Research On Change Detection Of Multispectral Remote Sensing Image Based On Semantic Learning

Change detection of remote sensing is a technology to detect the changes of landform,by observing the multitemporal images of the same area.It is widely used in land monitoring and utilization,disaster management and regional planning.With the development of remote sensing technology,the resolution has been improved.The spectral heterogeneity of high resolution remote sensing image brings challenge to the traditional change detection technology.The process of change detection can be regarded as the segmentation problem of binary classification,which is based on the pixel pairs of multitemporal images.The semantic segmentation technology based on deep learning provides a new way to solve the problem of spectral heterogeneity.In this paper,the deep learning framework based on semantic segmentation is proposed to achieve the pixel by pixel end-to-end change detection.The main contents of this paper include the following work:(1)In the process of remote sensing change detection,small objects usually have lower priority in order to achieve the best overall accuracy.In this paper,the problem of small object processing in complex background is studied.Firstly,Multi-Scale Unet(MSUnet)from the perspective of semantic segmentation is proposed.By using the multi-scale features of the objects in remote sensing image,we can improve the segmentation performance and accurately identify the small objects in the complex background.Then,an end-to-end multi-scale change detection network(Siam-MSUnet)is established to achieve multi-scale change detection,by using Siamese network to obtain features of multitemporal images at the same time,to detect small changed areas effectively.In addition,in order to further improve the change detection,and solve the problem of insufficient sample data in the dataset,we use the attribute profile to filter the multitemporal images and achieve data enhancement.The experimental results show that these methods can make great improvement on change detection,even the kappa coefficient increased by 5%,and can accurately capture the edge information of small target changed area,and improve the change detection accuracy.(2)With the improvement of the resolution of remote sensing image,the change of uninterested area seriously interferes with change detection and reduces the accuracy.Therefore,this paper focuses on the interference problem caused by the change of uninterested region.In this paper,the Diff-Selective-Transform-Unet(DSTUnet)is proposed to detect the changes of the interested region in remote sensing image.We design an adaptive selection transform unit(DSTU)based on the difference attribute to fuse the shallow features and deep features in Unet,suppress the change information of the uninterested region in the encoding process,enhance the change information of the interested region,and avoid the interference caused by the change of the uninterested region.In addition,in order to solve the problem of sample imbalance between the changed and unchanged classes in the dataset,we propose a loss function that can effectively improve the accuracy.The loss function improves the performance by realizing the weighted combination of class balanced cross entropy loss function,dice coefficient loss function and L1 regularization.The experimental results show that the change detection based on DSTUnet can effectively identify the changes of interested region in remote sensing image,and by using the proposed loss function,the accuracy evaluation indexes of change detection are significantly improved,F1 score increased by 9%.In a word,this paper mainly studies these two problems: the detection difficulty of small target changed area in complex background and the interference of uninterested change area of change detection,and proposes corresponding solutions respectively,and proves that the proposed methods improve the change detection accuracy through experiments.