“The use of optical and SAR remote sensing images for building height estimation is of great significance for understanding urban form and optimizing urban stock space. However, existing datasets have many limitations: due to the small number of samples, it is difficult to meet the requirements of remote sensing information extraction based on deep learning, and the area covered by the samples is limited, which cannot provide sufficient geographical diversity and spatial feature representativeness, especially for large-scale building height datasets in China. In addition, the insufficient openness of the dataset limits its application and validation in a wider range of research. To address these issues, this paper constructs a building height dataset based on Sentinel images for deep learning BHDSI（Building Height Estimation Dataset Based on Sentinel Imagery）， This dataset covers the central urban areas of 62 cities in China, with a total of 5606 samples, covering urban, rural, and other scenarios. It is currently the largest building height dataset in China's regional coverage area. This dataset contains remote sensing images of Sentinel-1 and Sentinel-2, as well as the true values of building height. The sample size is 256 × 256, which provides an important supplementary option for building height estimation research compared to the 64 × 64 dataset. Compared to other datasets, this dataset has the characteristics of large sample size, wide coverage, availability, and reasonable distribution of building heights, which can better meet the training needs of deep learning networks. On this basis, this article evaluates the BHDSI dataset and other similar datasets using the same deep learning network, and compares the performance of multiple networks in building height regression tasks when using the BHDSI dataset, deeply analyzing the advantages and disadvantages of each network. The results indicate that compared to other datasets, the BHDSI dataset performs better in building height regression tasks. Further analysis reveals that when using the BHDSI dataset, the estimation accuracy is relatively high in areas with lower building heights. In addition, the U-Net decoder can achieve higher accuracy when used for training building height estimation networks. In summary, the BHDSI dataset provides important support for future research in the field of building height estimation.”