Objectives Existing population spatialization methods mainly use administrative-unit-level data to train regression model, and transfer it to grid cell-level to achieve population allocation. However, the significant scale difference between the analytical units in training and estimation leads to the issues of cross-scale model transfer. Meanwhile, only the attributes of current cell are considered in cell-level feature modeling, which causes the innate spatial association between cells to be eliminated and cells to be isolated.
Methods This paper proposes a novel population spatialization based on random forest by considering pixel-level attribute grading and spatial association (PAG-SA). In the cell-level feature modeling, we firstly construct the night light grading features embedded with building category constraints based on natural breaks, and count the grid proportion of each grading level at the administrative-unit-level as the training input to reduce the cross scale error; secondly, the influence and distance attenuation of neighborhood point of interests (POIs) upon the current cell is modelled by using kernel density estimation; thirdly, based on overlay analysis, the numbers of POIs in the contours of different building types are counted to improve the precision of feature modeling.
Results To verify the effectiveness of the proposed method, we selected Wuhan city as the experimental area and compared its spatialization accuracy with the datasets of WorldPop, GPW and PopulationGrid_China at street scale. The results show that the mean absolute error of PAG‐SA is only 1/6-1/3 of the comparison datasets. In addition, the influence of feature composition, grid size and kernel density bandwidth on the accuracy is also discussed.
Conclusions By fusing multi‐source data and considering pixel‐level attribute grading and spatial association, the proposed method PAG‐SA is effective for achieving population spatialization in urban areas with finer grid sizes and higher accuracy. It can also provide references for spatialization applications of other geographic attributes that also face with scale mismatch issue in spatial regression modeling.