«上一篇
 文章快速检索 高级检索

 智能系统学报  2018, Vol. 13 Issue (4): 550-556  DOI: 10.11992/tis.201706078 0

### 引用本文

LI Yafei, DONG Hongbin. Classification of remote-sensing image based on convolutional neural network[J]. CAAI Transactions on Intelligent Systems, 2018, 13(4), 550-556. DOI: 10.11992/tis.201706078.

### 文章历史

Classification of remote-sensing image based on convolutional neural network
LI Yafei, DONG Hongbin
College of Computer Science and Technology, Harbin Engineering University, Harbin 150001, China
Abstract: The classification of remote-sensing images is a specific application of pattern recognition technology in the remote-sensing domain. In this paper, we propose a method for the classification of remote-sensing images based on convolutional neural networks (CNN). In addition, to address the difficulty of providing effective information regarding a single-source feature in convolutional neural networks, we propose a multi-source and multi-feature fusion method. We combine the spectral, texture, and spatial-structure features of remote-sensing images in the form of vectors or matrices according to their spatial dimensions, and train the CNN model using these combined features. The experimental results show that multi-source and multi-feature fusion can effectively improve the model convergence speed and classification accuracy, in comparison with traditional classification methods, and that the CNN method achieves higher classification accuracy and classification effect.
Key words: remote-sensing image    classification of land cover    convolutional neural networks    feature fusion

2006年，加拿大多伦多大学教授，机器学习领域的泰斗Hinton[9]和他的学生Salakhutdinov在《科学》上发表了一篇文章，掀起了深度学习在学术界和工业界的浪潮。深度学习是通过建立一种类似人脑分层的模型结构，对输入信息逐层进行特征提取，层级越深，提取的特征越抽象复杂，称为深度神经网络(deep neural networks，DNN)[10]。如今，深度学习作为机器学习的一个重要分支，已在图像识别、语音识别和自然语言处理等领域取得了巨大的成功[11-13]

1 卷积神经网络简介

 ${{{Y}}_i} = f({{{W}}_i} \cdot {{{Y}}_{i - 1}} + {{{b}}_i})$ (1)

 $f(x) = \left\{ {\begin{array}{*{20}{c}} {0,}&{x < 0} \\ {x,}&{x \geqslant 0} \end{array}} \right.$ (2)

 ${{{Y}}_i} = {\rm{subsample}}({{{Y}}_{i - 1}})$ (3)

 $L({{W}},{{b}}) = {\rm{CE}}({{W}},{{b}}) = - \sum\limits_{i = 1}^N {\sum\limits_{j = 1}^C {1\{ {{\hat y}_i} = j\} \log p_i^j} }$ (4)

 ${{{W}}_i} = {{{W}}_i} - \eta \frac{{\partial L({{W}},{{b}})}}{{\partial {{{W}}_i}}}$ (5)
 ${{{b}}_i} = {{{b}}_i} - \eta \frac{{\partial L({{W}},{{b}})}}{{\partial {{{b}}_i}}}$ (6)

2 基于CNN的遥感图像分类方法 2.1 CNN分类模型

AlexNet网络模型是由Alex Krizhevsky于2012年设计的一种深度卷积神经网络模型[15]，鉴于该模型层数不是很深，并且有很好的分类性能，因此本文以AlexNet模型为基础，构建了适用于遥感图像分类的CNN模型，其模型结构如图2所示。

2.2 多源多特征融合

 Download: 图 3 基于CNN模型的分类方法流程 Fig. 3 Flowchart of the CNN-based classification approach
3 实验与分析 3.1 实验环境

3.2 实验数据及样本选取

3.3 实验结果及分析