融合地面实景图像分析与遥感影像数据的景观视觉偏好评价模型研究

A Model for Assessing Landscape Visual Preference by Integrating Ground-level Image Analysis andRemote Sensing Data

关键词:  风景园林  景观视觉偏好  Elo评分算法  景观特征  遥感影像  随机森林

Key words:  landscape architecture  landscape visual preference  Elo scoring algorithm  landscape feature  remote sensing imagery  random forest

摘要:

：景观审美价值评估对生态保护与可持续规划至关重要，但传统方法存在主客观割裂、尺度局限等问题。以井冈山风景区为案例，提出融合地面实景照片 与遥感影像数据的景观视觉偏好评估框架，通过Elo评分算法量化173张实地照片的美学偏好，结合DeepLabV3+语义分割模型提取景观要素的像素特征，并整合 Sentinel-2遥感数据与土地利用信息构建随机森林预测模型。主要贡献在于构建了多源数据融合模型，其预测精度显著提升(R2=0.49)，其中图像特征主导预测 (Pic_forest，importance=22%，r=0.48)，天空图像特征(importance=18%)与森林遥感特征(r=0.40)协同影响审美偏好，建筑要素呈显著负效应(r =-0.21)。 关键发现包括：1)识别出高美学得分(Elo>0.75)的典型条件为“遥感森林覆盖率≥80%且地面影像中天空像素占30%~40%、森林像素≥35%”，表明审美偏好 依赖于自然要素在宏观与微观尺度的双重协调；2)DeepLabV3+迁移学习实现较高精度(mIoU=0.72)有效解析景观构成要素。通过耦合像素级视觉解析与景观 格局分析，突破了传统美学评价的尺度壁垒，所提框架虽基于山地景区验证，但可推广至高原梯田、滨水生态带及城市山地边缘等多种地貌，为景观的生态保护 与可持续开发提供了可量化的空间决策依据。

Abstract:

Landscape visual preferences assessment is crucial for ecological conservation and sustainable spatial planning, as landscape aesthetic value emerges from the interplay of biophysical characteristics and human perception and can significantly influence human well-being and support for conservation. However, traditional approaches often suffer from a disconnect between subjective scenic preferences and objective measures, as well as limited spatial scales of analysis. This study addresses these gaps by developing an integrated methodology that combines ground-level photography with remote sensing data to quantitatively evaluate scenic landscape quality, applied to a mountainous scenic area (Jinggang Mountain, China). A total of 173 ground-level photographs were collected to represent diverse landscapes, and an Elo scoring algorithm was used to derive comparative aesthetic preference scores from observer evaluations of these images. Key visual features were then extracted from each photograph using a DeepLabV3+ semantic segmentation model to identify major landscape components (sky, forest, water, grass/field, and buildings), achieving a high segmentation accuracy (mIoU=0.72). The proportion of each component in the images was quantified as a set of micro-scale visual metrics, which were combined with macro-scale landscape variables derived from Sentinel-2 satellite imagery and land use data. Using this fused dataset, a Random Forest regression model was constructed to predict the Elo-based aesthetic score of each scene. The integrated model achieved a reasonable predictive performance (R2=0.49), outperforming separate models that used only image features or only remote-sensing features and underscoring the value of multiscale data fusion. Analysis of feature importance indicated that image-derived variables dominated the prediction: most notably, the fraction of forest visible in the photo was the strongest predictor (importance 22%, r=0.48), and other significant image features included the visible sky fraction (18% importance) and grass/field cover (12%), all positively correlated with aesthetic appeal. In comparison, the most influential remote-sensing factor was the overall forest cover in the landscape (importance 8%, r=0.40). Built-environment indicators, by contrast, showed negative influence on scenic quality (e.g., a higher proportion of buildings in view corresponded to lower scores, r=–0.21), suggesting that human-made structures generally detract from perceived beauty. Based on the model results, high-scoring scenes (Elo>0.75) were predominantly found where remote sensing forest cover exceeded 80% and the ground-view composition included roughly 30%~40% open sky and at least 35% forest pixels. This finding highlights that human scenic perceptions are maximized when natural features are plentiful and harmoniously represented at both the macro-scale (landscape context) and micro-scale (immediate view), demonstrating a strong coupling between overall landscape pattern and on-site visual experience. The successful use of DeepLabV3+ and other AI-based analysis in this framework shows the potential of advanced computer vision techniques to precisely quantify visual attributes of landscapes. By coupling pixel-level visual analysis with geospatial landscape data, the proposed framework effectively bridges scale-related gaps in landscape aesthetic evaluation. The practical implications are significant: this quantitative visual assessment tool can support planners and conservationists in identifying and prioritizing high-scenic-value areas for protection, inform sustainable tourism planning by pinpointing picturesque sites and routes, and guide the placement of development or infrastructure in ways that minimize visual intrusion in sensitive landscapes. Future research should validate this framework in other geographic regions to ensure its generalizability and should explore integrating broader ecological indicators (such as biodiversity metrics or cultural ecosystem services) to further strengthen the link between visual quality and holistic conservation goals.