SwinH-Fuse: A Dual-Stage Multisensor Transformer Framework for Urban Building Identification and Height Estimation

Accurate building footprint delineation and height estimation are critical for urban mapping, infrastructure monitoring, and risk assessment. We presentSwinH-Fuse, a hierarchical satellite remote sensing framework that integrates data from the European Space Agency's (ESA) Sentinel-1 synthetic aperture radar (SAR) mission and Sentinel-2 multi-spectral optical mission to jointly extract building footprints and estimate heights. The framework employs a dual-branch Swin-UNet transformer for city-scale classification, a UNet for neighborhood-scale refinement, log-scale regression, and bias correction for height estimation. Monte Carlo dropout is integrated to quantify predictive uncertainty. Across diverse test regions in Europe and North America,SwinH-Fuseachieved a height RMSE of 1.22 m with a Pearson correlation coefficient of 0.73 with reference values. In an out-of-sample case study on the King Abdullah University of Science and Technology (KAUST) campus, the regression module maintained strong correlation despite classification challenges in sparse desert conditions. The results demonstrate thatSwinH-Fuseprovides accurate, scalable, and uncertainty-aware urban mapping suitable for sustainable planning and risk assessment.