测绘通报 ›› 2021, Vol. 0 ›› Issue (8): 14-21.doi: 10.13474/j.cnki.11-2246.2021.0233

• 学术研究 • 上一篇    下一篇

基于LDCNN特征提取的多核SVM高分辨率遥感影像场景分类

耿万轩, 周维勋, 金双根   

  1. 南京信息工程大学遥感与测绘工程学院, 江苏 南京 210044
  • 收稿日期:2020-09-17 修回日期:2021-03-28 发布日期:2021-08-30
  • 通讯作者: 周维勋。E-mail:zhouwx@nuist.edu.cn
  • 作者简介:耿万轩(1998-),男,硕士,主要研究方向为遥感图像智能处理。E-mail:20191211014@nuist.edu.cn
  • 基金资助:
    中科院先导A专项课题(XDA23040100);南京信息工程大学人才启动经费(2019r085)

Scene classification of high-resolution remote sensing image based on multi-kernel SVM using features extracted from LDCNN

GENG Wanxuan, ZHOU Weixun, JIN Shuanggen   

  1. School of Remote Sensing & Geomatics Engineering, Nanjing University of Information Science & Technology, Nanjing 210044, China
  • Received:2020-09-17 Revised:2021-03-28 Published:2021-08-30

摘要: 针对卷积神经网络特征维度高且单层特征不能准确表达复杂高分辨率遥感影像语义信息的问题,本文提出了一种提取低维卷积神经网络(LDCNN)深层次特征进行多核SVM分类的场景分类方法。首先将预训练的卷积神经网络改造成低维网络结构,其次提取低维网络的不同深层特征并进行不同核函数的SVM分类,找到对应的最优核函数;然后将多种最优核函数加权融合成为一个新的合成核;最后进行多核SVM分类。试验表明,本文方法不仅特征维度低,且通过多核SVM能够充分结合各层特征的优点,在两个标准数据集上均取得了99%以上的分类精度。此外,该试验还证明了本文方法具有较强的迁移学习能力。

关键词: 高分辨率遥感影像, 场景分类, 卷积神经网络, 特征提取, 多核SVM

Abstract: Due to the problem that the feature dimension of convolutional neural network is high and the single-layer feature cannot accurately express the complex semantic information of high-resolution remote sensing image, a scene classification method of the low dimension of convolutional neural network (LDCNN) based on multi-kernel SVM is proposed in this paper. Firstly, the pre-trained convolutional neural network is modified into a low-dimensional network structure. Then, different high-level features extracted from the low-dimensional network is performed to find the corresponding optimal kernel function via SVM classification using different kernel functions, and these multiple optimal kernel functions are fused into a new composite kernel. Finally, multi-kernel SVM classification is carried out. Experimental results show that the proposed method has low feature dimension, and can combine the advantages of features extracted from each layer via multi-kernel SVM, thus achieving more than 99% classification accuracy on two benchmark datasets. In addition, the experiment also proves that this method has strong transfer learning ability.

Key words: high-resolution remote sensing image, scene classification, convolutional neural network, features extraction, multi-kernel SVM

中图分类号: