基于DOCnet的强对流天气分类识别  

Classification of Severe Convective Weather Based on DOCnet

作　　者：陈小兰 杨昊 陈敏 邹茂扬[3] 周航 CHEN Xiaolan;YANG Hao;CHEN Min;ZOU Maoyang;ZHOU Hang(School of Computing Science,Chengdu University of Information Technology,Chengdu 610225,China;Chengdu Institute of Computer Applications,Chinese Academy of Sciences;School of Blockchain Industry,Chengdu University of Information Technology,Chengdu 610225,China)

机构地区：[1]成都信息工程大学计算机学院,四川成都610225 [2]中国科学院成都计算机应用研究所 [3]成都信息工程大学区块链产业学院,四川成都610225

出　　处：《软件导刊》2024年第11期39-46,共8页Software Guide

基　　金：国家重点研发计划项目(2021YFC3000902);四川省科技计划项目(2022YFS0542);四川省自然科学基金项目(2022NSF⁃SC0964)。

摘　　要：强对流天气对人民生命财产有严重威胁,精确识别强对流天气有益于更好地预测强对流天气。为了解决传统机器学习方法忽视不同强对流天气在雷达图上的不同形状表达和传统机器学习方法计算量大等问题,提出一种基于DOCnet的多类强对流天气分类识别方法。该网络使用深度倍频卷积分别提取雷达图像的高频特征和低频特征,在去除低频特征图中冗余空间减少参数量的同时,增大了提取低频特征图和高频特征图卷积层的感受野,能充分提取雷达图的图像特征,提高模型对强对流天气的分类准确度。通过泛洪法训练模型,提高了模型泛化能力。在风暴事件图像(SEVIR)数据的测试集中,DOCnet模型对强降水、雷暴大风、冰雹、龙卷风4类强对流天气分类的平均命中率为90.54%,平均临界成功指数为81.2%,平均空报比率为11.9%。与基线模型相比,DOCnet的命中率提高了15.02个百分点;与表现最好的MobileNetV2相比,DOCnet的命中率高出5.87%。实验结果表明,DOCnet能够有效提高强对流天气分类效果。Severe convective weather will pose a serious threat to people's lives and property,accurate identification of it is beneficial to bet⁃ter forecasting it.In order to solve the problems that traditional machine learning methods ignore the different shape expressions of different se⁃vere convective weather on radar and the traditional machine learning methods have a large amount of calculation,a classification and identifi⁃cation method of multi-category severe convective weather based on DOCnet is proposed.This network uses deepwise octave convolution to ex⁃tract high-frequency features and low-frequency features of radar respectively.This not only removes the redundant space in low-frequency feature maps and reduces the amount of parameters,but also increases the receptive field of the convolution layer for extracting low-frequency feature maps and high-frequency feature maps.which enables the network to fully extract the image features of the radar and improve the mod⁃el's classification accuracy for severe convective weather.Finally,the model is trained through the flooding method to improve the generaliza⁃tion ability of the model.In the test set of the storm event image(SEVIR)data,the DOCnet model has an average probability of detection of 90.54%,an average critical success index of 81.2%and the average false alarm ration of 11.9%for the four types of severe convective weather:heavy precipitation,thunderstorms,hail,and tornadoes.Compared with the baseline model,DOCnet achieves a probability of detection improvement of 15.02 percentage points,and compared with the best performing MobileNet V2,DOCnet achieves a hit rate 5.87%higher than that of MobileNet V2.The experimental results show that DOCnet can effectively improve the classification effect of strong convective weather.

关 键 词：分类识别 强对流天气 倍频卷积 深度卷积 雷达图 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]