基于多尺度特征提取的遥感图像分类  

Remote Sensing Image Classification Based on Multi-scale Feature Extraction

作　　者：罗豪 李先锋[1] LUO Hao;LI Xianfeng(School of Information Engineering,Yancheng Institute of Technology,Yancheng 224000,China)

机构地区：[1]盐城工学院信息工程学院,江苏盐城224000

出　　处：《计算机与现代化》2025年第3期86-92,共7页Computer and Modernization

基　　金：国家自然科学基金资助项目(62076215)。

摘　　要：为了提高卷积神经网络在遥感图像分类任务上的准确率,提出一种基于多尺度特征提取的分类方法。针对遥感图像中目标物体尺度差异大和普通卷积提取特征容易产生冗余特征的问题,提出普通卷积与空洞卷积共同作用的多尺度混合卷积模块,有效增强模型特征提取能力;在特征融合方面,提出一种特征交叉融合模块,有效融合各支路的语义信息,充分利用各个尺度特征之间的信息进行深度融合与交互。针对遥感图像中地物信息复杂的问题,设计了并行注意力模块对每一个支路运用注意力机制,使其更加关注图像的关键部分,忽略冗余信息。与现有方法对比,本文方法的分类性能有较为明显的提升。在数据集WHU-RS19、RSSCN7和SIRI-WHU上,本文方法的整体分类准确率分别达到了98.19%、94.18%和96.37%。In order to improve the accuracy of convolutional neural network in the classification task of remote sensing image,a method based on multi-scale feature extraction for classification is proposed.Aiming at the problems that the scale difference of target objects in remote sensing images is large and ordinary convolutional extraction tends to produce redundant features,a multi-scale hybrid convolution module with ordinary convolution and atrous convolution is proposed,which can effectively en⁃hance the feature extraction ability of the model.In terms of feature fusion,a feature cross fusion module is proposed,which can effectively fuse the semantic information of each branch and make full use of the information between various scale features for deep fusion and interaction.Aiming at the problem of complex land cover information in remote sensing images,a parallel atten⁃tion module is proposed to apply attention mechanisms to each branch,which makes it pay more attention to the key parts of the image and ignore the redundant information.Compared with the existing methods,the classification performance of the proposed method is significantly improved.On the data sets WHU-RS19、RSSCN7 and SIRI-WHU,the overall accuracy of the proposed method reaches 98.19%、94.18%and 96.37%,respectively.

关 键 词：遥感图像 注意力机制 卷积神经网络 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]