Remote sensing-derived building footprint data has been widely applied in population mapping. However, nonresidential buildings typically occupy larger areas or volumes, and obtaining building footprint data with type information is also a great challenge. Furthermore, existing methods largely disregard the dynamic optimization and spatial heterogeneity of population estimation. To address these issues, we introduced a novel framework that estimates population distributions using multiple spatial prediction methods. First, a building habitability index (BHI) was constructed by integrating building footprints with residential quarters, road networks, and digital elevation models. Next, a BHI-based iterative method was proposed to dynamically optimize population distribution estimates. Finally, an ensemble approach combining the iterative method with a machine-learning model via a geographically weighted regression model was developed to produce 100-m gridded population maps that consider spatial heterogeneity. The results showed that the BHI-based iterative method substantially reduced population misallocation arising from reliance on building footprint data, with relative root mean square errors (rRMSEs) of 0.54 and 0.32, respectively, yielding smaller errors than those of benchmark models. The ensemble method further reduced the error in population estimation, particularly when combined with the extreme gradient boosting model, which achieved the lowest rRMSE of 0.23 (approximately one-third of that of WorldPop datasets). This study highlights the effectiveness of fusing remote and social sensing data for population mapping and the importance of employing multiple spatial prediction methods to build ensemble models that generate high-accuracy population distribution estimates.
Ma et al. (Fri,) studied this question.