干旱气象 ›› 2025, Vol. 43 ›› Issue (5): 810-819.DOI: 10.11755/j.issn.1006-7639-2025-05-0810

• 研究论文 • 上一篇    下一篇

基于人工和地基器测数据的AI云识别方法

张德玉1(), 胡树贞2(), 秦三杰3, 张强1, 白明1, 庞成1, 魏荣妮1   

  1. 1.甘肃省张掖市气象局,甘肃 张掖 734000
    2.中国气象局气象探测中心,北京 100081
    3.甘肃省气象局观测与网络处,甘肃 兰州 730020
  • 收稿日期:2024-11-28 修回日期:2025-05-02 出版日期:2025-10-31 发布日期:2025-11-09
  • 通讯作者: 胡树贞(1985—),男,山东德州人,高级工程师,主要从事毫米波雷达资料分析及应用研究。E-mail: 052310421hu@163.com
  • 作者简介:张德玉(1978—),男,甘肃民乐人,高级工程师,主要从事气象雷达与大气探测研究。E-mail: zydeyu@163.com
  • 基金资助:
    中国气象局气象探测中心观测试验项目(GCSYJH23-23);甘肃省气象局科研项目(Zd2024-A-1-A)

AI cloud classification based on manual and ground-based instrumental data

ZHANG Deyu1(), HU Shuzhen2(), QIN Sanjie3, ZHANG Qiang1, BAI Ming1, PANG Cheng1, WEI Rongni1   

  1. 1. Zhangye Meteorological Bureau of Gansu Province, Zhangye 734000, Gansu, China
    2. CMA Meteorological Observation Center, Beijing 100081, China
    3. Department of Observation and Network,Gansu Meteorological Bureau, Lanzhou 730020, China
  • Received:2024-11-28 Revised:2025-05-02 Online:2025-10-31 Published:2025-11-09

摘要:

为弥补“天气现象视频智能观测仪”在云状识别中存在的纯视觉观测局限,以2023年5月1日至2024年4月30日张掖国家气候观象台试验外场的毫米波云雷达、全天空成像仪等器测数据为基础,结合人工观测云状记录及地面自动站气象资料,构建多源融合的人工智能(Artificial Intelligence,AI)云状识别样本库。选取多种类型机器学习算法开展训练与性能评估,结果表明,支持向量机模型在综合识别精度与稳定性方面表现最佳,可实现对卷积云、卷云、高积云、高层云、雨层云、层云、层积云、积雨云、积云9种云状及降水天气的自动识别。通过4个典型日云分类个例的验证显示,模型能精准识别多层云结构,识别结果与人工观测高度一致。本文在数据集构建的多源融合性及算法适配性方面均有明显改进,云状识别种类增加33%,准确率提升15%。

关键词: 地基器测云, AI云识别, 毫米波云雷达, 全天空成像仪, 人工云状

Abstract:

To compensate for the limitations of the Weather Phenomena Video Intelligent Observation Instrument in purely visual cloud classification, an AI (Artificial Intelligence) cloud classification sample library was constructed by integrating millimeter-wave cloud radar and all-sky imager measurements from the Zhangye National Climate Observatory experimental field during the period from May 1, 2023 to April 30, 2024, together with manual cloud observations and ground automatic station meteorological data. Multiple machine learning algorithms were applied for training and performance evaluation, and the Support Vector Machine model was identified as the most effective in terms of overall accuracy and stability. This model enables the automatic recognition of nine cloud types, cirrocumulus, cirrus, altocumulus, altostratus, nimbostratus, stratus, stratocumulus, cumulonimbus, and cumulus, as well as precipitation weather. Verification based on four typical daily cloud classification cases demonstrated that the model can accurately identify multilayer cloud structures, with results highly consistent with manual observations. This study achieves significant improvements in both data integration and algorithm adaptability, increasing the number of identifiable cloud types by 33% and enhancing classification accuracy by 15%.

Key words: ground-based cloud observation, AI cloud classification, millimeter-wave cloud radar, all-sky imager, artificial cloud shape

中图分类号: