内蒙古冰雹特征及基于机器学习的冰雹识别方法研究

doi:10.12118/j.issn.1000-6060.2024.057

摘要/Abstract

摘要： 利用1959—2021年内蒙古人工观测冰雹记录，分析冰雹分布的时空特征，并基于机器学习算法构建了冰雹识别方法。结果表明：（1）时间分布上，冰雹事件出现的站数和站日数均呈现下降趋势；空间分布上，冰雹多集中在阴山山脉和大兴安岭一带，冰雹多发区沿山脉伸展分布。（2）冰雹发生具有明显的季节变化和日变化特征，每年5—9月是冰雹频发月份，占全年雹日的91.79%，雹日中12:00—19:00是冰雹的多发时段。（3）利用随机森林、LightGBM、K近邻和决策树4种机器学习算法，通过数据预处理、预报因子选择、模型训练、模型调优等步骤，对内蒙古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，各模型的TS评分均达到0.83以上，命中率达到92%以上，随机森林算法在测试集上识别效果最优。研究结果可为内蒙古冰雹预报预警和人工防雹工作提供参考。

关键词: 冰雹站日数, 时空特征, 机器学习, 冰雹识别

Abstract:

Based on the manual observation of hail records in Inner Mongolia, China, from 1959 to 2021, the spatial and temporal characteristics of hail distribution are analyzed, and a hail recognition method is constructed based on machine learning algorithms. The results are as follows: (1) Regarding temporal distribution, the number of hail days and affected stations in Inner Mongolia shows a decreasing trend. In terms of spatial distribution, hail events are predominantly concentrated in the Yinshan Mountains and the Greater Hinggan Mountains, with hail-prone areas extending along these mountain ranges. (2) Hail exhibits distinct seasonal and diurnal characteristics. The peak hail months in Inner Mongolia are from May to September, accounting for 91.79% of the annual hail days. The most frequent period for hail occurrences is between 12:00 BST and 19:00 BST. (3) Four machine learning algorithms (random forest, LightGBM, K-proximity, and decision tree) are used to model and evaluate hail events in Inner Mongolia through data preprocessing, predictor selection, model training, and tuning. Verification results indicate that machine learning methods effectively identify hail events, with the threat score of each model exceeding 0.83 and hit rates surpassing 92%. Among these, the random forest algorithm demonstrates the best recognition performance on the test set. These findings provide useful references for hail forecasting and artificial hail prevention in Inner Mongolia.

Key words: the number of hail station days, temporal and spatial characteristics, machine learning, hail identification

辛悦, 苏立娟, 郑旭程, 李慧, 衣娜娜, 靳雨晨. 内蒙古冰雹特征及基于机器学习的冰雹识别方法研究[J]. 干旱区地理, 2025, 48(1): 11-19.

XIN Yue, SU Lijuan, ZHENG Xucheng, LI Hui, YI Nana, JIN Yuchen. Hail characteristics and hail recognition method based on machine learning in Inner Mongolia[J]. Arid Land Geography, 2025, 48(1): 11-19.

图/表 9

图1

表1

图2

表2

图3

表3

表4

表5

表6

参考文献 26

[1]	姚展予, 屠琦, 安琳, 等. 冰雹形成过程及人工防雹研究综述[J]. 气象学报, 2022, 80(6): 835-863.
	[Yao Zhanyu, Tu Qi, An Lin, et al. Review of advances in hail formation process and hail suppression and hail suppression research[J]. Acta Meteorologica Sinica, 2022, 80(6): 835-863.]
[2]	陶涛, 张立新, 桑建人, 等. 六盘山区一次非典型冰雹天气过程微物理量特征的分析[J]. 干旱区地理, 2020, 43(2): 299-307.
	[Tao Tao, Zhang Lixin, Sang Jianren, et al. A case analysis of microphysical characteristics of atypical hail formation over Liupan Mountain, China[J]. Arid Land Geography, 2020, 43(2): 299-307.]
[3]	王昀, 谢向阳, 马禹, 等. 天山北侧成灾雹云移动路径及预警指标的研究[J]. 干旱区地理, 2017, 40(6): 1152-1164.
	[Wang Yun, Xie Xiangyang, Ma Yu, et al. Moving paths and nowcasting indicators of radar of hail cloud in northern Tianshan Mountains[J]. Arid Land Geography, 2017, 40(6): 1152-1164.]
[4]	张芳华, 高辉. 中国冰雹日数的时空分布特征[J]. 南京气象学院学报, 2008, 31(5): 687-693.
	[Zhang Fanghua, Gao Hui. Temporal and spatial features of hail days in China[J]. Journal of Nanjing Institute of Meteorology, 2008, 31(5): 687-693.]
[5]	汤兴芝, 黄治勇, 张荣, 等. 2010—2020年全国冰雹灾害事件时空分布特征[J]. 暴雨灾害, 2023, 42(2): 223-231.
	[Tang Xingzhi, Huang Zhiyong, Zhang Rong, et al. Temporal and spatial distribution characteristics of hail disaster events in China from 2012 to 2020[J]. Torrential Rain and Disasters, 2023, 42(2): 223-231.]
[6]	尉英华, 花家嘉, 王莹, 等. 近11年天津冰雹统计特征及对流参数指标分析[J]. 气象, 2023, 49(2): 213-223.
	[Wei Yinghua, Hua Jiajia, Wang Ying, et al. Statistical characteristics and convection indicators of hailstorm over Tianjin in recent 11 years[J]. Meteorological Monthly, 2023, 49(2): 213-223.]
[7]	钟敏, 郭英莲, 陈璇, 等. 基于客观分型的冰雹概率预报方法研究[J]. 高原气象, 2022, 41(4): 934-944. doi: 10.7522/j.issn.1000-0534.2021.00030
	[Zhong Min, Guo Yinglian, Chen Xuan, et al. Study on hail probability forecast method based on objective classification[J]. Plateau Meteorology, 2022, 41(4): 934-944.] doi: 10.7522/j.issn.1000-0534.2021.00030
[8]	刘新伟, 黄武斌, 蒋盈沙, 等. 基于LightGBM算法的强对流天气分类识别研究[J]. 高原气象, 2021, 40(4): 909-918. doi: 10.7522/j.issn.1000-0534.2020.00075
	[Liu Xinwei, Huang Wubin, Jiang Yingsha, et al. Study of the classified identification of the strong convective weathers based on the LightGBM algorithm[J]. Plateau Meteorology, 2021, 40(4): 909-918.] doi: 10.7522/j.issn.1000-0534.2020.00075
[9]	刘新伟, 蒋盈沙, 黄武斌, 等. 基于雷达产品和随机森林算法的冰雹天气分类识别及预报[J]. 高原气象, 2021, 40(4): 898-908. doi: 10.7522/j.issn.1000-0534.2020.00063
	[Liu Xinwei, Jiang Yingsha, Huang Wubin, et al. Classified identification and nowcast of hail weather based on radar products and random forest algorithm[J]. Plateau Meteorology, 2021, 40(4): 898-908.] doi: 10.7522/j.issn.1000-0534.2020.00063
[10]	汤兴芝, 黄兴友. 冰雹云的多普勒天气雷达识别参量及其预警作用[J]. 暴雨灾害, 2009, 28(3): 261-265.
	[Tang Xingzhi, Huang Xingyou. Doppler radar identification parameters and their effect on early warning of hail clouds[J]. Torrential Rain and Disasters, 2009, 28(3): 261-265.]
[11]	韩经纬, 王海梅, 乌兰, 等. 内蒙古雷暴、冰雹灾害的评估分析与防御对策研究[J]. 干旱区资源与环境, 2009, 27(7): 31-38.
	[Han Jingwei, Wang Haimei, Wu Lan, et al. The analysis and assessment on thunderstorm and hail disasters and the countermeasures in Inner Mongolia[J]. Journal of Arid Land Resources and Environment, 2009, 27(7): 31-38.]
[12]	顾润源, 孙永刚, 韩经纬, 等. 内蒙古自治区天气预报手册[M]. 北京: 气象出版社, 2012: 277-281.
	[Gu Ruiyuan, Sun Yonggang, Han Jingwei, et al. Weather forecast manual of Inner Mongolia[M]. Beijing: China Meterological Press, 2012: 277-281.]
[13]	李文娟, 赵放, 郦敏杰, 等. 基于数值预报和随机森林算法的强对流天气分类预报技术[J]. 气象, 2018, 44(12): 1555-1564.
	[Li Wenjuan, Zhao Fang, Li Minjie, et al. Forecasting and classification of severe convective weather based on numerical forecast and random forest algorithm[J]. Meteorological Monthly, 2018, 44(12): 1555-1564.]
[14]	周康辉. 基于深度卷积神经网络的强对流天气预报方法研究[D]. 北京: 中国气象科学研究院, 2021.
	[Zhou Kanghui. Convective weather forecasting with convolutional neural networks[D]. Beijing: University of Chinese Academy of Sciences, 2021.]
[15]	王芝兰, 陈录元, 尚可政, 等. 青海强对流天气时空特征及其对气候变暖的响应[J]. 干旱气象, 2011, 29(4): 439-445, 454.
	[Wang Zhilan, Chen Luyuan, Shang Kezheng, et al. Characteristics of temporal and spatial distribution of severe convective weather and its response to climate warming in Qinghai Province[J]. Arid Meteorology, 2011, 29(4): 439-445, 454.]
[16]	加勇次成, 次旺顿珠, 措姆. 气候变化背景下羌塘自然保护区冰雹日数时空变化特征[J]. 中国农学通报, 2019, 35(18): 103-109. doi: 10.11924/j.issn.1000-6850.casb18120118
	[Jiayongcicheng, Ciwangdunzhu, Cuomu. Hail days in Qiangtang National Nature Reserve of Tibet: Spatial and temporal variation characteristics under the background of climate change[J]. Chinese Agricultural Science Bulletin, 2019, 35(18): 103-109.] doi: 10.11924/j.issn.1000-6850.casb18120118
[17]	邸强, 穆超, 孔文甲. 呼和浩特地区冰雹时空特征分析[J]. 内蒙古科技与经济, 2023, 4: 92-93, 157.
	[Di Qiang, Mu Chao, Kong Wenjia. Analysis of spatiotemporal characteristics of hail in Hohhot area[J]. Inner Mongolia Science Technology and Economy, 2023, 4: 92-93, 157.]
[18]	哈青辰. 近32年赤峰地区冰雹的分布特征及防御[J]. 农业灾害研究, 2021, 11(7): 83-84.
	[Ha Qingchen. Distribution characteristics and defense of hail in Chifeng area in recent 32 years[J]. Journal of Agricultural Catastropholgy, 2021, 11(7): 83-84.]
[19]	罗喜平, 廖波, 张小娟, 等. 1961—2020年贵州冰雹气候特征[J]. 干旱气象, 2022, 40(6): 1024-1032. doi: 10.11755/j.issn.1006-7639(2022)-06-1024
	[Luo Xiping, Liao Bo, Zhang Xiaojuan, et al. Climatic characteristics of hail in Guizhou from 1961 to 2020[J]. Journal of Arid Meteorology, 2022, 40(6): 1024-1032.]
[20]	Friedman J H. Greedy function approximation: A gradient boosting machine[J]. Annals of Statistics, 2001, 29(5): 1189-1232.
[21]	张勇, 刘慧, 郑颖菲, 等. 人工智能模型的分类临近预报产品效果检验与分析[J]. 沙漠与绿洲气象, 2023, 17(1): 115-121.
	[Zhang Yong, Liu Hui, Zheng Yingfei, et al. Effect validation and analysis of classified products outputted by artificial intelligent nowcasting model[J]. Desert and Oasis Meteorology, 2023, 17(1): 115-121.]
[22]	刘瑞亮, 贾科利, 李小雨, 等. 组合光学和微波遥感的耕地土壤含盐量反演[J]. 干旱区地理, 2024, 47(3): 433-444. doi: 10.12118/j.issn.1000-6060.2023.375
	[Liu Ruiliang, Jia Keli, Li Xiaoyu, et al. Inversion of soil salt content by combining optical and microwave remote sensing in cultivated land[J]. Arid Land Geography, 2024, 47(3): 433-444.] doi: 10.12118/j.issn.1000-6060.2023.375
[23]	Jordan M I, Mitchell T M. Machine learning: Trends, perspectives, and prospects[J]. Science, 2015, 349(6245): 255-260. doi: 10.1126/science.aaa8415 pmid: 26185243
[24]	朱思华, 罗继, 曲良璐. 新疆阿克苏地区冰雹时空分布及雷达回波特征[J]. 沙漠与绿洲气象, 2021, 15(2): 81-88.
	[Zhu Sihua, Luo Ji, Qu Lianglu. The spatial-temporal distribution and radar echo signatures of hail in Aksu, Xinjiang[J]. Desert and Oasis Meteorology, 2021, 15(2): 81-88.]
[25]	虎雅琼, 边宇轩, 黄梦宇, 等. 基于灾情信息的1981—2017年北京地区降雹特征[J]. 应用气象学报, 2019, 30(6): 710-721.
	[Hu Yaqiong, Bian Yuxuan, Huang Mengyu, et al. Characteristics of hailstone distribution based on disaster in Beijing from 1981 to 2017[J]. Journal of Meteorological Science, 2019, 30(6): 710-721.]
[26]	唐文苑, 周庆亮, 刘鑫华, 等. 国家级强对流天气分类预报检验分析[J]. 气象, 2017, 43(1): 67-76.
	[Tang Wenyuan, Zhou Qingliang, Liu Xinhua, et al. Anlyisis on verification of national severe convective weather categorical forecasts[J]. Meteorological Monthly, 2017, 43(1): 67-76.]

冰雹特征参数	1月	2月	3月	4月	5月	6月	7月	8月	9月	10月	11月	12月
冰雹站日数	0	0	30	448	1419	2645	2134	1697	1644	364	9	2
冰雹站日数频率/%	0.00	0.00	0.29	4.31	13.65	25.45	20.54	16.33	15.82	3.50	0.09	0.02

预报因子类型	预报物理量
指数类	K指数、修正K指数、总指数、强天气威胁指数、沙氏指数
水汽因子	柱总水量、柱水汽量、850 hPa相对湿度、700 hPa相对湿度、850 hPa露点温度差、700 hPa露点温度差、瞬时最大湿通量
动力因子	850~500 hPa风切变、850~300 hPa风切变、10 m瞬时最大风速
高度层因子	0 ℃层高度、-20 ℃层高度、-20~0 ℃厚度
热力因子	850~500 hPa温度差、850 hPa假相当位温、700 hPa假相当位温、500 hPa假相当位温、850~500 hPa假相当位温差、700~500 hPa假相当位温差、对流有效位能
其他信息	冰雹发生的年份、月份、日期、当地时、经度、纬度

模型名称	命中数/个	漏报数/个	空报数/个	命中率/%	空报率/%	漏报率/%	TS评分
随机森林	1854	137	158	93.11	7.82	6.88	0.8627
LightGBM	1820	171	168	91.55	8.59	8.45	0.8429
K近邻	1777	214	154	89.25	7.97	10.74	0.8284
决策树	1759	232	217	88.34	10.98	11.65	0.7966

超参数	初始值	取值范围
n_neighbors	3	[1, 11]
n_estimators	100	[10, 500]
max_depth	2	[1, 11]
min_sample_leaves	10	[1, 50]

模型名称	命中数/个	漏报数/个	空报数/个	命中率/%	空报率/%	漏报率/%	TS评分
随机森林	1856	135	155	93.22	7.71	6.78	0.8648
LightGBM	1848	143	158	92.82	7.88	7.18	0.8599
K近邻	1855	136	196	93.17	9.56	6.83	0.8481
决策树	1854	137	226	93.12	10.87	6.88	0.8362