A novel framework for urban geometry rapid reconstruction utilizing high-resolution stereo satellite imagery for wind environment assessment

Rapid and accurate construction of regular urban building models, which can be directly used in computational fluid dynamics (CFD), is of great significance for urban wind environment assessment. However, traditional methods for generating building geometry, which rely on manual modeling or on-site point cloud acquisition, consume massive amounts of time for data collection and struggle to meet the demand for rapid city-scale reconstruction in emergency situations. To address this issue, this paper proposes a framework for rapidly generating urban geometric models suitable for CFD simulations utilizing high-resolution stereo satellite imagery. Taking Shenzhen as an example, the GaoFen-7 (GF-7) stereo images were first processed through image fusion and ortho-rectification to construct a semantic segmentation dataset. Subsequently, the remote sensing mamba (RS-Mamba) network was trained to extract building contours. Simultaneously, disparity was estimated using the digital surface model network (DSM-Net), and point clouds were generated via a forward intersection algorithm. These point clouds were then projected into a digital surface model (DSM) to calculate building heights. To satisfy the geometric quality requirements for CFD, this study developed a contour simplification and regularization algorithm to rapidly generate high-quality Level of Detail 1 (LoD1) building and vegetation models at the urban scale. Finally, validation across Shenzhen and Dongguan using UAV-LiDAR yielded R²=0.91, MAE=2.72m, and RMSE=4.09m, outperforming SGM, MGM, and CSF/Top-hat methods. These results demonstrated that the proposed framework effectively mitigates GF-7 tailing effects and ensures a numerically stable foundation for high-fidelity urban wind environment assessment.