Scale Heterogeneity Avoided Dasymetric Mapping for the Gridded Population

The gridded population, crucial for resource allocation and emergency support, is mainly downscaled from the census data with administrative divisions. A common dasymetric mapping approach is building a regression model between aggregated geospatial properties and population potential at the administrative level and then applying this model directly to the grid level. The aggregation of geospatial properties often relies on statistical methods like averaging. However, the difference in scale between the two levels can lead to the heterogeneity of geospatial properties, which causes a gap between the training domain and the target domain and makes these methods fail to preserve the physical meaning of geographic properties. To address this issue, we propose a deep learning-based approach, in which a sophisticated loss function involving tripartite elements, gridded geospatial properties, gridded population potential, and administrative population potential, is designed. In this way, scale heterogeneity both in aggregation and domains can be avoided. In this study, a 30-meter resolution population density map of Hong Kong is produced through the proposed approach. The validation result shows that compared with both the machine learning-based or the artificial neural network-based one, the proposed approach gets a lower RMSE and potentially provides a more accurate reference for detailed urban management.