This study takes the Melbourne Central Business District as the research area, focusing on the three stages before the epidemic, during the epidemic, and after the epidemic, and constructs a multi-source high-resolution data set that integrates the community scale, street scale, and time-weather dimensions. By collecting multi-platform multi-source data and preprocessing such as outlier removal and data distribution adjustment, 7 machine learning models were tested and the influence of features was analyzed with the help of the SHAP algorithm. Finally, the deep learning model (especially MLP) has the highest prediction accuracy, and the dependence of pedestrian flow on the built environment has shifted from being dominated by community-scale features before the epidemic to being dominated by street-scale features during and after the epidemic. The interaction of cross-scale features has been enhanced. At the same time, behavioral changes caused by the epidemic such as reduced reliance on public transportation and increased preference for green, open and walkable spaces are durable, which can provide empirical support for improving urban resilience and planning updates.
From neighborhoods to streetscapes: Pandemic-era shifts in built-environment effects on pedestrian mobility
从社区到街景：大流行时代建筑环境对行人流动性影响的变化
1.Introduction
2.Literature review
2.1. Built environment and pedestrian mobility
2.2 Impacts of COVID-19 on travel behavior
2.3 Evolution of data collection and modeling approaches
3. Data and methods
3.1 Research framework
3.2 Study area and dataset
3.2.1 Pedestrian count data
3.2.2 Weather and neighborhood environment data
3.2.3 Streetscape environment data
3.3 Data preprocessing
3.3.1 Temporal encoding and outlier removal
3.3.2 Data distribution adjustment
3.3.3 Buffer range determination
3.3.4 Multicollinearity check
3.4 Model development
3.4.1 Model selection
3.4.2 Evaluation metrics
4. Results and analysis
4.1 Model fitting and prediction error
4.2 Feature contributions across time and space
4.3 Differential effects across periods
4.4 Interaction effects between neighborhood and streetscape features
5. Discussion
5.1 Model performance
5.2 Built environment effects on pedestrian flow
5.3 Structural shifts in travel behavior
5.4 Implications for urban planning and policy
5.5 Limitations
6. Conclusion
研究内容
Research contents
该研究以墨尔本中央商务区为研究区域，聚焦疫情前、疫情中、疫情后三个阶段，构建整合社区尺度、街道尺度及时间 - 天气维度的多源高分辨率数据集，旨在探究步行流动与建成环境特征的交互作用及演变趋势，解决现有步行流动研究在特征选择、数据精度和空间尺度上的局限，为提升城市韧性、优化城市更新与规划策略提供实证支撑。
This study takes the Melbourne Central Business District as the research area, focusing on three stages before the epidemic, during the epidemic, and after the epidemic, and constructs a multi-source high-resolution data set that integrates community scale, street scale, and time-weather dimensions. It aims to explore the interaction and evolution trends of pedestrian flow and built environment characteristics, solve the limitations of existing pedestrian flow research in feature selection, data accuracy, and spatial scale, and provide empirical support for improving urban resilience and optimizing urban renewal and planning strategies.
研究方法
Research methods
研究通过多平台采集步行传感器、天气、社区环境及街道景观等多源数据，经异常值剔除、数据分布调整、缓冲区优化和多重共线性检验等预处理后，测试岭回归、KNN、随机森林、深度学习等 7 种机器学习模型，采用 MAE、RMSE、R² 等指标评估模型性能筛选最优模型，并借助 SHAP 算法量化特征贡献度、分析特征交互效应及不同阶段影响差异，增强模型解释性以揭示建成环境对步行流动的动态影响机制。
The study collected multi-source data such as walking sensors, weather, community environment and street landscape through multiple platforms. After pre-processing such as outlier removal, data distribution adjustment, buffer optimization and multi-collinearity testing, 7 machine learning models such as ridge regression, KNN, random forest and deep learning were tested. Indicators such as MAE, RMSE and R² were used to evaluate the model performance to select the optimal model, and with the help of SHAP The algorithm quantifies feature contribution, analyzes feature interaction effects and impact differences at different stages, and enhances model interpretability to reveal the dynamic impact mechanism of the built environment on walking flow.
研究结果
Research conclusions
研究发现深度学习模型（尤其是 MLP）在各阶段步行流量预测中表现最优，疫情期间非稳定场景下优势更显著；步行流动对建成环境的依赖呈现从疫情前社区尺度特征主导到疫情中及疫情后街道尺度特征主导的转变，且跨尺度特征交互作用增强；疫情引发步行行为结构性转变，包括对公共交通依赖降低、对绿色开放可步行空间偏好提升，这些变化在疫情后持续存在，呈现出持久影响。
The study found that deep learning models (especially MLP) performed best in pedestrian flow prediction at all stages, and the advantages were more significant in unstable scenarios during the epidemic; the dependence of walking flow on the built environment showed a change from community-scale features before the epidemic to street-scale features during and after the epidemic, and the interaction of cross-scale features increased; the epidemic triggered structural changes in walking behavior, including a decrease in reliance on public transportation and an increase in preference for green, open and walkable spaces. These changes persisted after the epidemic, showing a lasting impact.
