基于深度学习的积层混合云对流泡降水粒子特征研究

doi:10.11755/j.issn.1006-7639(2023)-06-0933

干旱气象 ›› 2023, Vol. 41 ›› Issue (6): 933-943.DOI: 10.11755/j.issn.1006-7639(2023)-06-0933

基于深度学习的积层混合云对流泡降水粒子特征研究

袁雅涵¹^,²(), 王烁¹^,³(), 王文青¹^,³, 张佃国¹^,³, 胡向峰⁴, 张荣⁵, 魏海文⁶, 孟金¹^,², 冯勇¹^,³

1.山东省气象局气象防灾减灾重点实验室，山东济南 250031
2.山东省气象数据中心，山东济南 250031
3.山东省人民政府人工影响天气办公室，山东济南 250031
4.河北省人工影响天气中心，河北石家庄 050021
5.中国气象局人工影响天气中心，北京 100081
6.山东省气象台，山东济南 250031

收稿日期:2022-11-14 修回日期:2023-10-10 出版日期:2023-12-31 发布日期:2024-01-03
通讯作者: 王烁（1992—），男，山东邹城人，工程师，主要从事云降水物理研究。E-mail:wskfg123@163.com。
作者简介:袁雅涵（1996—），女，山东东明人，工程师，主要从事机器学习算法研究。E-mail:foreverlamaara@163.com。
基金资助:
中国气象局创新发展专项(CXFZ2022J034);山东省气象局科研项目(2022sdqxz13);山东省气象局科研项目(2022SDQN03);中部区域积层混合云人工增雨（雪）研究试验（商丘）项目(ZQC-H22256)

A study on particle characteristics of generating cells in stratiform-cumulus mixed cloud based on convolution neural network

YUAN Yahan¹^,²(), WANG Shuo¹^,³(), WANG Wenqing¹^,³, ZHANG Dianguo¹^,³, HU Xiangfeng⁴, ZHANG Rong⁵, WEI Haiwen⁶, MENG Jin¹^,², FENG Yong¹^,³

1. Shandong Key Laboratory for Meteorological Disaster Prevention and Reduction， Ji'nan 250031， China
2. Shandong Meteorological Data Center， Ji'nan 250031， China
3. Shandong Weather Modification Office， Ji'nan 250031，China
4. Hebei Weather Modification Center， Shijiazhuang 050021， China
5. CMA Weather Modification Center， Beijing 100081， China
6. Shandong Meteorological Observatory， Ji'nan 250031， China

Received:2022-11-14 Revised:2023-10-10 Online:2023-12-31 Published:2024-01-03

摘要/Abstract

摘要：

为实现对降水粒子的高精准分类，整理3 a机载探测降水粒子图像，构建山东省降水粒子图像数据集（Shandong Province Precipitation Particle Image Dataset， SD-PPID）。结合多维度混和的模型放缩方法，提出一种基于EfficientNet卷积神经网络的降水粒子识别模型（A Precipitation particle Recognition model based on EfficientNet convolutional neural Network，PREN）。通过多模型、多指标评价对比，验证了PREN模型具有较好的性能和分类识别能力，模型的识别准确率、精准率和召回率均为98%。使用PREN模型分析对流泡降水粒子特征，选取2次典型积层混合云降水过程的3个时段，结合机载Ka波段云雷达（Airborne Ka-Band Precipitation Cloud Radar，KPR）和DMT粒子测量系统（Droplet Measurement Technologies）分析对流泡内部与外部、不同强度和不同高度的降水粒子形状占比，并研究其降水机制。结果表明，PREN可有效识别对流泡降水粒子的特征。对流泡内，主要是球状、针状、不规则状和柱状降水粒子，而对流泡外降水粒子主要为球状和针状。不同强度的对流泡云微物理参数各不相同。降水成熟阶段对流泡内霰粒子和针状粒子占比高于消散阶段，降水成熟阶段降水粒子平均弦长415 μm，而消散阶段粒子平均弦长367 μm。对流泡上部降水粒子以球状和六边形板状为主，主要通过凝华过程增长。在0 ℃层，不规则状粒子和柱状粒子的比例增加，融化过程与动力条件有利于碰并增长形成不规则状粒子，柱状粒子主要来自于高层掉落。

关键词: 图像数据集, 识别模型, 对流泡, 降水粒子特征

Abstract:

Arranging 3 years’ worth of airborne precipitation particle images to construct a precipitation particle image dataset in Shandong Province. Building a precipitation particle recognition model based on EfficientNet convolutional neural network, named PREN(Precipitation particle Recognition model based on EfficientNet convolutional neural Network).The accuracy rate is 98%, and the multi-model and multi-index evaluation and comparison experiments verify that PREN demonstrates excellent robustness and generalization ability. Taking typical stratiform-cumulus mixed cloud precipitation as two examples (total 3 time periods), PREN is applied to the particle characteristics analysis of generating cells. Combined with airborne Ka-band cloud radar and DMT particle measurement system, an analysis conducted on the shape proportion of precipitation particles inside and outside the generating cells and indifferent intensity generating cells, revealing the precipitation mechanism. The results show that the shapes of precipitation particles in the generating cells are mainly spherical, needle-like, irregular and columnar. Precipitation particles outside the generating cells are mostly spherical and needle-like. The cloud microphysical parameters in the generating cells with different intensities vary. The proportion of graupel and needle particles in the precipitation maturity stage is higher than that in the dissipation stage. The average chord length of precipitation particles in the maturity stage is 415 μm. While the average chord length of particles in dissipation stage is 367 μm. The particles on the top of generating cells are mainly spherical and hexagonal, primarily growing through the process of deposition. The ratio of irregular particles and columnar particles in the 0 ℃ are increasing, and the melting process and dynamic conditions favor aggregation and growth, forming irregular particles, while columns mainly originate from the upper levels of the atmosphere.

Key words: SD-PPID, PREN, generating cells, precipitation particle characteristics

中图分类号:

P426.5+1

袁雅涵, 王烁, 王文青, 张佃国, 胡向峰, 张荣, 魏海文, 孟金, 冯勇. 基于深度学习的积层混合云对流泡降水粒子特征研究[J]. 干旱气象, 2023, 41(6): 933-943.

YUAN Yahan, WANG Shuo, WANG Wenqing, ZHANG Dianguo, HU Xiangfeng, ZHANG Rong, WEI Haiwen, MENG Jin, FENG Yong. A study on particle characteristics of generating cells in stratiform-cumulus mixed cloud based on convolution neural network[J]. Journal of Arid Meteorology, 2023, 41(6): 933-943.

图/表 13

图1 粒子及隔断栏典型图像

Fig.1 Typical image of particles and bars

图2 PREN模型（a）和 MBConv结构（b）

Fig.2 PREN model (a) and structural diagram of MBConv (b)

表1 模型混淆矩阵

Tab.1 Confusion matrix

样本真实类别	样本识别类别
样本真实类别	正例	反例
正例	TP（真正例）	FN（假反例）
反例	FP（假正例）	TN（真反例）

图3 PREN模型准确率

Fig.3 Accuracy of PREN model

图4 各类粒子的准确率(a)和损失值(b)的模型训练实验结果

Fig.4 Results of model training experiments for accuracy (a) and loss values (b) for each type of particle

图5 各模型准确率对比

Fig.5 Comparison of the accuracies of each model

图6 不同模型的识别混淆矩阵

Fig.6 Confusion matrix of different models

表2 不同参数对不同模型性能综合评估表

Tab.2 The comprehensive evaluation table of the performance of different models with different parameters

模型	MiPre/%	MaPre/%	MaRa/%	F1-score/%
AlexNet	92.4	93.3	92.4	92.8
ResNet50	97.5	97.7	97.5	97.5
ResNet101	95.7	96.2	95.7	95.9
GoogleNet	94.0	94.7	94.0	94.3
MoblieNet	92.8	93.0	92.8	92.8
PREN	98.0	98.0	98.0	98.0

图7 2018年4月22日10：17—10：21降水消散阶段云雷达回波图（单位：dBZ）（红色双实线为飞行轨迹。下同）

Fig.7 Cloud radar echo of precipitation dissipation stage from 10：17 to 10：21 on 22 April， 2018 （Unit： dBZ）（The red double solid line shows the flight path. Same as below）

表3 不同区域云宏观和微观物理参数信息

Tab.3 Cloud macro and micro physical parameter information in different regions

区域	云宏观和微观物理参数
区域	高度/m	温度/℃	垂直风速/ （m·s^-1）	液态水含量/（g·m^-3）	云滴浓度/cm^-3	云滴平均直径/μm	粒子平均弦长/μm	粒子最大弦长/μm
a	4 330	-0.3	0.2	0.002	11.2	6.1	209	756
b	4 333	0.1	0.19	0.007	16.1	7.4	367	1281
c	4 331	0.2	0.16	0.004	8.9	7.9	278	1223

图8 2018年4月22日10：50—11：00降水成熟阶段云雷达回波（单位：dBZ）

Fig.8 Cloud radar echo of precipitation maturity stage from 10：50 to 11：00 on 22 April， 2018 （Unit： dBZ）

图9 2019年4月27日14：15—14：45降水发展阶段云雷达回波（单位：dBZ）

Fig.9 Cloud radar echo of precipitation developmental stage from 14：15 to 14：45 on 27 April， 2019 （Unit： dBZ）

表4 不同高度云宏观和微观物理参数信息

Tab.4 Cloud macro and micro physical parameter informations in different heights

区域	云宏观和微观物理参数
区域	高度/m	均一化高度	温度/℃	垂直风速/（m·s^-1）	液态水含量/（g·m^-3）	云滴浓度/cm^-3	云滴平均直径/μm
d	5 850	0.79	-9	0.97	0.01	7.3	19.02
e	4 000	0.55	0	升速：1.1 降速：0.82	0.03	125.5	11.30

参考文献 30

[1]	陈梓桐, 李昀英, 2022. 基于云雷达和降水雷达资料的一次典型暖云降水成因及粒子类型演变[J]. 大气科学学报, 45(1): 87-98.
[2]	黄兴友, 陆琳, 洪滔, 等, 2020. 利用毫米波云雷达数据反演层云微物理参数和云内湍流耗散率[J]. 大气科学学报, 43(5): 908-916.
[3]	李宝东, 孙玉稳, 孙霞, 等, 2014. 河北春季一次飞机人工增雪的综合分析[J]. 干旱气象, 32(5): 819-829. DOI
[4]	庞朝云, 张丰伟, 张建辉, 2013. 西北干旱地区一次降水性层状云的飞机观测分析[J]. 干旱气象, 31(2): 272-277.
[5]	唐林, 李琼, 黎祖贤, 等, 2020. 一次积层混合云云系微物理结构数值模拟与增雨条件分析[J]. 干旱气象, 38(1): 100-108.
[6]	王昂生, FUKUTA N, 1984. 冰晶增长规律的定量研究[J]. 大气科学, 8(3): 242-251.
[7]	王磊, 李成才, 赵增亮, 等, 2014. 二维粒子形状分类技术在云微物理特征分析中的应用[J]. 大气科学, 38(2): 201-212.
[8]	袁凯, 庞晶, 李武阶, 等, 2023. 深度学习模型在2021年汛期武汉市雷达回波临近预报中的应用评估[J]. 干旱气象, 41(1): 173-185. DOI
[9]	张佃国, 王烁, 郭学良, 等, 2020. 基于机载Ka波段云雷达和粒子测量系统同步观测的积层混合云对流泡特征[J]. 大气科学, 44(5): 1 023-1 038.
[10]	张荣, 李宏宇, 周旭, 等, 2021. DMT机载云粒子图像形状识别及其应用[J]. 应用气象学报, 32(6): 735-747.
[11]	周志华, 2016. 机器学习[M]. 北京: 清华大学出版社.
[12]	朱士超, 郭学良, 2014. 华北积层混合云中冰晶形状、分布与增长过程的飞机探测研究[J]. 气象学报, 72(2): 366-389.
[13]	朱文刚, 李昌义, 曲美慧, 等, 2020. 深度神经网络方法在山东降水相态判别中的应用[J]. 干旱气象, 38(4): 655-664.
[14]	EVANS A G, LOCATELLI J D, STOELINGA M T, et al, 2005. The IMPROVE-1 storm of 1-2 February 2001. part II: cloud structures and the growth of precipitation[J]. Journal of the Atmospheric Sciences, 62(10): 3 456-3 473.
[15]	HE K M, ZHANG X Y, REN S Q, et al, 2016. Deep residual learning for image recognition[C]// 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). June 27-30, 2016, Las Vegas, NV, USA. IEEE, 770-778.
[16]	HERZEGH P H, HOBBS P V, 1980. The mesoscale and microscale structure and organization of clouds and precipitation in midlatitude cyclones. II: warm-frontal clouds[J]. Journal of the Atmospheric Sciences, 37(3): 597-611.
[17]	HEYMSFIELD A J, PARRISH J L, 1979. Techniques employed in the processing of particle size spectra and state parameter data obtained with the T-28 aircraft platform[M]. Convective Storms Division, National Center for Atmospheric Research, 34-37.
[18]	HEYMSFIELD A, WILLIS P, 2014. Cloud conditions favoring secondary ice particle production in tropical maritime convection[J]. Journal of the Atmospheric Sciences, 71(12): 4 500-4 526.
[19]	HINTON G E, SALAKHUTDINOV R R, 2006. Reducing the dimensionality of data with neural networks[J]. Science, 313(5 786): 504-507.
[20]	HOBBS P V, LOCATELLI J D, 1978. Rainbands, precipitation cores and generating cells in a cyclonic storm[J]. Journal of the Atmospheric Sciences, 35(2): 230-241.
[21]	HOLROYD E W III, 1987. Some techniques and uses of 2D-C habit classification software for snow particles[J]. Journal of Atmospheric and Oceanic Technology, 4(3): 498-511.
[22]	IKEDA K, RASMUSSEN R M, HALL W D, et al, 2007. Observations of freezing drizzle in extratropical cyclonic storms during IMPROVE-2[J]. Journal of the Atmospheric Sciences, 64(9): 3 016-3 043.
[23]	ILLINGWORTH A J, HOGAN R J, O'CONNOR E J, et al, 2007. Cloudnet: continuous evaluation of cloud profiles in seven operational models using ground-based observations[J]. Bulletin of the American Meteorological Society, 88(6): 883-898.
[24]	KRIZHEVSKY A, SUTSKEVER I, HINTON G E, 2012. ImageNet classification with deep convolutional neural networks[C]// Proceedings of the 25th International Conference on Neural Information Processing Systems-Volume 1. December 3-6,2012, Lake Tahoe, Nevada. New York: ACM, 1 097-1 105.
[25]	PLANK V G, ATLAS D, PAULSEN W H, 1955. The nature and detectability of clouds and precipitation as determined by 1.25-centimeter radar[J]. Journal of Meteorology, 12(4): 358-378.
[26]	RAHMAN M M, QUINCY E A, JACQUOT R G, et al, 1981. Feature extraction and selection for pattern recognition of two-dimensional hydrometeor images[J]. Journal of Applied Meteorology, 20(5): 521-535.
[27]	ROSENOW A A, PLUMMER D M, RAUBER R M, et al, 2014. Vertical velocity and physical structure of generating cells and convection in the comma head region of continental winter cyclones[J]. Journal of the Atmospheric Sciences, 71(5): 1 538-1 558.
[28]	SIMONYAN K, ZISSERMAN A, 2014. Very deep convolutional networks for large-scale image recognition[J]. CoRR, abs/1 409.1 556.
[29]	TAN M X, LE Q V, 2019. EfficientNet: rethinking model scaling for convolutional neural networks[EB/OL]. arXiv: 1905.11946. https://arxiv.org/abs/1905.11946.pdf.
[30]	YANG J, WANG Z E, HEYMSFIELD A J, et al, 2016. Characteristics of vertical air motion in isolated convective clouds[J]. Atmospheric Chemistry and Physics, 16(15): 10 159-10 173.

基于深度学习的积层混合云对流泡降水粒子特征研究

A study on particle characteristics of generating cells in stratiform-cumulus mixed cloud based on convolution neural network

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 13

参考文献 30

相关文章 1

编辑推荐

Metrics