基于图像语义分割与ELO算法的乡村视觉质量评价研究

A Study on Rural Visual Quality Evaluation Based on Image Semantic Segmentation and the ELO Algorithm

关键词:  风景园林  视觉偏好  ELO评价算法  视觉评价  乡村景观质量

Key words:  landscape architecture  visual preference  ELO evaluation algorithm  visual evaluation  rural landscape quality

摘要:

在建设和美乡村的政策背景下，如何提升乡村景观吸引力已成为推动乡村旅游可持续发展的关键议题。采用图像分割技术、ELO算法及XGBoostSHAP模型，分析各景观要素对视觉评价偏好的贡献度，旨在探讨乡村旅游中景观要素对游客视觉偏好的影响，进而为提升乡村景观视觉质量提供支撑。结果表 明：1)建筑、天空和乔灌木的重要性较高，是乡村景观环境中不可忽视的核心要素；2)水体与建筑对村落景观偏好具有促进作用，而铺装和田野地则起到抑制作 用；3)K-means聚类结果进一步揭示，高评分景观主要依赖视觉吸引力强的水体与传统赣派建筑，低评分景观则主要受到不协调建筑与杂乱田野地的负面影响； 4)ArcGIS分析结果显示，游客景观偏好集中在各村落群中心区域，而村落群中部的田埂、花海区域评分较低，景区提升需重点关注此类低评分区域。研究方法拓 展了现有乡村景观领域的视觉评价理论体系，为视觉偏好研究提供了新的证据支持，同时对针对性提升景区吸引力具有重要的实践指导价值

Abstract:

Under the policy framework of constructing beautiful and livable rural areas, enhancing the attractiveness of rural landscapes has become a critical and urgent task for advancing the sustainable development of rural tourism. A highquality rural landscape not only satisfies the aesthetic demands of tourists but also contributes to the economic, cultural, and ecological vitality of rural regions. However, systematic and quantitative assessments of the visual quality of rural landscapes, particularly at the element level, remain underexplored. To address this gap, this study combines advanced machine learning techniques with spatial analysis to identify, quantify, and interpret the contribution of individual landscape components to tourists' visual preferences. The study employed a non-linear XGBoost model, in conjunction with SHAP (SHapley Additive exPlanations) analysis, to conduct a detailed quantitative assessment of rural visual landscape elements extracted from high-resolution images. The image segmentation process categorized major visual elements - such as buildings, water bodies, pavements, farmland, sky, trees, and shrubs - into distinct categories. XGBoost was then employed to model the relationship between these variables and visual preference scores, while SHAP values provided an in-depth interpretation of each element's positive or negative influence on perceived landscape quality. Furthermore, we analyzed the interaction effects between elements to uncover how specific combinations might enhance or diminish aesthetic appeal. The results reveal that: 1) Buildings within villages consistently show high SHAP values, indicating that architectural elements, particularly those with traditional stylistic features, significantly enhance tourists' visual preferences. In contrast, large proportions of pavements and farmland tend to suppress preference scores. 2) Interaction analysis shows that buildings and water bodies exhibit a strong synergistic effect, collectively producing a positive impact on perceived attractiveness. Notably, when the proportions of farmland and pavements remain low, their SHAP values are positive - indicating that a moderate presence of these elements can contribute to visual diversity - whereas excessive coverage results in negative SHAP values, reflecting a detrimental impact. 3) K-means clustering further distinguishes high-scoring images, typically characterized by visually appealing water bodies and well-preserved Gan-style architecture, from lowscoring images, where discordant buildings and cluttered farmland dominate. 4) ArcGIS spatial analysis shows that tourists' preferences are spatially concentrated in the central areas of village clusters, where visual coherence is higher. Conversely, peripheral zones such as ridges, flower fields, and scattered agricultural plots within village clusters receive consistently lower ratings, indicating areas that should be prioritized for improvement. This integrated approach represents a methodological innovation, combining interpretable machine learning with spatial clustering and GIS-based mapping, thus delivering actionable insights for rural landscape enhancement. The application of SHAP enables a transparent interpretation of model predictions, bridging the gap between computational analysis and practical landscape planning. Our findings not only advance the theoretical framework of visual evaluation in landscape architecture but also provide empirical evidence for the development of targeted design and management strategies. Specifically, optimizing the spatial arrangement and proportion of key elements - such as buildings, water bodies, and vegetation - can significantly enhance the visual quality of rural landscapes without requiring substantial additional resource input. Overall, the study offers a replicable, data-driven methodology for visual landscape assessment and provides practical guidance for rural tourism development. By identifying both the strengths and deficiencies of existing landscapes, this research supports more informed decision-making, thereby contributing to the sustainable and aesthetic revitalization of rural environments.