The precise assessment of earthquake disasters requires high accuracy in population spatialization data. With the rapid development of technologies such as remote sensing and social sensing (social perception), and the growing availability of large-scale geographic data, high-resolution remote sensing data and vast geographical spatial data provide new opportunities for inferring more accurate population data. This paper proposes a method that integrates social perception big data, building attribute data, and multi-source remote sensing data. Using Random Forest, XGBoost, LGBM, Gradient Boosting, AdaBoost, and CatBoost models as base learners, and RidgeCV regression as the secondary learner, a higher-performing Stacking ensemble learning prediction model is constructed. The method system is employed to spatialize the population data of Yuxi City for the year 2020 at a 50m resolution using the Partition Density Mapping technique. The results show that: (1) The population estimation model proposed in this paper outperforms the WorldPop dataset and LandScan dataset by more than 40% in terms of accuracy at the administrative village scale, providing higher simulation accuracy for regions with different population densities. This result highlights the refined spatial features of the population depicted in satellite remote sensing images, offering richer and more realistic population distribution information. (2) By measuring the SHAP feature contributions and importance of the variables, it is found that residential building area, POI, and nighttime light intensity are the most significant indicators of population distribution in Yuxi City, suggesting a higher correlation of social factors in population spatial redistribution.
Zhang et al. (Fri,) studied this question.