基于深度学习的地基云图分割研究  

Research on Ground-Based Cloud Image Segmentation Based on Deep Learning

作　　者：官禹 李守智 Guan Yu;Li Shouzhi(Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing,China;Key Laboratory of Technology in Geo-Spatial information Processing and Application System,Chinese Academy of Sciences,Beijing,China;School of Electronic,Electrical and Communication Engineering,University of Chinese Academy of Sciences,Beijing,China)

机构地区：[1]中国科学院空天信息创新研究院,北京 [2]中国科学院空间信息处理与应用系统技术重点实验室,北京 [3]中国科学院大学电子电气与工程学院,北京

出　　处：《科学技术创新》2024年第16期102-105,共4页Scientific and Technological Innovation

摘　　要：地基云图分割本身易受天、光线和太阳直射角等因素影响,已有深度学习分割方法在不进行领域适配的前提下,往往对云层边界分割效果不佳。基于以上因素,本文选取了对边界识别能力较强的CloudSegNet、DeepLabV3以及U-Net模型。另外,为了选择出最优的特征抽取网络,本文通过调研选择了VGG19、ResNet101、SE_Resnext101以及mobilenet_v2作为候选的特征抽取网络。最后,为了进一步提升模型对云层边界的分割能力,本文在已有的深度分割模型基础上,引入多任务学习,实现对云层边界单独建模,提高模型的云层边界识别能力。ground-based cloud segmentation is susceptible to various environmental factors such as weather conditions,lighting,and solar zenith angle.Existing deep learning segmentation methods often yield unsatisfactory results in cloud boundary segmentation without domain adaptation.Considering these factors,this study selects the CloudSegNet,DeepLabV3,and U-Net models known for their strong boundary recognition capabilities.Furthermore,to identify the optimal feature extraction networks,VGG19,ResNet101,SE_Resnext101,and Mobilenet_v2 are chosen through research.Lastly,to further enhance the segmentation capability of cloud boundaries,multi-task learning is introduced on top of existing deep segmentation models to independently model cloud layer boundaries and improve the recognition ability of cloud layer boundaries.

关 键 词：地基云图 深度学习 特征抽取网络 多任务学习网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]