Research on High Resolution Remote Sensing Image Classification Based on Convolution Neural Network

Wenwen Gong,Xin Fan,Junmeng Hao,Yong Liang,Zhuqing Wang

doi:10.1007/978-3-030-06137-1_9

Wenwen Gong, Xin Fan + Show 3 more

Open Access

https://doi.org/10.1007/978-3-030-06137-1_9

Copy DOI

Publication Date: Jan 1, 2019
Citations: 2	License type: cc-by

Affiliation: Shandong Agricultural University, Taishan University

Abstract

Traditional classification method based on machine learning algorithm has been widely adopted in very high resolution remote sensing image classification, yet the problem that could not effectively convey a higher level of abstract feature still need to be improved. This paper, relying on the convolution neural network algorithm, has conducted on the high-resolution remote sensing image classification method. Firstly, structure of convolution neural networks was analyzed. The prediction model of convolution neural networks was discussed, and the core of structure was the alternation of the convolution layer and the down sampling layer. Then, the training model of convolution neural networks was researched. By using weights sharing and local connection, convolution neural network, that image could directly entered into, avoids to a certain extent caused by image displacement, dimension change and so on. On this basis, basing on different phase GF-1 remote sensing data and MATLAB development environment under Windows10 operating system, then combining with object-oriented classification technology in image segmentation, this paper built the high resolution remote sensing image classification model based on convolution neural network. Finally, the parameters of the model were tested and analyzed repeatedly, and more accurate model parameters were obtained in this paper. Results show that the mode can effectively improve the classification accuracy, and provide technical support for improving remote sensing image interpretation and formulating sustainable development strategy.

Highlights

Remote sensing image classification, which is one of the important ways of remote2 sensing image analysis and interpretation, is a concrete application of pattern recognition in the field of remote sensing
In order to solve this problem, a lot of space geometric information has been integrated into the process of high resolution remote sensing image classification; at the same time, Machine Learning (ML) algorithms are introduced, such as Support Vector Machine (SVM) [3], Random Forest(RF)[4], and Neural Networks(NN)[5]
Because the traditional classification method based on spectral characteristics and multi-feature classification method based on ML algorithm belong to shallow learning algorithm, which only could extract the lower level features, such as the spectrum, texture and shape, and couldn’t effectively express the abstract characteristics of higher level

Summary

Introduction

Remote sensing image classification, which is one of the important ways of remote2 sensing image analysis and interpretation, is a concrete application of pattern recognition in the field of remote sensing. Because the traditional classification method based on spectral characteristics and multi-feature classification method based on ML algorithm belong to shallow learning algorithm, which only could extract the lower level features, such as the spectrum, texture and shape, and couldn’t effectively express the abstract characteristics of higher level. These methods are difficult to cross the semantic gap between low-level and high-level, and no longer applicable in current big data environment

Methods

Results

Discussion

Conclusion