High-resolution population mapping by fusing remote sensing and social sensing data considering the spatial scale mismatch issue

High-resolution population data are crucial for various applications, from developing regional plans to disaster risk management. Current population spatialization methods typically apply population mapping relationships established at the regional level to the grid level using multi-source data. However, the significant scale difference between the regional and grid levels, combined with the simple integration of multi-source data features without considering the spatial dependence of the population, results in lower accuracy. To address the scale mismatch issue in the downscaling process, we first construct a spatially heterogeneous population label by combining census data with gridded population datasets. Then, we establish a relationship mapping between population covariates and population at a low-resolution scale (100 m) and apply it to a neighboring high-resolution scale (25 m) to reduce the prediction bias resulting from directly downscaling from the regional level to the grid level. Meanwhile, a deep learning model based on transformer feature attention convolution net (TFACNet) is employed to aggregate each geographic unit's global and local spatial relationships, integrating complementary features learned from multi-source heterogeneous data in an end-to-end manner. The experimental results in Wuhan and Guilin show that our method achieved a more accurate population spatialization (overall [Formula: see text] 0.92) at the street level.