Building Change Detection With High-Resolution Remote Sensing Imagery Based On Deep Self-training

With the acceleration of global urbanization,the phenomenon of illegal and unauthorized building construction has gradually come to the fore,which seriously hinders the cities from sustainable development.Since China is now in a stage of rapid urban development,it is necessary to introduce more change detection techniques based on high-resolution remote sensing images for high-efficiency building monitoring.In recent years,it is more and common for the researchers to apply deep learning algorithms to remote sensing imagery change detection.The methods proposed in this paper aim to address the following problems,such as the blurred edges and discontinuous geometric structures of building change detection results as well as the high requirements for training data and annotation cost.Accordingly,a novel double-branch network W-Net(Double U Network)and a revised S~3W-Net(Self-training and Semi-Supervised learning Double U network)are designed to improve change detection accuracies in large scenes,which combines super-pixel constraints with traditional change detection structures based on deep learning.Self-training and semi-supervised learning are introduced into the S~3W-Net to better the model’s performance.The main research of the paper is as follows:(1)A double-branch structure model coupled with a change detection branch and a super-pixel branch is constructed to address the problem of information loss and geometric structure fragmentation of building details during multi-layer feature extraction in deep networks.A dual-branch multi-task coupling framework of change detection and superpixels is innovated to effectively improve the edge blurring phenomenon of detecting buildings through the auxiliary inference of superpixels.The method proposed in the paper is validated on two public datasets(LEVIR-CD dataset and WHU-building dataset).W-Net obtained the best performance with F1-score of0.9031 and Kappa coefficient of 0.8969 on LEVIR-CD dataset,and F1-score of 0.9172and Kappa coefficient of 0.9142 on WHU-building dataset.The experimental results show that the proposed algorithm can better detect the edges of changing buildings,maintain the complete geometric structures,and outperform other change detection methods.(2)To address the problems of unstable performance and inadequate training of remote sensing imagery change detection algorithms in large scenes with limited labeled samples,a self-training semi-supervised learning algorithm is introduced,and a remote sensing image change detection model(S3W-Net)based on deep semi-supervised learning is proposed,which can expand the training prior knowledge of the detection network by mining the dual-temporal image change difference features and improve the change detection algorithm in large-scene application performance.The proposed method is validated on two public datasets(LEVIR-CD dataset and WHU-building dataset),with F1-score of 0.9170 and 0.9219,and Kappa coefficient of 0.9089and 0.9152 respectively.Its performance is better than other methods compared with other semi-supervised algorithms.In addition,we self-labelled and constructed a new change detection dataset based on remote sensing images on Jiading district in Shanghai,and tested the algorithm on this dataset with F1-score of 0.9321 and Kappa coefficient of 0.9281.The experimental results show that the proposed deep semi-supervised learning change detection method can achieve higher detection accuracy with limited labeled samples.(3)Facing the application problem of remote sensing image change detection algorithm in practical engineering,a prototype system with interactive interface for building change detection of high-resolution remote sensing images is developed based on Python3.7.The interactive interface of the software is built with Py Qt5 framework,and the processing functions of remote sensing image data are realized based on Pytorch,Numpy,Open-cv,Matplotlib and other libraries.The software integrates the newly-proposed W-Net,S3W-Net network and the classical change detection algorithm,which solves the problems of complex change detection process.