基于多特征融合和通道注意力的深度学习云层检测方法  

Deep Learning Cloud Detection Method Based on Multi-feature Fusion and Channel Attention

作　　者：杜晓凤 傅莘莘 朱祎 DU Xiaofeng;FU Shenshen;ZHU Yi(School of Computer and Information Engineering,Xiamen University of Technology,Xiamen,Fujian 360000,China)

机构地区：[1]厦门理工学院计算机与信息工程学院,福建厦门360000

出　　处：《遥感信息》2023年第1期121-129,共9页Remote Sensing Information

基　　金：福建省自然科学基金项目(2019J01855、2019J01854);福建省教育厅A类基金项目(JT180437、JAT190681)。

摘　　要：目前,基于传统图像处理的云层检测方法多依赖于来自特定传感器的纹理、颜色以及温度等物理特征,这类方法在遥感图像下垫面与云层具有相似的特性时,通常会出现较多错、漏检的情况。在深度学习方法良好性能的启发下,文章为解决这一问题提出了一种新的云层检测方法FANet(feature fusion attention network)。FANet整体由特征提取、解码以及注意力模块组成。在特征提取模块中,FANet不仅保留了自顶向下连接,也新增了多条自底向上和同层连接来缩短信息单向流动和融合的路径,以缓解特征信息丢失的问题。此外,模型引入注意力模块计算各级融合特征图所占的权重,使多层级特征更精准地融合并形成丰富的语义和空间特征。最后,在使用了两个云层检测数据集的对比实验中,FANet分别达到了96.07%和97.03%的准确率,与传统图像方法和其他基于深度学习检测的方法相比,FANet具有准确率高、云层检测边界清晰等优点。Currently,traditional image processing-based cloud detection methods rely on physical features such as texture,color,and temperature from specific sensors,which usually result in more errors and omissions when the subsurface of remotely sensed images has similar characteristics to the cloud layer.Inspired by the good performance of deep learning methods,this paper proposes a new cloud detection method FANet(feature fusion attention network)to solve this problem.FANet as a whole consists of feature extraction,decoding and attention modules.In the feature extraction module,FANet not only retains the top-down connections,but also adds several bottom-up and same-layer connections to shorten the one-way information flow and fusion path to alleviate the problem of feature information loss.In addition,the model introduces an attention module to calculate the weights of the fused feature maps at each level,so that the multi-layer features can be fused more accurately and form rich semantic and spatial features.Finally,in the comparison experiments using two cloud detection datasets,FANet achieves the accuracy of 96.07%and 97.03%,respectively,which has the advantages of high accuracy and clear cloud detection boundaries compared with traditional image methods and other deep learning-based detection methods.

关 键 词：云层检测 深度学习 注意力机制 特征融合 图像分割 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]