卷积神经网络和特征融合的遥感图像场景分类  

Remote sensing image scene classification based on convolution neural network and feature fusion

作　　者：何婧媛[1] 谢生龙 田原[1] 田琴琴 HE Jingyuan;XIE Shenglong;TIAN Yuan;TIAN Qinqin(School of Mathematics and Computer Science,Yan’an University,Yan’an 716000,China)

机构地区：[1]延安大学数学与计算机科学学院,陕西延安716000

出　　处：《电子设计工程》2022年第16期60-64,共5页Electronic Design Engineering

基　　金：2018年延安大学校级科研计划项目(YDQ2018-05);2018年延安大学校级科研计划项目(YDQ2018-08);2019年度陕西省教育厅科学研究专项计划项目(19JK0964)。

摘　　要：针对遥感图像场景分类中多尺度特征提取困难和特征表征能力弱等问题,提出一种基于卷积神经网络和多尺度特征融合MFF-CNN(Multiscale Feature Fusion-Convolution Neural Network)的遥感图像场景分类方法。对MFF-CNN网络的高层特征和低层特征进行提取,采用跳跃式特征融合方式对其进行特征融合,将融合后的特征与全连接层输出作为高层特征,利用多核支持向量机MKSVM(Multi-Kernel Support Vector Machine)进行精确分类。将现有基于深度学习方法与MFF-CNN在3个公开的遥感图像场景分类基准数据集上进行验证,结果表明,MFF-CNN算法总体精度更高。Aiming at the difficulty of multi-scale feature extraction and weak feature representation ability in remote sensing image scene classification,a method based on Multiscale Feature Fusion-Convolution Neural Network(MFF-CNN) is proposed. the high-level features and the low-level features of the MFFCNN network are extracted. the jump feature fusion method is adopted. The fused features and the output of the full-connection layer are regarded as the high-level features to accurately classified by Multi-Kernel Support Vector Machine(MKSVM). The existing method based on deep learning was tested with MFFCNN on three public benchmark data sets of remote sensing image scene classification. The experimental results show that the MFF-CNN algorithm has a higher overall accuracy.

关 键 词：遥感图像 场景分类 深度学习 多尺度特征融合 

分 类 号：TN919.81[电子电信—通信与信息系统]