融合多源空间数据的城镇人口分布估算

精细尺度的城镇人口空间分布是分析人类-资源-环境相互关系的重要指标。本文提出了一种融合地理空间大数据和高分辨率遥感数据估计精细尺度城镇人口分布的方法。通过对比各指标与人口相关性,选取R2>0.7的建筑面积、到道路距离、夜间灯光强度、商服中心、EAHSI指数、幼儿园、公园、小学、加油站、医院、公交车站、长途汽车站作为影响人口分布的变量因子。结合城市功能区数据确定人口分布区域,利用随机森林模型对宁波市2018年人口数据进行了500 m格网空间化,从而得出宁波市城镇人口空间分布图。最后,基于随机森林模型的变量因子重要性分析宁波市人口空间分布的影响因素。研究结果表明,本文所提出的城镇人口分布模型在街道尺度的估算精度为81.2%,平均相对误差MRE为0.29、RMSE为3279.89;网格级别的MRE为17.16,RMSE为1149.9,因此模型能精确地反演城镇内部街道人口分布信息。通过对变量因子重要性进行比较,发现建筑面积重要性约为0.22,对宁波市人口估算影响最大;到道路的距离、夜间灯光强度、商服中心、EAHSI（Elevation-Adjusted Human Settlement Index）、幼儿园、公园对宁波市人口估算具有重要作用。本文在格网级别进行的人口分布精度验证对于研究城市精细人口分布具有重大意义。

Estimating Population Distribution in Cities and Towns though Fusing Multi-source Spatial Data

The finer-scale spatial distribution of population within cities and towns is of great significance for studying the human-resource-environment interrelationships and supporting smart city construction and resource allocation. It also helps the government to assist disaster assessments and land use planning, manage the distribution of population and resource, and promote urban sustainable development. However, existing population spatialization methods are insufficient to spatialize population in cities and towns at fine scales. With the rapid development of geospatial big data and the popularity of high-resolution remote sensing data, this study proposes a method to estimate urban population distribution at fine scales through fusing multi-source spatial data. First, a total of 12 variables having large correlations （R2>0.7） with the population were selected to estimate the population distribution in Ningbo city, including the build-up area, distance to the road, nighttime lights, business service center, EAHSI index, kindergarten, park, primary school, gas station, hospital, and bus station and coach station. First, the population distribution areas are determined by urban functional zones, then a random forest model was used to train a population estimation model with the selected 12 variables; finally, the 2018 population data of the Ningbo were redistributed into 500 m grids by the trained estimation model. The importance of the chosen variables were analyzed using the random forest model. The results demonstrate that the presented population estimation model reaches an accuracy of 81.2% at sub-district scale with the MRE of 0.29 and the RMSE of 3279.89. Therefore, the population estimation model presented in this study can accurately predict the population distribution at the sub-district. This study also conducted the accuracy verification at the grid scale with the MRE of 17.16 and the RMSE of 1149.9. According to the importance of variables computed by the random forest model, it is found that the importance of the variable building area is about 0.22, which has the largest influence on the population distribution, followed by the variables, distance to road, nighttime lights, business service center, EAHSI ( Elevation-Adjusted Human Settlement Index), kindergarten, and park. The accuracy verification at the grid level is of great significance for studying the fine population distribution in cities. However, the estimation accuracy is still not very high in some cases where the populations of some grids are either overestimated or underestimated. The lack of building height information is a possible reason. In addition, deep learning methods will be explored to improve accuracy in future.