Deep learning-based building height mapping using Sentinel-1 and Sentinel-2 data

Abstract

Accurately mapping building height at a fine scale is crucial for comprehending urban systems. However, existing methods suffer from limitations such as coarse resolutions, long delays, and limited applicability for large-scale mapping. This challenge is particularly significant in China, where rapid urbanization has led to complex urban scenario. To address this issue, we propose a novel approach that capitalizes on publicly available Sentinel-1/-2 and crowdsourced data. Our method employs a dual-branch structure building height estimation network (BHE-NET) and an improved multi-modal Selective-Kernel (MSK) module to fuse optical and SAR features. The validation results, derived from building height data across 63 cities, demonstrate strong performance with a root mean square error (RMSE) of 4.65 m. We further test the scalability of our approach by mapping three most developed urban agglomerations in China. In comparison with four recent studies, our method captures finer details of building height while mitigating the overestimation in urban high-density building clusters. Moreover, we further investigate the relationship between population and mean building height as well as the building volume at city level. Our work opens up new possibilities for producing fine-scale building height map of China at a 10-m resolution using remote sensing data.

Keywords

• Building height
• Sentinel
• Deep learning
• Urban agglomeration
• Data fusion