中国农业气象 ›› 2023, Vol. 44 ›› Issue (03): 228-237.doi: 10.3969/j.issn.1000-6362.2023.03.006

• 农业气象信息技术 栏目 • 上一篇    下一篇

多源强降雨灾情可信度智能判别方法

司丽丽,赵亮,魏铁鑫,霍治国,李姣   

  1. 1. 河北省气象与生态环境重点实验室,石家庄 050021;2.中国气象局雄安大气边界层重点开放实验室,雄安新区 071800;3. 河北省气象灾害防御和环境气象中心,石家庄 050021;4. 中国气象科学研究院,北京 100081
  • 收稿日期:2022-01-16 出版日期:2023-03-20 发布日期:2023-03-14
  • 通讯作者: 霍治国,二级研究员,主要从事农业气象灾害预测与评估研究。 E-mail: huozg@cma.gov.cn
  • 作者简介:丽丽,E-mail: sll_0312@163.com
  • 基金资助:
    河北省气象局科研开发项目(19ky06);2022年山洪地质灾害防治气象保障工程建设项目

An Intelligent Method for Discriminating the Reliability of Multi-Source Heavy Rainfall Disaster Information

SI Li-li , ZHAO Liang, WEI Tie-xin, HUO Zhi-guo, LI Jiao   

  1. 1. Key Laboratory of Meteorology and Ecological Environment of Hebei Province, Shijiazhuang 050021, China; 2.China Meteorological Administration Xiong'an Atmospheric Boundary Layer Key Laboratory, Xiong'an New Area 071800;3. Hebei Meteorological Disaster Prevention and Environment Meteorology Center, Shijiazhuang 050021; 4. Chinese Academy of Meteorological Sciences, Beijing 100081
  • Received:2022-01-16 Online:2023-03-20 Published:2023-03-14

摘要: 真实的灾情信息是有效防范和减轻强降雨灾害损失的重要参考。本研究以过程降雨强度(R)为指标,构建1984−2020年河北省县级多源气象灾情与致灾过程相匹配的强降雨灾害事件库。经过人工质控,获得真实灾情信息2305组,伪灾情信息263组。采用相关分析法确定与灾害发生程度(灾度)显著相关的降雨关键特征因子,基于单类支持向量机和十折交叉检验法,随机抽取10次样本,建立强降雨灾情气象因子致灾判别模型,并进行检验优化,以探索智能化、易用性的多源强降雨灾情可信度智能判别方法。结果表明:(1)与灾度显著相关的降雨关键特征因子共计11个,分别为最大降雨量、最小降雨量、过程平均降雨量、日均降雨量、平均小时雨强、1h最大雨量、3h最大雨量、6h最大雨量、12h最大雨量、24h最大雨量及前10日降雨总量,均通过了0.01水平显著性检验。(2)采用11个因子建立10个致灾判别模型(M1−M10),依据真实灾情判别准确率确定最优模型为M9,其证真率为96.4%,证伪率为67.6%,表明该模型对灾情真伪判定较为片面,应进一步优化。(3)通过自相关检验,以最大降雨量、平均小时雨强、1h最大雨量及前10日降雨总量4个因子作为输入因子,重新构建强降雨灾情气象因子致灾判别模型(M11−M20),最优模型为M20,其证真率和证伪率分别达到96.2%和82.9%。综合分析认为,由4个因子构建的气象因子致灾判别模型评估强降雨灾情可信度比11个因子建立的模型更可靠。

关键词: 强降雨灾害, 支持向量机, 相关分析, 致灾因子, 可信度判别

Abstract: The real disaster information is an important reference for preventing and reducing heavy rainfall disaster losses effectively. Taking process rainfall intensity as the index, this paper constructs a heavy rainfall disaster event database matching the multi-source meteorological disasters and disaster-causing processes in Hebei Province during 1984−2020. After manual quality control, 2305 groups of real disaster information and 263 groups of false information are obtained. In this study, correlation analysis was used to determine and select the rainfall eigenfactors that are significantly related to the disaster degree. Based on the One-class support vector machine (OCSVM) and 10 folds cross-validation method, 10 samples were randomly selected to establish the meteorological factor disaster discriminant model and test and optimize, so as to explore the intelligent and easy-to-use intelligent discriminant method of multi-source heavy rainfall disaster credibility. The results showed that: (1) there are eleven rainfall eigenfactors related to disaster degree at 0.01 significant level, which are maximum rainfall, minimum rainfall, average process rainfall, average daily rainfall, average hourly rainfall intensity, hourly maximum rainfall, 3-hour maximum rainfall, 6-hour maximum rainfall, 12-hour maximum rainfall, 24-hour maximum rainfall and rainfall in the first 10 days. (2) Ten models (M1-M10) were established using 11 rainfall eigenfactors. According to the identification accuracy of the real disaster, the optimal model was determined to be M9, with the authenticity rate of 96.4% and the falsification rate of 67.6%, which indicated that the model is one-sided in determining the authenticity of the disaster situation and should be further optimized. (3) Through the autocorrelation test, maximum rainfall, average hourly rainfall intensity, hourly maximum rainfall and rainfall in the first 10 days were taken as input factors. Ten models (M11-M20) were reconstructed, and the optimal model is M20 with 96.2% proof rate and 82.9% false rate. Based on comprehensive analysis, the model established by 4 factors is more reliable than the model established by 11 factors.

Key words: Heavy rainfall disaster, OCSVM, Correlation analysis, Disaster-causing factor, Reliability discriminant