Developing an annual building volume dataset at 1-km resolution from 2001 to 2019 in China

Abstract

ABSTRACTUrban vertical features are crucial for understanding urban morphology. However, long-term information on three-dimensional buildings, which are important fundamental data for studying on the historical urbanization processes, remains scarce in China. In this study, we proposed a Random Forest model to generate an annual 1-km resolution building volume dataset covering mainland China from 2001 to 2019, by integrating the nighttime light data, population demographics, electricity consumption records, carbon dioxide emissions data, and various optical and statistical datasets. This new building volume data are highly consistent with that derived from Baidu Maps on 1-km scale, with Pearson’s correlation coefficient (R) of 0.847, root mean square error (RMSE) of 9.17 × 105 m3/km2 and mean absolute error (MAE) of 5.86 × 105 m3/km2. Notably, cross-validation indicate that the blooming problem was greatly improved when compared with previous model-based building three-dimensional data. The proposed method holds significant advantages, benefiting form low-cost implementation based on free open-source data and providing extendable algorithm to estimate the 3D shape of cities in the future. The time-series historical building volume data offer comprehensive insights into the historical development of urban structures, and provide valuable fundmental data for future urban planning, urban climate models and land use projections.

Keywords

• Urbanization
• building volume
• time-series information
• China
• machine learning
• urban planning