Research On Building Change Detection In High-resolution Remote Sensing Images Based On Twin Network

With the continuous development of aerospace remote sensing technology,the spatial resolution of remote sensing images is increasing while the acquisition of image data is becoming easier.High resolution remote sensing images are gradually becoming the preferred data source for remote sensing research due to their rich feature texture information,more complex shape information.Buildings are the main indicator to measure the urbanization of a region.Rapid and accurate identification of building changes plays a vital role in urban unauthorised building management,urban sprawl and disaster analysis.The detection of building change areas using high resolution remote sensing image is one of the current branches of remote sensing image change detection research.Unlike other types of change detection,buildings have different shapes,heights,materials and colors,when using traditional change detection methods to identify buildings in high resolution images,the phenomenon of "salt and pepper noise" is prone to occur.The traditional method also poses certain challenges due to the pseudo-variation of the image caused by solar altitude angle,light,cloud cover,atmosphere and the sensor itself.In recent years,computer software and hardware are constantly being updated,and artificial intelligence technology represented by deep learning has emerged and shone in the field of remote sensing image processing,therefore,this paper uses deep learning technology to study the problem of building change detection in high-resolution remote sensing images.The main research is as follows.(1)Building change detection method based on convolutional neural network.In this paper,a Siamese neural network based on convolutional operation is constructed for building change detection in high-resolution remote sensing images,and the deep features of the before and after dual-temporal images are extracted by Siamese networks respectively.In the process of extracting features,short-circuit connections are used to deepen the connection between shallow and deep layers to avoid the phenomenon of gradient disappearance,and a multi-scale feature extraction mechanism composed of null convolution is used in the deep layer of the decoder to extract deep features In the decoding process,a learning up-sampling method is used to recover the image features,and in the output process,a depth supervision strategy is used to fuse the output between different layers.(2)Building change detection method based on Transformer.Considering that the convolution operation cannot take into account the relationship between the surrounding pixels of the image and cannot establish long-distance modeling,this paper constructs a twin network based on Transformer.The twin sub-network uses the multi-head attention of the Transformer to extract image features.Although the attention operation can extract image features globally,it is insufficient in local fine-grained feature extraction.Therefore,after the Transformer structure,the convolution operation composed of dilated convolutions will be used to enhance local features.Transformer and convolution operations Complementary,the global and local features of the image can be obtained.During the decoding process,the features extracted by the Transformer and the convolution will be connected at the same time using the skip connection mechanism.(3)Use the LEVIR-CD dataset and the WHU building dataset to conduct comparative experiments and ablation experiments respectively,and compare the network proposed in this paper with FC-EF,Seg Net,FC-Siam-Diff,FC-Siam-Conv,UNet++＿MSOF,IFN,SNUNet and BIT and other advanced models are tested.The results show that the network in this paper has good performance in the building change detection task compared with the same type of network.The average F1 score of the Siamese convolutional network on the two data sets has reached 87.85%,the F1 score of the twin Transformer network on the two data sets reached 91.05%,both networks are higher than the comparison experiment network,and the false detection and missed detection areas in the visualization results are also the least.