<b>Characterizing dynamics of building height in China from 2005 to 2020 based on GEDI, Landsat, and PALSAR data</b>

The unprecedented urbanization in China has driven rapid urban and rural development in recent decades. While existing studies have extensively focused on horizontal urban expansion, research on vertical urban expansion patterns in China remains limited. To address this gap, we proposed a Multi-Temporal Building Height estimation network (MTBH-Net) to estimate building heights at a 30 m spatial resolution in China for 2005, 2010, 2015, and 2020 by integrating Global Ecosystem Dynamics Investigation (GEDI), Landsat, and PALSAR data. Specifically, we introduced sample migration to generate reference building height data and utilized the Continuous Change Detection and Classification (CCDC) disturbance feature to ensure consistency in unchanged built-up areas. Validation with GEDI L2A V2 data demonstrated that MTBH-Net achieved RMSEs of 5.38 m, 5.73 m, 6.26 m, and 6.36 m for the respective years. Further validation with field-measured data and GF-7 building height data yielded RMSEs of 9.13 m and 10.99 m, respectively. The proposed 30-m China Multi-Temporal Building Height (CMTBH-30) dataset reveals an increase in average building heights in China from 10.48 m in 2005 to 11.37 m in 2020, reflecting an upward trend in urban development. Additionally, the standard deviation of building heights rose from 3.87 m in 2005 to 6.35 m in 2020, indicating increased height variation nationwide. Regional analysis from 2005 to 2020 shows significant vertical growth on stable impervious surfaces in Chongqing (+3.6 m), Guizhou (+3.0 m), and Qinghai (+3.0 m), while Macau (+14.9 m), Hong Kong (+13.9 m), and Guangdong (+13.5 m) experienced notable growth on newly expanded impervious surfaces. Minimal growth was observed in Jilin, Heilongjiang, and Xinjiang. CMTBH-30 offers a more refined and accurate depiction of building heights, effectively capturing height variations and mitigating the underestimation of high-rise buildings. It fills the gap in multi-temporal building height products. Overall, this study provides a new dimension for urban research and is valuable for urban planning, disaster management, and environmental sustainability.

近数十年来，中国史无前例的城市化进程推动了城乡的快速发展。现有研究已广泛关注横向城市扩张，但针对中国城市竖向扩张模式的探究仍较为有限。为填补这一研究空白，我们提出了多时序建筑高度估算网络（Multi-Temporal Building Height estimation network, MTBH-Net），通过整合全球生态系统动态调查（Global Ecosystem Dynamics Investigation, GEDI）、Landsat与PALSAR数据，以30米空间分辨率估算中国2005、2010、2015及2020年的建筑高度。具体而言，我们引入样本迁移方法生成参考建筑高度数据，并利用持续变化检测与分类（Continuous Change Detection and Classification, CCDC）扰动特征，保障未变化建成区的一致性。借助GEDI L2A V2数据开展的验证结果表明，MTBH-Net在对应年份的均方根误差（Root Mean Square Error, RMSE）分别为5.38米、5.73米、6.26米与6.36米。进一步通过实地测量数据与GF-7建筑高度数据开展验证，所得均方根误差分别为9.13米与10.99米。本研究构建的30米分辨率中国多时序建筑高度（China Multi-Temporal Building Height, CMTBH-30）数据集显示，中国建筑平均高度从2005年的10.48米升至2020年的11.37米，反映出城市发展的向上趋势。此外，建筑高度的标准差从2005年的3.87米升至2020年的6.35米，表明全国范围内建筑高度差异有所扩大。2005至2020年的区域分析显示，重庆（+3.6米）、贵州（+3.0米）与青海（+3.0米）的稳定不透水面区域实现了显著的竖向增长；而澳门（+14.9米）、香港（+13.9米）与广东（+13.5米）则在新增不透水面区域实现了显著增长。吉林、黑龙江与新疆的建筑高度增长则较为有限。CMTBH-30数据集能够更精细准确地刻画建筑高度分布，有效捕捉高度差异并缓解了对高层建筑的低估问题，填补了多时序建筑高度产品的研究空白。总体而言，本研究为城市研究提供了全新视角，其成果对城市规划、灾害管理与环境可持续发展均具有重要价值。