基于随机森林模型的西藏人口分布格局及影响因素

在乡镇尺度下厘清人口分布格局及其影响因素与区域差异,对在生态脆弱区制定可持续发展政策具有重大指导意义。基于2010年西藏自治区的乡镇尺度人口普查数据,提取人口密度和空间因子,利用空间统计方法分析了人口分布的疏密特征和集聚特征,对比运用多元线性回归方法和随机森林回归方法探索该地区人口分布的影响因素及其区域差异。结果表明：① 西藏乡镇人口密度在空间上表现出极强的非均衡性,其总体趋势是东南高西北低,高密度区与大江大河及主要交通干线具有较强的空间耦合性;② 大致以波绒乡（聂拉木县）—岗尼乡（安多县）为西藏的人口分界线,人口集聚的“核心—边缘”特征明显;③ 多元线性回归方法中,人造地表指数对人口分布的影响程度最大,随后依次为夜间灯光指数和路网密度;④ 利用随机森林方法进行的人口密度预测比多元线性回归方法精度高,可以用来对影响因子的重要性进行排序;排序在前六位的影响因子由高到低依次为夜间灯光指数、人造地表指数、路网密度、工业总产值、GDP和多年平均气温,它们与人口密度均呈正相关关系;地形地貌要素中以海拔和坡度的贡献率最大且与人口密度均呈负相关关系;⑤ 西藏人口分布格局的影响因素及其相互作用呈现出明显的区域差异特征,河谷是西藏地区人口的集聚区,主要分布在拉萨河谷、年楚河谷以及三江河谷;⑥ 通过随机森林回归分析,可以利用概念模型来表达人口分布影响因素,将主导因素概括为土地利用结构、道路通达度及城镇化水平。

Population distribution pattern and influencing factors in Tibet based on random forest model

Clarifying the spatial pattern of population distribution, its influencing factors and regional differences at the township level is of great guiding significance for formulating sustainable development policies in ecologically fragile areas. Based on the population census data of Tibet at the township level in 2010, the population density and spatial factors were extracted. The density and clustering characteristics of the population distribution were analyzed by spatial statistical method. The multiple linear regression method and the random forest regression method were used to explore the population influencing factors and their regional differences of population distribution. The results showed that: (1) The population density of Tibet at the township level showed a strong spatial non-equilibrium. The general trend was high in the southeast and low in the northwest, and there was a strong spatial coupling between the main rivers and the main traffic trunks in high density area. (2) The "core-edge" characteristic of population clustering was obvious, and roughly to the wave of Borong (Nyalam County)-Gangni (Anduo County) as the demarcation line. (3) In the multiple linear regression method, the artificial surface index had the greatest influence on the population distribution, followed by the nighttime light index and road network density. (4) Random forest method was more accurate than multiple linear regression method to predict the population density, which can be used to sort the importance of the influencing factors. The influencing factors of the first six factors were the night light index, artificial surface index, road network density, industrial output value, GDP and multi-year average temperature, and these factors were positively correlated with population density. Among topographic factors, the contribution rate of elevation and slope was the largest, which was negatively correlated with population density. (5) The influencing factors and their interactions of population distribution in Tibet showed obvious regional differences. The valley was a gathering area for population in the study region, mainly in Lhasa River Valley, Nianchu River Valley and Sanjiang River Valley. (6) Through the analysis of random forest regression, the conceptual model can be used to express the influencing factors of population distribution, and the dominant factors were summarized as land use structure, road accessibility and urbanization level.