研究内容

Research contents

（1）背景与动机：COVID-19大流行对全球公共卫生和经济造成了巨大冲击。研究表明，建成环境在控制病毒传播中扮演着重要角色。然而，以往的研究大多集中在区域级建成环境上，对街道级建成环境（SLBE）的关注较少。本研究旨在填补这一空白，探讨SLBE对SARS-CoV-2传播风险（SSTR）的具体影响。

（2）研究区域与时间段：研究区域选择了香港，一个具有高密度和高度连通性的城市。研究时间段覆盖了香港疫情的前五波，从第一波到第五波早期。

（3）核心问题：研究主要关注以下三个问题：（1）如何更好地揭示SLBE与SSTR之间的关联？（2）哪些SLBE特征更具影响力，且它们在疫情不同波次中的影响是否有所不同？（3）从政策和管理角度来看，应优先关注哪些SLBE特征以有效降低街道级传播风险？

(1) Background and motivation: The COVID-19 pandemic has had a huge impact on global public health and economy. Research shows that the built environment plays 

an important role in controlling the spread of the virus. However, most previous studies focused on the regional-level built environment and paid less attention to the 

street-level built environment (SLBE). This study aims to fill this gap and explore the specific impact of SLBE on SARS-CoV-2 transmission risk (SSTR).

(2) Research area and time period: Hong Kong was selected as the research area, a city with high density and high connectivity. The research time period covers the first 

five waves of the Hong Kong epidemic, from the first wave to the early stage of the fifth wave.

(3) Core questions: The research mainly focuses on the following three questions: (1) How to better reveal the association between SLBE and SSTR? (2) Which SLBE 

characteristics are more influential, and do their impacts differ in different waves of the epidemic? (3) From a policy and management perspective, which SLBE 

characteristics should be prioritized to effectively reduce street-level transmission risks?

研究方法

Research methods

（1）从公开数据源下载了香港COVID-19确诊病例的访问和居住建筑信息。

（2）使用Google Street View（GSV）图像获取这些建筑周围800米步行服务范围内的SLBE特征。

（3）对GSV图像进行分类和物理特征分割，将SLBE特征分为多个类别，如街道绿化、街道建筑、街道人行道等。

（4）利用全局和局部空间自相关分析（如Moran’s I指数和LISA集群图）识别高风险（HH）和低风险（LL）聚类区域。

（5）采用随机森林算法（RFA）结合SHapley Additive exPlanations（SHAP）模型，以揭示SSTR与SLBE特征之间的非线性关联。

（6）比较了多种机器学习模型（如线性回归、支持向量机、梯度提升机等）的性能，发现随机森林模型表现最佳。

（7）还与地理加权回归（GWR）模型进行了比较，以评估不同模型在揭示SLBE与SSTR关联方面的差异。

(1) The visit and residential building information of confirmed COVID-19 cases in Hong Kong were downloaded from public data sources.

(2) Use Google Street View (GSV) images to obtain SLBE features within an 800-meter walking service range around these buildings.

(3) Classify GSV images and segment physical features, and divide SLBE features into multiple categories, such as street greening, street buildings, street sidewalks, etc.

(4) Use global and local spatial autocorrelation analysis (such as Moran’s I index and LISA cluster plot) to identify high-risk (HH) and low-risk (LL) cluster areas.

(5) Use the Random Forest Algorithm (RFA) combined with the SHapley Additive exPlanations (SHAP) model to reveal the nonlinear association between SSTR and SLBE 

features.

(6) Compared the performance of various machine learning models (such as linear regression, support vector machine, gradient boosting machine, etc.) and found that 

the random forest model performed best.

(7) were also compared with geographically weighted regression (GWR) models to evaluate the differences between different models in revealing the association of SLBE 

with SSTR.

研究结果

Research conclusions

（1）在高风险区域，街道人行道、街道卫生设施和人工结构是SSTR的主要风险因素。这些特征的存在和密度与SSTR呈正相关。

（2）在低风险区域，交通控制设施与SSTR显著正相关，而其他一些特征（如街道未铺路面）则与SSTR负相关。

（3）不同疫情波次中，SLBE特征对SSTR的影响存在差异。例如，在疫情初期，自然特征对SSTR的影响较大；而在疫情高峰期，街道基础设施的影响更为显著。

（4）随机森林模型在解释SSTR与SLBE之间的非线性关系上表现最佳，尤其是在高风险区域。其预测精度和稳健性均优于其他模型。

（5）GWR模型在局部区域的预测性能较好，但在高风险区域的表现不如随机森林模型。

（6）基于研究结果，提出了几项具体的政策建议，包括优先管理高风险区域的街道卫生设施、合理增加低风险区域的街道障碍物以减少SSTR、关闭高风险区域的户外商业和广告设施等。

（7）这些建议旨在通过优化SLBE来降低SARS-CoV-2的传播风险，为城市规划者和政策制定者提供了有价值的参考。

(1) In high-risk areas, street sidewalks, street sanitation facilities, and artificial structures are the main risk factors for SSTR. The presence and density of these features 

are positively correlated with SSTR.

(2) In low-risk areas, traffic control facilities are significantly positively correlated with SSTR, while some other characteristics (such as unpaved streets) are negatively 

correlated with SSTR.

(3) There are differences in the impact of SLBE characteristics on SSTR in different epidemic waves. For example, in the early stages of the epidemic, natural features had

 a greater impact on SSTR; while at the peak of the epidemic, the impact of street infrastructure was more significant.

(4) The random forest model performs best in explaining the nonlinear relationship between SSTR and SLBE, especially in high-risk areas. Its prediction accuracy and 

robustness are better than other models.

(5) The GWR model has better prediction performance in local areas, but its performance in high-risk areas is not as good as the random forest model.

(6) Based on the research results, several specific policy recommendations were put forward, including prioritizing the management of street sanitation facilities in 

high-risk areas, rationally increasing street barriers in low-risk areas to reduce SSTR, and closing outdoor commercial and advertising facilities in high-risk areas. wait.

(7) These recommendations aim to reduce the risk of SARS-CoV-2 transmission by optimizing SLBE, providing a valuable reference for urban planners and policymakers.