乡村景观偏好与环境特征的非线性影响效应——基于VR 和可解释性机器学习的新解读

Nonlinear Impact Effects of Rural Landscape Preferences andEnvironmental Features: A New Explanation Using VR andInterpretable Machine Learning

关键词:  风景园林  乡村景观偏好  VR  图像语义分割  非线性关系  可解释性机器学习

Key words:  landscape architecture  rural landscape preference  VR  image semantic segmentation  nonlinear relationship  interpretable machine learning

摘要:

在美丽中国与乡村振兴背景下，探索公众视觉审美偏好与乡村景观环境之间的关系、识别其关键影响因素，具备重要理论与实践价值。结合8种非线性 模型和SHAP可解释性机器学习技术，深入分析影响138名受试者乡村景观感知满意度的主客观因素及其非线性影响趋势和阈值效应。结果显示：1)PDP曲线验 证了感知满意度与环境特征之间存在非线性关系，多数环境特征存在激活水平和阈值点，而非简单的线性相关；2)主观感知特征和客观环境特征均显著影响受试 者的景观偏好，前者的影响等级往往超过后者；3)8种非线性模型在预测感知满意度的性能方面，XGBoost效果最好；4)7个主观感知指标(色彩满意度、环境 整洁度、独特性、宜居性、安全感、归属感、原真性)和13个客观环境指标(乔灌草比率、水域面积比例、房屋像素占比、道路面积比例、植被覆盖度、道路可达 性、公共活动空间面积、水景质量、围栏像素占比、高程、公众关注度、植物丰富度、蓝色视野指数)的标准化重要性值总和达96.27%，是影响受试者感知满意 度的关键因素。研究结果为理解乡村景观偏好与环境特征之间的非线性关系提供了新视角，为乡村景观提质优化过程中“精明增长”和“适度发展”的建设实践 提供理论参考。

Abstract:

Against the backdrop of the "Beautiful China" initiative and rural revitalization strategy, improving rural landscape quality and understanding public visual preferences have become critical issues in landscape planning and design. However, existing studies are still largely constrained by linear analytical assumptions, which fail to adequately capture the complex, dynamic, and threshold-based relationships between environmental features and human perceptual responses. This study, through the effective combination of 8 nonlinear models and SHAP interpretable machine learning technology, conducted an in-depth analysis of the subjective and objective factors influencing the landscape perception satisfaction of 138 participants and the nonlinear influence effects involved. The results revealed that: 1) PDP curves verified the general nonlinear relationships between satisfaction and environmental features, with most environmental factors exhibiting activation levels and thresholds, rather than simple linear correlations; 2) Both subjective perceptions and objective environmental features significantly influenced landscape preferences, with the former often having a higher importance ranking; 3) Among eight nonlinear machine learning models used to predict satisfaction, XGBoost performed the best, with R2=0.893; 4) Seven subjective perception features (color satisfaction, cleanliness, uniqueness, livability, safety, belonging, and authenticity) and 13 objective indicators (ratio of trees to grass, the ratio of water area, pixel ratio of buildings, the ratio of road area, vegetation coverage, road accessibility, public activity space area, water scenery quality, pixel ratio of fences, elevation, public attention index, plant richness, and blue view index) accounted for 96.27% of the total standardized importance, making them key factors affecting satisfaction. Overall, this study demonstrates that rural landscape preference is a multidimensional, threshold-sensitive, and nonlinear perceptual process jointly influenced by human perceptual responses and physical landscape characteristics. These findings provide a novel perspective and detailed insights for understanding the nonlinear relationship between landscape preferences and environmental characteristics. The results offer valuable implications for "smart growth" in the process of optimizing rural landscapes, supporting more precise, evidence-based, and human-centered rural planning strategies.