Open AccessDetection of Small Changed Regions in Remote Sensing Imagery Using Convolutional Neural Network
Author(s) -
Zhaobin Cao,
Mengmeng Wu,
Rui Yan,
Fa Zhang,
Xiaohua Wan
Publication year - 2020
Publication title -
iop conference series. earth and environmental science
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.179
H-Index - 26
eISSN - 1755-1307
pISSN - 1755-1315
DOI - 10.1088/1755-1315/502/1/012017
Subject(s) - change detection , computer science , fuse (electrical) , remote sensing , satellite imagery , convolutional neural network , artificial intelligence , high resolution , pattern recognition (psychology) , deep learning , image resolution , geology , engineering , electrical engineering
As a crucial application in remote sensing, change detection can detect semantic change of the land use in multi-temporal aerial imagery. Especially, with the improvement of the spatial resolution of remote sensing imagery, more and more small objects have been taken into account. Although several methods based on deep learning have been proposed to solve change detection in satellite imagery, most of existing methods cannot properly handle the small changed regions and the class imbalance, which are extremely serious in satellite imagery. In this paper, we propose a Si amese H igh- R esolution Net work (Si-HRNet) to detect the small changed regions in remote sensing imagery. To our best knowledge, this is the first time to transfer High Resolution (HR) module into the Siamese network, so that our network can maintain high-resolution representation throughout the whole process and repeatedly fuse multi-resolution representations to obtain rich feature representations, especially can reduce information loss for small objects. In addition, to handle the class imbalance issues, we combine weighted binary cross entropy (BCE) and inverse volume weighted generalized dice loss (GDL) in small objects change detection. Experimental results show that the proposed Si-HRNet achieves state-of-the-art performance in both DigitalGlobe dataset and WHU Building change detection dataset, and the F1 score is improved by 3.35% ∼ 3.43%.