A key indicator of urban change is construction, demolition, and renovation. Although these development activities are often interrelated, they are typically studied independent of one another. Analytic methods relying on a strict set of modeling assumptions limit our ability to understand this change holistically. Machine learning has demonstrated the potential when combined with big data to discover patterns and relationships between seemingly unrelated variables. This research explores urban change through net construction, a composite value that treats demolition as a deductive process that is subtracted from construction activity which provides for a more holistic and nuanced understanding of development activity. Once validated through a visual analysis of its reliability as a measure of urban change, we then used a series of random forest regression models to evaluate the predictive accuracy of net construction compared with independent models of construction and demolition. Applying the approaches to an urban county in the United States, we compiled 122 independent variables to provide a comprehensive view of individual neighborhoods from multi-disciplinary data sources such as socioeconomic, built environment characteristics, and landscape metrics. We then analyze the feature importance scores derived from the random forest models in an effort to assess the similarities and differences between the variables that have the greatest influence on model accuracy. The net construction model produced more accurate results than models that used construction and demolition activity independently. While many of the most important features aligned with those from the independent models, land use mix drawn from landscape metrics appeared as the most important, representing a departure from previous studies. This study provides a scalable method for modeling urban change using machine learning techniques and reveals the importance of applying data-driven algorithms that can help communities become more informed about their pressing issues.

城市变化的一个关键指标是建设、拆除和改造。尽管这些开发活动通常是相互关联的，但它们的研究通常是相互独立的。依赖于一组严格建模假设的分析方法限制了我们全面理解这种变化的能力。机器学习已经证明了与大数据相结合的潜力，可以发现看似不相关的变量之间的模式和关系。本研究通过网络建设探索城市变化，一种综合价值，将拆除视为从建筑活动中减去的演绎过程，从而提供对开发活动的更全面和细致入微的理解。通过对其作为城市变化度量的可靠性的可视化分析进行验证后，我们然后使用一系列随机森林回归模型来评估网络建设的预测准确性与独立的建筑和拆除模型相比。将这些方法应用于美国的一个城市县，我们编制了 122 个自变量，以从社会经济、建筑环境特征和景观指标等多学科数据源中提供各个社区的综合视图。然后，我们分析从随机森林模型得出的特征重要性分数，以评估对模型准确性影响最大的变量之间的异同。净施工模型比独立使用施工和拆除活动的模型产生更准确的结果。虽然许多最重要的特征与独立模型中的特征一致，但从景观指标得出的土地利用组合似乎是最重要的，代表与以往研究的背离。这项研究提供了一种使用机器学习技术对城市变化进行建模的可扩展方法，并揭示了应用数据驱动算法的重要性，这些算法可以帮助社区更加了解他们的紧迫问题。