CD-TransUnet Based SAR Image Urban Building Change Detection

As cities expand and man-made disasters occur,the complexity of urban scenes is increasing.Buildings are an important part of a city,and their changes reflect the process of urbanization to a large extent.Accurate and efficient detection of buildings is a powerful way to obtain urban change information,which is also of great significance to urban management and planning.Compared with traditional optical remote sensing,SAR remote sensing is independent of atmosphere and sunlight,and has strong penetration.SAR remote sensing can penetrate vegetation canopy to detect blocked buildings.Especially in areas with dense vegetation and changeable weather,SAR remote sensing has unique advantages.Most traditional SAR image change detection algorithms suffer from slow speed and low accuracy.Deep learning-based algorithms have better feature extraction capabilities,but they also suffer from blurred building edges and wrong detection and omission of small area building changes.Moreover,deep learning networks are mostly accompanied by a large number of parameters and computational effort,with complex models and low detection efficiency.In response to the above problems,this paper uses L-band SAR(ALOS-2)image as the data source to carry out corresponding research on the deep learning-based SAR image urban building change detection method.The main research of this paper is as follows:(1)Production of data set.In this paper,a change detection dataset is produced based on dual-time L-band SAR images.Based on the characteristics of SAR images with high scattering noise and low contrast,a differential image is first constructed using the log-ratio algorithm,and then the original image information is introduced and superimposed with the differential image layer by layer to generate a three-channel color differential image,which can easily distinguish different changes of buildings according to the RGB color.Then the difference image is clipped and the sample is expanded by means of data augmentation.Finally,the data set used for change detection in this paper is obtained.(2)Research on building change detection algorithm based on improved Trans UNet network.The traditional deep learning model directly applied to this research paper does not achieve better detection results.Therefore,to address the problems of blurred building edges,error detection and omission of small building changes and sample imbalance in change detection,this paper improves the hybrid coded Trans UNet network,which mainly includes the following three aspects: 1)Adding coordinate attention module in the encoder and decoder parts to focus on the specific locations of the features;2)An improved ASPP module is added before up-sampling to obtain multi-scale information of the target,so that small and narrow buildings can be detected;3)Adopt the joint loss function.The experimental results show that the improved Trans UNet model is more suitable for SAR image urban building change detection compared with the original network,and can achieve better detection results.(3)Research on change detection algorithm based on lightweight CD-Trans UNet network.Since most of the CNN-based networks have the problems of large number of parameters and occupy many computational resources,in order to achieve more efficient change detection,this paper further optimizes the algorithm based on the improved Trans UNet network.Ghost bottleneck is used to replace traditional convolution in the encoder part,and standard convolution is replaced by depthwise separable convolution in the decoder part.A lightweight CD-Trans UNet model is constructed and applied to the detection of urban building changes in SAR images.The experiments demonstrate that the CD-Trans UNet model constructed in this paper has significantly reduced the number of parameters and computation,but it almost does not affect the accuracy of building change detection and achieves the lightweight model to a great extent.The comparison experiments also show that the CD-Trans UNet network proposed in this paper outperforms other commonly used change detection networks with higher accuracy and is more suitable for the research in this paper.