PS-DenseNet下的代数模型遥感图像场景分类研究  

Research on scene classification of remote sensing image based on algebraic model with PS-DenseNet

作　　者：陈垚[1] 张明波 CHEN Yao;ZHANG Ming-bo(Guang′an Vocational and Technical College,Guangan 638000,China;School of Statistics,Jiangxi University of Finance and Economics,Nanchong 330000,China)

机构地区：[1]广安职业技术学院,四川广安638000 [2]江西财经大学统计学院,江西南充330000

出　　处：《激光与红外》2022年第3期442-450,共9页Laser & Infrared

基　　金：国家自然科学基金项目(No.61601382)资助。

摘　　要：遥感图像能在短时间内获取大范围丰富的地表数据和细节,有效完成遥感图像场景的分类成为分析相关信息的重要依据。基于此,本文提出了PS-DenseNet下的代数模型遥感图像场景分类方法,为提升遥感图像的分类准确性,引入Lie group代数模型分析工程问题可获取底层特征并降低特征维度,并在Densenet基础上设计PS-DenseNet提取高层特征,进而采用焦点代价函数完成网络训练,所设计的深度网络模型可训练海量遥感图像样本特征,使模型具备较强的自学习功能;为校验本文方法的有效性,采用两类数据集完成本文方法和文献[5]、[6]方法的验证,实验结果表明,本文方法的OCA参数和Kappa参数均优于其他两种方法,并能在较少的epoch中达到分类精准度较高的稳定状态。Remote sensing images can obtain a wide range of rich land surface data and details in a short time.The effective completion of remote sensing image scene classification has become an important basis for the analysis of relevant information.Based on this,a remote sensing image scene classification method based on BMFD-NET algebraic model is proposed.In order to improve the classification accuracy of remote sensing images,Lie group algebraic model is introduced to analyze engineering problems to obtain low-level features and reduce feature dimensions,and PS-DenseNet is designed on the basis of DenseNet network to extract high-level features.Then the focus cost function is used to complete the network training.The designed deep network model can train the features of massive remote sensing image samples,so the model has a strong self-learning function,and the rich image features can be obtained automatically.In order to verify the effectiveness of the proposed method,two kinds of data sets were used to validate the proposed method and methods of reference[5]and[6].The experimental results show that the OCA parameters and Kappa parameters of the proposed method are better than those of the other two methods,and it could achieve a stable state with high classification accuracy in fewer epochs.

关 键 词：深度学习 分类 Densenet 代价函数 网络训练 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]