顾及POI人口吸引力异质性的城市人口空间化方法

人口空间化是提升人口统计数据空间分辨率的常用手段，现有研究多基于统计建模思想建立多源数据与统计人口的数学模型以预测格网人口。兴趣点（Point of Interest, POI）作为精细人口估算的重要数据源，通常以数量/密度型指标形式参与回归建模，该方式忽略了类型相同但个体规模不同的POI与人口之间数量关系的差异，特征均质化处理造成POI语义细节的损失，导致中心城区人口低估与远城区高估。为此，本文基于随机森林模型，提出一种顾及POI人口吸引力异质性的城市人口空间化方法。该方法在表征POI空间多尺度重要性的基础上，引入移动定位数据构建人口吸引力指标；并基于非欧式滤波修正格网人口权重，建模人口空间自相关，刻画水体等障碍物对局部空间连通性的影响。本文以武汉市为研究区域开展100 m格网验证，通过与POI密度型回归模型、公开人口数据集的对比和消融实验，展现了人口吸引力指标与权重修正的有效性。结果表明，本文方法平均绝对误差为WorldPop、GPW及对比模型的1/4～2/3，在精细人口空间化场景具有精度优势。此外，本文还讨论了移动定位数据采样率及格网粒度对建模精度的影响。

Urban Population Spatialization by Considering the Heterogeneity on Local Resident Attraction Force of POIs

Population spatialization is a common method to refine the spatial resolution of census data. Existing studies are mostly based on the idea of statistical modeling to establish the association between ancillary data and population at the administrative-unit-level, and then transfer it to predict the gridded population. As an important data input for fine-grained population estimation, Point of Interests (POIs) are usually in the form of quantity or density indexes for regression modeling, which ignores the heterogeneity in the association between population and POIs with same type but different sizes. Such modeling methods cause the loss of semantic details of POIs, in turn leading to the population underestimation in main urban areas and overestimation in suburban areas. To tackle this problem, this paper proposes an urban population spatialization method based on random forest model by considering the heterogeneity of population attraction of POIs. More specifically, on the basis of establishing a multi-scale representation of the spatial importance of POIs, this method constructs population attraction indexes by integrating mobile positioning data. Meanwhile, the spatial autocorrelation of population is modeled based on non-Euclidean spatial filter for weight correction, which considers the influence of obstacles such as water body on local spatial connectivity. We select Wuhan city as the study area to conduct population spatialization experiment at 100 m spatial resolution. Through the comparison with traditional density-based model, popular gridded datasets, and the ablation experiments, the results verify the effectiveness of population attraction indexes and weight correction. The mean absolute error of our method is about 1/4-2/3 of the WorldPop, GPW, and the comparison model (i.e., Ye's model), demonstrating the advantages of our method in fine-grained population spatialization. In addition, the influences of the sampling rate and size of grid of mobile positioning data on the modeling accuracy are also discussed.