediction errors. This study provides region-specific insights for generating large-scale, fine-grained population map.
Gao et al. (Thu,) studied this question.