APSIPA Transactions on Signal and Information Processing > Vol 8 > Issue 1

Large-scale Landsat image classification based on deep learning methods

Xuemei Zhao, Aerospace Information Research Institute, Chinese Academy of Sciences, China AND School of Electronic Engineering and Automation, Guilin University of Electronic Technology, China, Lianru Gao, Aerospace Information Research Institute, Chinese Academy of Sciences, China, gaolr@radi.ac.cn , Zhengchao Chen, Aerospace Information Research Institute, Chinese Academy of Sciences, China, Bing Zhang, Aerospace Information Research Institute, Chinese Academy of Sciences, China AND College of Resources and Environment, University of Chinese Academy of Sciences, China, Wenzhi Liao, University of Strathclyde, UK AND IMEC-Ghent University, Belgium
 
Suggested Citation
Xuemei Zhao, Lianru Gao, Zhengchao Chen, Bing Zhang and Wenzhi Liao (2019), "Large-scale Landsat image classification based on deep learning methods", APSIPA Transactions on Signal and Information Processing: Vol. 8: No. 1, e26. http://dx.doi.org/10.1017/ATSIP.2019.18

Publication Date: 06 Nov 2019
© 2019 Xuemei Zhao, Lianru Gao, Zhengchao Chen, Bing Zhang and Wenzhi Liao
 
Subjects
 
Keywords
Large-scale image classificationLandsat image classificationCNNtransfer learning
 

Share

Open Access

This is published under the terms of the Creative Commons Attribution licence.

Downloaded: 1507 times

In this article:
I. INTRODUCTION 
II. METHODS 
III. RESULTS AND ANALYSES 
IV. CONCLUSIONS 

Abstract

Deep learning has demonstrated its superiority in computer vision. Landsat images have specific characteristics compared with natural images. The spectral and texture features of the same class vary along with the imaging conditions. In this paper, we extend the use of deep learning to remote sensing image classification to large geographical regions, and explore a way to make deep learning classifiers transferable for different regions. We take Jingjinji region and Henan province in China as the study areas, and choose FCN, ResNet, and PSPNet as classifiers. The models are trained by different proportions of training samples from Jingjinji region. Then we use the trained models to predict results of the study areas. Experimental results show that the overall accuracy decreases when trained by small samples, but the recognition ability on mislabeled areas increases. All methods can obtain great performance when used to Jingjinji region while they all need to be fine-tuned with new training samples from Henan province, due to the reason that images of Henan province have different spectral features from the original trained area.

DOI:10.1017/ATSIP.2019.18