The article focuses on five waves of epidemics in Hong Kong from January 2020 to March 2022, and studies the impact of street-level construction environment (SLBE) on SARS-CoV-2 spatial transmission risk (SSTR). Research focuses include street environmental factor analysis, spatial heterogeneity and dynamic epidemic response. Research methods include data collection and processing, collecting case-related building data and obtaining images through Google Street View to segment street environmental elements, and correlating street environmental characteristics and building transmission risks; spatial analysis, using global and local spatial autocorrelation to identify the spatial aggregation mode of the epidemic and divide the building transmission risk cluster; machine learning modeling, using random forest (RF) model to analyze the nonlinear relationship between SLBE and SSTR, explaining the importance of features and interactions through the SHAP model and comparing the performance of other models. The research results show that street sidewalks, sanitary facilities, and artificial structures in high-propagation areas are positively correlated with transmission risks, while traffic control facilities and unpaved roads are negatively correlated with risks; street infrastructure is more significant in the third and fourth waves of the epidemic, and natural factors play a significant role in the first wave; the SHAP model reveals the threshold effect and interaction of environmental characteristics; the RF model performs best in high and low transmission risk clusters, significantly better than the traditional GWR model.
研究内容
Research contents
文章聚焦于香港街道层面的建成环境（Street-level Built Environment, SLBE）对SARS-CoV-2空间传播风险（SSTR）的影响，覆盖了2020年1月至2022年3月的五波疫情。研究重点包括：
街道环境要素分析：如人行道（SS）、卫生设施（SSF）、人工结构（AS）、交通控制设施（TCF）、未铺设道路（SURS）等对传播风险的作用。
空间异质性：通过空间自相关分析将病例相关建筑划分为高传播风险集群（HH）和低传播风险集群（LL），探讨不同区域的环境特征差异。
动态疫情响应：分析不同疫情波次中环境因素对传播风险的动态影响。
The article focuses on the impact of Street-level Built Environment (SLBE) on SARS-CoV-2 spatial transmission risk (SSTR) at the street level in Hong Kong, covering five
waves of the epidemic from January 2020 to March 2022. Research priorities include:
Analysis of street environmental factors: the role of sidewalks (SS), sanitation facilities (SSF), artificial structures (AS), traffic control facilities (TCF), unpaved roads
(SURS), etc. on transmission risks.
Spatial heterogeneity: Case-related buildings are divided into high transmission risk clusters (HH) and low transmission risk clusters (LL) through spatial autocorrelation
analysis, and the differences in environmental characteristics in different regions are explored.
Dynamic epidemic response: Analyze the dynamic impact of environmental factors on transmission risks in different epidemic waves.
研究方法
Research methods
数据收集与处理：
收集3693个COVID-19病例相关建筑的地理坐标及时间分布数据。
通过Google街景（GSV）获取84,045张图像，利用PSPNet和Mask R-CNN算法分割街道环境要素（如绿化、建筑、人行道等）。
基于800米步行服务范围，结合欧氏距离衰减权重，将街道环境特征与建筑传播风险关联。
空间分析：
采用全局和局部空间自相关（Moran’s I和LISA）识别疫情的空间聚集模式。
将建筑划分为高传播风险（HH）和低传播风险（LL）集群。
机器学习建模：
使用随机森林（RF）模型分析SLBE与SSTR的非线性关系，并通过SHAP模型解释特征重要性及交互作用。
对比线性回归（LR）、支持向量机（SVM）、地理加权回归（GWR）等模型的性能，验证RF的优越性。
Data collection and processing:
Data on geographical coordinates and temporal distribution of buildings related to 3693 COVID-19 cases were collected.
84,045 images were obtained through Google Street View (GSV), and street environmental elements (such as greening, buildings, sidewalks, etc.) were segmented using
PSPNet and Mask R-CNN algorithms.
Based on the 800-meter walking service range and combined with the Euclidean distance attenuation weight, the street environmental characteristics are correlated with
the risk of building transmission.
Spatial analysis:
Global and local spatial autocorrelation (Moran’s I and LISA) were used to identify spatial aggregation patterns of epidemics.
Buildings are divided into high transmission risk (HH) and low transmission risk (LL) clusters.
Machine Learning Modeling:
The nonlinear relationship between SLBE and SSTR was analyzed using the random forest (RF) model, and the feature importance and interaction were explained through
the SHAP model.
Comparing the performance of models such as linear regression (LR), support vector machine (SVM), geographic weighted regression (GWR), and other models, we verify
the superiority of RF.
研究结果
Research conclusions
关键环境风险因素：
高传播区域（HH）：街道人行道（SS）、卫生设施（SSF）、人工结构（AS）与传播风险显著正相关；街道障碍物（SO）在低密度时增加风险。
低传播区域（LL）：交通控制设施（TCF）、未铺设道路（SURS）与风险负相关，绿化（SG）的抑制作用较弱。
动态差异：
街道基础设施（如SSF、AS）在第三、第四波疫情中影响更显著，而自然要素（如SURS）在第一波作用明显。
SHAP模型揭示了环境特征的阈值效应（如SSF密度>0.5%时风险陡增）及交互作用（如AS与TCF协同降低风险）。
模型性能：
RF模型在HH和LL集群中均表现最优（测试集R²分别为0.86和0.66），显著优于传统GWR模型（R²=0.74）。
Key environmental risk factors:
High transmission areas (HH): Street sidewalks (SS), sanitation facilities (SSF), and artificial structures (AS) are significantly positively correlated with transmission risk;
street obstacles (SO) increase the risk at low density.
Low transmission areas (LL): Traffic control facilities (TCFs), unpaved roads (SURS) are negatively correlated with risks, and the inhibitory effect of greening (SG) is weak.
Dynamic Differences:
Street infrastructure (such as SSF, AS) has more significant impacts in the third and fourth waves of the epidemic, while natural elements (such as SURS) have obvious
effects in the first wave.
The SHAP model reveals the threshold effects of environmental characteristics (such as the risk increase sharply when the SSF density is >0.5%) and interactions (such
as the risk reduction of AS and TCF in concert).
Model Performance:
The RF model performed optimally in both HH and LL clusters (test set R² is 0.86 and 0.66, respectively), significantly better than the traditional GWR model (R²=0.74).
