Research On Remote Sensing Image Change Detection Based On Deep Learning

The remote sensing image change detection task aims to obtain the location and change category of change regions in dual-time remote sensing image pairs,which has been widely used in the fields of disaster monitoring and unauthorized building monitoring.In order to solve the problem of high labor and time cost of early remote sensing image change detection,researchers have proposed traditional remote sensing image change detection methods such as algebraic method,machine learning based change detection method and object-oriented image analysis.However,such methods have very strict requirements on remote sensing image preprocessing,and the quality of image pre-processing directly affects the detection results,resulting in the generally low accuracy and cumbersome operation of such methods.With the improvement of remote sensing image resolution and the development of deep learning technology,the change detection methods based on deep learning have received more and more attention.This study starts from two aspects of improving the accuracy of remote sensing image change detection and introducing semi-supervised learning for the problem of difficulty in acquiring high-quality datasets,improving on the existing methods and proposing new solutions,and the specific work is as follows:(1)In order to improve the accuracy of remote sensing image semantic change detection,it is necessary to locate the location of remote sensing image change regions and their change categories more accurately,and an improved remote sensing image change detection method based on twin HRNet V2 network is proposed.Firstly,the twin network is combined with HRNet V2 to avoid the problems of diachronic image feature mixing and difficulty of difference map reconstruction brought by the earlier diachronic remote sensing image pair fusion method;then,the attention mechanism is introduced to improve the network performance in order to better utilize the semantic information in pixels.In addition,pseudo-labeled data enhancement is used to increase the number of pixels with specific semantic annotations to alleviate the problem of data imbalance;finally,the effectiveness of the proposed method is verified using the SECOND dataset,the effects of different feature extraction networks on the change detection accuracy are explored,the effectiveness of the attention mechanism and pseudolabeled data enhancement are verified by ablation experiments,and compared with advanced semantic change detection methods A comparison is made with advanced semantic change detection methods.Compared with the comparison method,the overall score of the proposed method is improved by nearly 6.7%,and the effectiveness of the proposed method is verified by effectively improving the problems of false detection,missed detection and hole phenomenon in local areas.(2)In order to overcome the problems of small data volume and low label quality of the current publicly available remote sensing image change detection datasets,an improved semisupervised remote sensing image change detection method based on consistent regularization is proposed to make full use of the information in unlabeled data to improve the performance of the change detection method.The method is based on twin networks,divided into two stages,supervised and unsupervised,with the same network architecture in both stages,and mainly improved with two modules of feature extraction and hidden feature difference extraction.The feature extraction module combines the twin network with the Conv Ne Xt network to have more powerful feature extraction capability;the ASPP module is added to the hidden feature difference extraction module to better fuse the multi-scale information of objects;the information in the unsupervised loss term is obtained in the unsupervised labeled data by using three random perturbations with consistency constraints.In the case of 5% labeled data,the proposed method improves the Io U score by 1.4% on the LEVIR-CD dataset and 2.2% on the WHU dataset compared to the comparison method,indicating that the proposed method can more effectively utilize the information from unlabeled data,proving the effectiveness of the proposed method.