###
气象:2023,49(4):454-466
本文二维码信息
码上扫一扫!
基于多任务学习的地基云图识别与分割技术
张雪,贾克斌,刘钧,张亮
(北京工业大学信息与通信工程学院,北京 100124; 先进信息网络北京实验室,北京 100124;华云升达气象科技有限公司,北京 102299)
Ground Cloud Image Recognition and Segmentation Technology Based on Multi-Task Learning
ZHANG Xue,JIA Kebin,LIU Jun,ZHANG Liang
(School of Information and Communication Engineering, Beijing University of Technology, Beijing 100124; Beijing Laboratory of Advanced Information Network, Beijing 100124;Huayun Shengda Meteorological Technology Co., Ltd., Beijing 102299)
摘要
图/表
参考文献
相似文献
本文已被:浏览 169次   下载 790
投稿时间:2021-12-16    修订日期:2022-09-28
中文摘要: 云在天气预报中扮演着一个至关重要的角色,准确识别和分割地基云图可以有效指导天气预报。针对大部分现有数据集只适用于单任务学习,地基云图识别和分割技术多以单任务实现,识别检测效率低且算法鲁棒性差等问题,构建了带标签且适合多任务学习的地基云图数据集(GBCD)和GT数据集(GBCD-GT),在此基础上设计了一种基于多任务学习的地基云图识别与分割联合网络模型(GCRSegNet)。该模型首先通过卷积神经网络提取共享特征,再为每个任务设计特定网络,提取更具辨识度的特征,分割网络通过学习共享特征实现地基云图分割,识别网络通过结合共享特征和分割特征实现地基云图识别。经过多组对比试验表明,该网络能准确表征地基云图特征,使识别任务准确率达到94.28%,分割任务像素准确率达到 93.85%,平均交并比达到71.58%,为实际应用提供了可能性。
Abstract:Clouds play an important role in weather forecasting. Accurate identification and segmentation of ground-based cloud images can effectively guide weather forecasting. Now most of the existing datasets are only suitable for single task learning, and ground-based cloud image recognition and segmentation technologies are mostly implemented by single task, thus identification and detection efficiency are low and the robustness of the algorithm is poor. Considering these problems, we construct the ground-based cloud image datasets GBCD and GBCD-GT with labels and suitable for multi-task learning, and then on this basis, a ground-based cloud image recognition and segmentation joint network model GCRSegNet based on multi-task learning is designed. The model firstly extracts shared features through convolutional neural network, then a special network is designed for each task to extract more recognizable features. The segmentation network learns shared features to achieve ground-based cloud image segmentation, and the recognition network combines sharing features and segmentation features to achieve ground-based cloud image recognition. Multiple groups of comparative experiments indicate that the network in this paper can accurately represent the features of ground-based cloud image. Meanwhile, the accuracy of the recognition task can reach 94.28%, the pixel accuracy of segmentation task can reach 93.85%, and mean intersection over union reach 71.58%. These results can provide a possibility for practical application.
文章编号:     中图分类号:    文献标志码:
基金项目:国家重点研发计划(2018YFF01010100)、北京市自然科学基金项目(4212001)共同资助
引用文本:
张雪,贾克斌,刘钧,张亮,2023.基于多任务学习的地基云图识别与分割技术[J].气象,49(4):454-466.
ZHANG Xue,JIA Kebin,LIU Jun,ZHANG Liang,2023.Ground Cloud Image Recognition and Segmentation Technology Based on Multi-Task Learning[J].Meteor Mon,49(4):454-466.