基于机器学习的冰雹天气识别研究

利用2017—2019年贵州省威宁彝族回族苗族自治县的C波段天气雷达数据,提取了多种与冰雹相关的雷达特征参量,并结合地面降雹记录建立了客观的冰雹样本标记方法。通过支持向量机、决策树和朴素贝叶斯三种机器学习方法,对冰雹高发区贵州省威宁的冰雹天气过程进行建模与评估。评估结果表明,采用机器学习方法可以有效地识别冰雹天气过程。支持向量机和决策树方法的命中率(Probability of Detection,POD)均较高,分别为88.9%和90.5%;临界成功指数(Critical Success Index,CSI)分别为73.1%和70.3%;误报率(False Alarm Rate,FAR)分别为19.6%和24.1%。朴素贝叶斯方法的POD和CSI相对偏低,分别为67.8%、62.8%,但FAR最低,为10.6%。

Hail identification based on machine learning methods

Three machine learning methods, namely support vector machine, decision tree and Naive Bayesian, were applied to model and predict hail events occurred in Weining, Guizhou Province from 2017 to 2019. A variety of hail-related features were extracted from the C-band weather radar and an improved objective hail sample labelling procedure was proposed. The verification results showed that the machine learning methods were able to effectively identify the hail events. Both the support vector machine and decision tree methods had a high POD of 88.9% and 90.5% respectively, CSI of 73.1% and 70.3%, and FAR of 19.6% and 24.1% respectively. The POD and CSI of Naive Bayesian method were lower than the other two methods at 67.8% and 62.8% respectively, but its FAR was the lowest at 10.6%.