| 摘要: |
| [目的]为应用多普勒天气雷达数据客观定量分析致洪强降雨回波特征，根据不同回波类型，有针对性地采取措施，规范化、标准化发布预警信息。[方法]利用无监督机器学习中的K-均值聚类分析方法，对2005-2022年呼和浩特地区68次洪涝灾害事件中的4个关键特征(回波强度、面积、梯度以及强回波面积占比)进行自动学习和分组分析。[结果]结合业务实践制定了致洪强降雨天气雷达回波的分类标准，标准将回波分为4类:大范围强降水、中尺度强降水、中尺度较强降水和稳定性降水回波。利用2023年6-8月呼和浩特降雨个例，对致洪强降雨分类标准进行验证。该分类标准在2023年7月2日的应用中，分析雷达回波面积增速与自动雨量计数据可知强降水主要发生在强回波面积增速较大时间段，综合各雷达特征数据表明回波强度均值、强回波面积占比能更早地反映强降水的性质，对预警有指示作用。[结论]经过2023年的应用检验，基于机器学习的致洪强降雨分类标准具有业务应用价值。为准确识别不同类型的致洪降雨提供科学依据。 |
| 关键词: 雷达特征;预警指标;山洪暴发 |
| DOI： |
| 投稿时间：2024-04-22修订日期：2024-08-24 |
| 基金项目:鄂尔多斯市重点研发计划项目（YF20240033）；内蒙古自治区气象局科技创新项目（nmqxkjcx202427、nmqxkjcx202441、nmqxkjcx202420）；中国气象局复盘总结专项（FPZJ2025-023、FPZJ2024-020）。 |
|
| Research on Classification Method of Heavy Rainfall Causing Floods Based on Machine Learning |
| XU Jing,KONG Wenjia,ZHAOFei,YAO Xiaojuan,ZHANG Lianxia,YANG Haidi |
| (Ordos Meteorological Office of Inner Mongolia Autonomous Region;Inner Mongolia Meteorological Service) |
| Abstract: |
| To objectively and quantitatively analyze the echo characteristics of flood-inducing heavy rainfall using Doppler weather radar data, to take targeted measures based on different echo types, and to release standardized and normalized early warning information, this paper utilizes clustering analysis, an unsupervised machine learning method, to automatically learn and analyze four key features(echo intensity, area, gradient, and proportion of strong echo area) from 68 flood disaster events in Hohhot from 2005 to 2022, and performs grouping analysis based on these characteristics. A classification standard for weather radar echoes of flood-inducing heavy rainfall has been established. Based on operational practices, the echoes are subdivided into four major categories:large-scale heavy precipitation, mesoscale heavy precipitation, mesoscale moderately heavy precipitation, and stable precipitation echoes. The classification standard of flood-inducing heavy rainfall is verified by the rainfall cases in Hohhot from June to August 2023. In the application analysis on the July 2, 2023 event, the area growth rate and automatic rain gauge data suggest that heavy precipitation primarily occurs in the period of obviously growing area of strong echoes. Comprehensive analysis of various radar characteristic data indicates that the mean echo intensity and the proportion of strong echo area can reflect the nature of heavy precipitation earlier, which can serving as indicators for early warning. The results show that the"Classification standard of flood-inducing heavy rainfall based on machine learning" has demonstrated operational application value in forecasting and early warning through the verification in 2023. The establishment of this classification standard provides a scientific basis for accurately identifying different types of flood-inducing rainfall. |
| Key words: radar feature; warning indicator; mountain torrents |
