SFobNet: An Improved Swin Transformer Integrating Urban Functional Zones for Object-Level Building Height Estimation from Sentinel-2 Images

Building height is a critical parameter in modeling urban spatial structures, serving as an essential basis for interpreting urban form and evaluating spatial efficiency. Despite the remarkable progress in building height estimation, most existing methods still rely on Convolutional Neural Network (CNN) and perform pixel-level estimation using physical features extracted from remote sensing imagery. However, these approaches often struggle to capture global structural patterns, fail to represent height heterogeneity at the individual building level, and overlook the intrinsic relationship between building functional types and height. To address these issues, this study proposes a novel deep learning method called the improved Swin Transformer integrated with urban Functional zones for Object-level Building height estimation Network (SFobNet). The proposed method utilizes an improved Swin Transformer to accurately extract the local and global features of buildings, effectively reducing systematic bias through the integration of urban functional zones, thereby achieving consistent representation of height information at the individual building level. The experimental results indicated that SFobNet achieved superior validation accuracy in Beijing, with R 2 = 0.7155 and RMSE = 8.0889 m, reducing error by 9.4% compared with the state-of-the-art SEASONet and showing clear advantages over other baseline models. Cross-city evaluations on Tianjin and Shijiazhuang further confirmed its generalization performance, achieving R 2 = 0.5058 and RMSE = 11.1161 m, while consistently outperforming SEASONet. Ablation experiments further verified the effectiveness of the proposed method in addressing the aforementioned challenges. In conclusion, SFobNet significantly enhances the precision and robustness of object-level building height estimation, offering a particularly promising and solid methodological foundation for future large-scale urban 3D morphological reconstruction.