Abstract
Contaminated sites pose serious threats to the soil environment and human health. However, the location and temporal changes of urban contaminated sites across China remain unknown due to data scarcity. Here, we developed a machine-learning model to identify the contaminated sites using public data. Results show that the trained model with 2,005 surveyed site samples and six variables can achieve a model performance evaluation value of 0.86. 43,676 contaminated sites were identified from 83,498 polluting enterprise plots in China. However, these contaminated sites have significant spatiotemporal heterogeneity, mainly located in economically developed provinces, urban agglomerations, and core urban areas. Moreover, the contaminated sites increased by 325% along with urban expansion from 1990 to 2018. The abandoned contaminated sites increased rapidly, but the contaminated sites in production decreased continuously. This methodological framework and our findings contribute to the precise management of contaminated sites and provide insights into urban sustainable development.