CT-CloudDetect:用于遥感卫星云检测的混合模型  

CT-CloudDetect:A Hybrid Model for Remote Sensing Satellite Cloud Detection

作　　者：方巍[1,2,3,4,5] 陶恩屹 FANG Wei;TAO Enyi(School of Computer,Nanjing University of Information Science and Technology,Nanjing 210044,China;Engineering Research Center of Digital Forensics,Ministry of Education,Nanjing University of Information Science and Technology,Nanjing 210044,China;Key Laboratory of Transportation Meteorology of China Meteorological Administration,Nanjing Joint Institute for Atmospheric Sciences,Nanjing 210041,China;Jiangsu Collaborative Innovation Center of Atmospheric Environment and Equipment Technology(CICAEET),Nanjing University of Information Science and Technology,Nanjing 210044,China;Provincial Key Laboratory for Computer Information Processing Technology,Soochow University,Suzhou,Jiangsu 215000,China)

机构地区：[1]南京信息工程大学计算机学院,南京210044 [2]南京信息工程大学数字取证教育部工程研究中心,南京210044 [3]南京气象科技创新研究院中国气象局交通气象重点开放实验室,南京210041 [4]南京信息工程大学江苏省大气环境与装备技术协同创新中心,南京210041 [5]苏州大学江苏省计算机信息处理技术重点实验室,江苏苏州215000

出　　处：《遥感信息》2024年第5期1-11,共11页Remote Sensing Information

基　　金：国家自然科学基金(42075007);苏州大学计算机信息处理技术重点实验室开放项目(KJS2275);南京气象科技创新研究院北极阁开放研究基金(BJG202306);江苏省研究生科研与实践创新计划(SJCX24_0476、SJCX24_0477)。

摘　　要：云检测是在遥感卫星云图中检测云的任务。近年来,人们提出了基于深度学习的云检测方法,并取得了良好的性能。然而,现有的基于深度学习的云检测模型大多还是基于卷积神经网络(convolutional neural network,CNN),由于卷积运算的固有局部性,难以捕获长距离依赖关系。针对上述问题,文章提出一个基于CNN和ViT(Vision Transformer)的混合型云检测模型,并提出一种基于CNN和ViT的编码器,使网络具备捕捉局部和全局信息的能力。为了更好地融合语义和尺度不一致的特征,提出了一个双尺度注意力融合模块,通过注意力机制有选择地融合特征。此外,提出了轻量级路由解码器,该解码器通过路由结构降低模型复杂度。在3个公开云检测数据集上对模型进行了评估。大量实验表明,所提出的模型具有比现有模型更好的性能。Cloud detection is the task of detecting clouds in remote sensing satellite cloud images.In recent years,deep learning-based cloud detection methods have been proposed and achieved good performance.However,most of the existing deep learning-based cloud detection models are still based on convolutional neural network(CNN),which is difficult to capture long-range dependencies due to the inherent localization of convolutional operations.To address the above problems,we propose a hybrid cloud detection model based on CNN and ViT(Vision Transformer),and an encoder based on CNN and ViT,which equips the network with the ability to capture both local and global information.In order to better fuse semantic and scale-inconsistent features,we propose a dual-scale attention fusion module,which selectively fuses features through the attention mechanism.In addition,we propose a lightweight routing decoder that reduces model complexity through routing structure.In this paper,the model is evaluated on three public cloud detection datasets.Extensive experiments show that the model proposed in this paper achieves better performance than existing models.

关 键 词：深度学习 卷积神经网络 空间Vision Transformer 混合模型 云检测 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]