Urban growth unveiled: Deep learning with satellite imagery for measuring 3D building-stock evolution in Urban China

Time-series information on building stock is of paramount importance to study cities in a host of disciplines ranging from economics to urban planning. Such data are lacking in a consistently measured way and especially among dynamically growing cities in developing countries. Due to their rapid change, building stock data in these cities can offer insights into the determinants and consequences of urbanization. To be able to analyze urban structures effectively, the building stock needs to be measured with sufficient detail – at a resolution that makes individual buildings or small conglomerates thereof visible – and it needs to consider building height (or volume) with a satisfactory scope across cities to cover both large numbers and multi-year sequences of data. This study aims to develop a comprehensive pipeline for predicting building volume – including both footprint and height – across 1,537 urban areas in mainland China, covering more than 60% of the Chinese population over a seven-year period (2017–2023). With the advancement of deep learning in remote sensing, we can leverage state-of-the-art techniques to efficiently produce large-scale data for Chinese cities across years, which could be very time-consuming with traditional remote-sensing techniques. We compare the performance of several deep learning architectures for the task at hand. We demonstrate that the best performing approach leads to credible metrics of both footprint and height predictions and performs very competitively with respect to existing building-volume predictions. We also benchmark our results against other data sources such as real-estate listings and demonstrate the out-of-sample prediction capability of the proposed model. • Combine Planet Basemaps and Sentinel-1 to estimate building volumes in Chinese urban areas. • Cover 1537 urban areas over 7 years (2017–2023). • Compare predictions with real-estate data. • Compare predictions with ones from Global Human Settlement Layer and Random Forests. • Achieve strong out-of-sample prediction performance with medium-high resolution of < 5 m.