A three-weight surface modeling approach for optimizing small-scale population disaggregation

In recent decades, gridded population data at fine scales has become a popular data source for assessing and monitoring the Sustainable Development Goals (SDGs). However, current population disaggregation methods are facing challenges in generating high-precision population grids for small areas with limited data. To fill this gap, we proposed a lightweight population gridding method that combines basic dasymetric mapping and point-based surface modeling, named three-weight surface modeling. In this method, there are three weights designed to describe the population spatial heterogeneity from different perspectives. The first weight is building-volume weight, which is equivalent to the preliminary results of population assignment based on building volume information. The second weight, POI-center weight, incorporates POI categories and aggregation patterns to express the centers with high population density, which is calculated based on the neighborhood accumulation rule of Spearman's correlation coefficients between POIs and population size. The third weight called POI-distance weight, indicates different rates of population decay with distance from high-density centers. The three-weight surface model allows us to dynamically adjust the parameters so as to correct the building-volume weight according to the remaining two POI-related weights for a more accurate population surface. After analyzing the census population and the disaggregation results of 544 villages in three counties (Huishui, Luodian and Pingtang) in southern Guizhou Province, China, we found that the customized three-weight model constructed using the local parameter groups demonstrated better accuracy performance compared to separate dasymetric mapping or point-based surface modeling. Meanwhile, the 10-m population grid generated by the local parameter model (LPTW-POP) exhibited higher resolution and lower errors (RMSE, MAE and MRE) than widely used gridded population datasets like LandScan, WorldPop and GHS-POP.