| 摘要: |
| 本研究选取位于成都平原的郫都区作
为研究区域，基于“生产—生活—生态”空间
理论构建了乡村聚落人居环境质量评价指标体
系。选取了50 个村级行政区进行分析，并使
用地理信息技术（GIS）计算了人口密度的空间
数据。通过张流量平台（TensorFlow）、梯度提
升决策树（GBDT）和SHAP 解释模型，构建
了影响人口密度空间数据的相关性模型。研究
结果表明，人均耕地面积、年生产总值以及年
平均降雨量对人口密度空间数据是正向促进作
用；而村民年人均可支配收入、NDVI （归一化
植被指数）以及水体和林地覆盖率对人口密度
空间数据是负向抑制作用；生活空间子系统中
的其他因素影响几乎可以忽略不计。本研究为
理解乡村振兴战略实施后，成都市郫都区乡村
聚落人居环境质量要素对乡村聚落社会的影响
提供了新的研究方法和视角，也为进一步提升
成都平原周边乡村聚落人居环境质量以及优化
人口空间分布提供了数据支撑。 |
| 关键词: 乡村聚落 人居环境质量 人口密
度 梯度提升决策树 SHAP解释模型 |
| DOI：10.13791/j.cnki.hsfwest.20241207003 |
| 分类号: |
| 基金项目:（通讯作者）：西南交通大学设计艺术学院，教授，shubo@swjtu.edu.cn |
|
| The impact of human settlement environment quality elements on spatial populationdensity in rural settlements of Pidu District, Chengdu Plain: An analysis based on gradientboosting decision tree and SHAP interpretation model |
|
LIU Yang,SHU Bo
|
| Abstract: |
| This study examines how human settlement environment quality shapes the spatial
distribution of population density across 50 village-level units in Pidu District on the Chengdu
Plain. Grounded in the “Production-Living-Ecological” (PLE) space framework, it builds a
comprehensive index system capturing key dimensions of rural habitat quality. Pidu is a pertinent
case: the Chengdu Plain is both a critical agricultural base and a rapidly urbanizing region, making
it ideal for exploring how production conditions, living amenities, and ecological endowments
jointly influence where people live. Using GIS, the authors compute spatial population density as
the dependent variable. The analytical pipeline integrates TensorFlow for modeling, Gradient
Boosting Decision Trees (GBDT) for prediction, and SHapley Additive exPlanations (SHAP) for
interpretation, yielding a powerful, interpretable model of complex non-linear relationships. Data
collection combines statistical yearbooks, remote sensing imagery, and field surveys. GBDT is
chosen for its ability to handle high-dimensional, non-linear interactions and resist overfitting—
traits crucial for mixed environmental and socio-economic features. Model training employs
parameter tuning and cross-validation to optimize performance. SHAP then provides a transparent
decomposition of predictions into feature-level contributions, delivering both global and local
interpretability and improving on traditional importance measures by ensuring consistency and
additivity. Model validation shows high predictive accuracy. On the production and ecological
fronts, three variables stand out with positive associations to population density: per capita arable
land area, annual gross production value, and average annual precipitation. Villages with more
farmland per person, stronger economic output, and ample rainfall tend to host denser populations.
In the Chengdu Plain’s agricultural context, higher precipitation directly supports yields,
reinforcing local production capacity and sustaining settlement concentrations. Conversely, several
variables display negative relationships with density: per capita disposable income, NDVI, and the
coverage of water bodies and forests. The income result likely reflects a mobility transition
dynamic: higher household incomes expand choices and can catalyze out-migration toward urban
opportunities, reducing rural density in origin villages. Elevated NDVI and extensive forest/water
coverage may denote areas prioritized for ecological conservation or constrained by natural
features, limiting buildable land and infrastructure expansion. These patterns underscore a planning
trade-off: stronger ecological protection can coincide with lower settlement concentration,
necessitating careful spatial zoning to balance conservation with habitation needs. Notably, most
indicators in the “living” subsystem. Exhibit minimal influence on density at this analytic scale.This suggests that, within Pidu, production capacity and first-order ecological conditions are the dominant drivers of where people cluster, outweighing finergrained
differences in residential comfort. The finding challenges the common assumption that living-environment upgrades alone can strongly reshape rural
population patterns; instead, it emphasizes the primacy of economic opportunity and foundational environmental suitability. Policy implications are substantial.
Methodologically, the study demonstrates how combining machine learning with explainable AI can illuminate the determinants of rural population distribution
in a way that is both accurate and decision-relevant. Practically, the positive ties between production indicators and density point to strategies that secure
agricultural productivity, support local industry, and leverage rainfall-driven advantages. Meanwhile, the negative links with ecological indicators caution
planners to calibrate development intensity in ecologically sensitive zones, using conservation-aware layouts and targeted service provision to avoid
undermining environmental goals. The study’s contribution is both substantive and methodological. Substantively, it clarifies the socio-environmental levers of
rural population density on the Chengdu Plain in the wake of Rural Revitalization Strategy, showing production and ecological baselines as decisive.
Methodologically, it offers a transparent, data-driven workflow that enhances trust and actionability. Future work could generalize across regions, incorporate
panel data to trace temporal dynamics, and test policy shocks or infrastructure upgrades to observe causal shifts. Together, these steps would deepen
understanding of how habitat quality and social behavior co-evolve in rural China, informing sustainable, balanced development around metropolitan
hinterlands like Pidu. |
| Key words: rural settlements habitat quality population density gradient boosting decision tree SHAP interpretation model |
