Scholarlink

<p>Short-duration heavy rainfall (SDHR) refers to rainfall events with 1-h accumulated rainfall of no less than 20 mm, characterized by sudden onset, rapid development and strong destructive potential. Accurate forecasting of SDHR remains a global challenge. Data-driven artificial intelligence (AI) techniques provide new avenues for SDHR forecasting. However, high-quality training datasets needed by the AI techniques are still lacking. Based on the data from multiple sources, including observations from 158 radar stations, 48,000 rain gauges, and 3-km China Meteorological Administration (CMA) regional reanalysis product, the National Meteorological Information Centre (NMIC) of CMA has developed a large-volume, well-labeled AI training dataset for SDHR (AIDA-SDHR), with minute-level temporal resolution and kilometer-level spatial resolution. Data processing techniques such as data cleansing based on multi-source data cross-validation, sample labeling via segmented inverse distance-weighted interpolation, and targeted feature extraction, were employed. The AIDA-SDHR dataset covers 14,392 SDHR events that occurred in central–eastern China since 2016, with a total of 1,181,308 samples. Each sample is annotated with a 6-min accumulated rainfall intensity label and supplemented with 10 radar-derived features as well as 30 atmospheric state variables, enabling direct deployment for training of AI models. Evaluations show that based on the AIDA-SDHR dataset, quantitative precipitation estimation by an AI model (AI-QPE) outperforms the algorithm of the radar reflectivity–rainfall (<i>Z–R</i>) in capturing the spatial distribution and intensity of rainfall, reducing the root-mean-square error (RMSE) by 12.19%. Furthermore, integrating more samples from AIDA-SDHR into the Yushi AI forecasting model improves its performance, with increases of 4.68% and 15.69% in the threat score (TS) for composite reflectivity and extreme composite reflectivity 50 dBZ) forecasts at 0–60-min lead time. Overall, the AIDA-SDHR dataset paves the way for development of AI-based monitoring and forecasting of SDHR in China. Moreover, it also holds substantial potential for a deeper understanding of the formation and evolution of SDHR.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An AI Training Dataset for Monitoring and Forecasting of Short-Duration Heavy Rainfall in China

  • 期刊论文

  • 发表日期：

    2026年6月25日

查看全文
  • Na Liu,
  • Wenming Xiao,
  • Anyuan Xiong,
  • Yujia Liu,
  • Qiang Zhang,
  • Yingrui Sun,
  • Shuo Zhao,
  • Zhongyan Hu

摘要

Short-duration heavy rainfall (SDHR) refers to rainfall events with 1-h accumulated rainfall of no less than 20 mm, characterized by sudden onset, rapid development and strong destructive potential. Accurate forecasting of SDHR remains a global challenge. Data-driven artificial intelligence (AI) techniques provide new avenues for SDHR forecasting. However, high-quality training datasets needed by the AI techniques are still lacking. Based on the data from multiple sources, including observations from 158 radar stations, 48,000 rain gauges, and 3-km China Meteorological Administration (CMA) regional reanalysis product, the National Meteorological Information Centre (NMIC) of CMA has developed a large-volume, well-labeled AI training dataset for SDHR (AIDA-SDHR), with minute-level temporal resolution and kilometer-level spatial resolution. Data processing techniques such as data cleansing based on multi-source data cross-validation, sample labeling via segmented inverse distance-weighted interpolation, and targeted feature extraction, were employed. The AIDA-SDHR dataset covers 14,392 SDHR events that occurred in central–eastern China since 2016, with a total of 1,181,308 samples. Each sample is annotated with a 6-min accumulated rainfall intensity label and supplemented with 10 radar-derived features as well as 30 atmospheric state variables, enabling direct deployment for training of AI models. Evaluations show that based on the AIDA-SDHR dataset, quantitative precipitation estimation by an AI model (AI-QPE) outperforms the algorithm of the radar reflectivity–rainfall (Z–R) in capturing the spatial distribution and intensity of rainfall, reducing the root-mean-square error (RMSE) by 12.19%. Furthermore, integrating more samples from AIDA-SDHR into the Yushi AI forecasting model improves its performance, with increases of 4.68% and 15.69% in the threat score (TS) for composite reflectivity and extreme composite reflectivity 50 dBZ) forecasts at 0–60-min lead time. Overall, the AIDA-SDHR dataset paves the way for development of AI-based monitoring and forecasting of SDHR in China. Moreover, it also holds substantial potential for a deeper understanding of the formation and evolution of SDHR.