Building Height Estimation at 10m across multiple countries using Sentinel Time-Series

With the rapid shift of urban population to cities, urbanization monitoring has become essential to ensure sustainable development. In the last decade, 2D urban monitoring such as building footprint extraction has received considerable attention resulting in multiple high and low-resolution products. But despite being the essential component of urbanization, the vertical dimension (height) has not been studied at a large scale. Accurate estimation of building height plays an important role in urban planning, as it is correlated with energy consumption, population, transportation, city planning, urban climate and many other monitoring and planning required for sustainable development.

Airborne LiDAR or high-resolution orthophotos can be used for accurate building height estimation but for large-scale monitoring applications, the data collection itself is extremely expensive. With a compromise of resolution, Earth observation data, especially free-of-cost data can be used for large-scale monitoring. Existing large-scale building height estimation methods operate at low resolution (1km to 100m). A few of the recent studies improved the resolution to 10m while operating in a few cities to few states of the country. In this study, we estimate building heights across four countries. We propose a DL model that operates on a time series of Sentinel-1 SAR and Sentinel-2 MSI data and estimates building height at 10m spatial resolution. Our model estimates building height with 1.89m RMSE (Root Mean Square Error) surpassing the best score of 3.73m reported in previous studies. 

To demonstrate the effectiveness of our approach, we tested it on data from four countries and compared it with a baseline and four recent DL networks. We evaluate the impact of time series input and individual input modality i.e., SAR and optical data on the performance of the proposed model. The model is also tested for generalizability. Furthermore, the predicted building heights are downsampled and compared with GHSL-Built-H R2023A, a state-of-the-art product at 100m spatial resolution. The results show an improvement of 0.3m RMSE.

References

[1] Yadav, R., Nascetti, A., & Ban, Y. (2022). BUILDING CHANGE DETECTION USING MULTI-TEMPORAL AIRBORNE LIDAR DATA. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XLIII-B3-2022, 1377–1383. https://doi.org/10.5194/isprs-archives-xliii-b3-2022-1377-2022

[2] Yang, C., &; Zhao, S. (2022). A building height dataset across China in 2017 estimated by the spatially-informed approach. Scientific Data, 9(1). https://doi.org/10.1038/s41597-022-01192-x

[3] Cai, B., Shao, Z., Huang, X., Zhou, X., & Fang, S. (2023). Deep learning-based building height mapping using Sentinel-1 and Sentinel-2 data. International Journal of Applied Earth Observation and Geoinformation, 122, 103399. https://doi.org/10.1016/j.jag.2023.103399

[4] Yadav, Ritu, Andrea Nascetti, and Yifang Ban. "A CNN regression model to estimate buildings height maps using Sentinel-1 SAR and Sentinel-2 MSI time series." arXiv preprint arXiv:2307.01378 (2023)

[5] Pesaresi, M., and P. Politis. "GHS-BUILT-H R2023A—GHS Building Height, Derived from AW3D30, SRTM30, and Sentinel2 Composite (2018)." (2018)