Research On Remote Sensing Change Detection Method Based On Lightweight Large Kernel Convolution And Distance Criss Cross Attention

Remote sensing images are one of the important means to detect changes on the Earth’s surface.The development of high-resolution remote sensing images has promoted the progress of deep learning change detection methods.Currently,the mainstream method is to build a shared-weight dual-branch network based on the idea of Siamese network,and improve the model detection performance by increasing the model width or depth,but also increase the model’s computational complexity.The development of high-resolution images has put forward higher requirements for change detection tasks,not only to determine whether there is a change in the object,but also to understand the type of change.In recent years,deep learning methods have performed well in change detection tasks.This thesis focuses on the research of twoclass and multi-class change detection methods based on deep learning,proposes a lightweight change detection model,and uses multiple datasets to train the model to verify its performance.The main research contents of this thesis are as follows:(1)In order to achieve a balance between the performance and parameter quantity of the change detection model,a change detection network based on lightweight large kernel convolution(LWLKC＿CD)is proposed.LWLKC＿CD is a fully convolutional neural network with a Siamese structure.It uses decomposed large kernel convolution to improve the Conv Next convolution block in the encoding end,extract multi-scale differential features while enhancing the model’s receptive field,and obtain more spatial information of the image context.The full-scale feature fusion module is used to fuse the multi-scale differential features extracted from the encoding end and send them to the decoding end to reconstruct the change binary image.Experimental results show that compared with the classic change detection model SNUNET,LWLKC＿CD can increase the IOU score by 3.19% in the large public dataset CDD(Change Detection Dataset),and reduce the computational complexity by 11.56 GFlops.The model has better detection effect on small target object changes and edge changes.In the small and medium-sized datasets SYSU(Sun Yat-Sen University)and XAHRCD(Xi’an High Resolution Change Detection Dataset),the performance gap between LWLKC＿CD and other change detection models will decrease as the dataset complexity decreases.(2)In order to reduce the problem of missed detection and false detection in multiclass change detection,the advantages and disadvantages of superimposed features and differential features in multi-class change detection tasks are analyzed,and the conclusion is drawn that a single feature fusion method cannot effectively detect multiclass object changes.Combining the superimposed features and differential features,the distance criss-cross attention mechanism is improved to obtain LWLKC＿DCCAM.The idea is to use the differential feature to generate the distribution of change attention in the image,allocate space weight to the superimposed feature to locate the change pixel,and judge the type of change through the original information-rich superimposed feature.The experimental results show that the MIOU score of LWLKC＿DCCAM is3.74% higher than that of the classic change detection model SNUNET.In the ablation experiment,it is 6.04% higher than LWLKC＿CD using differential features,and 1.49%higher than LWLKC＿CAT using superimposed features.In the visualization results,it shows certain advantages in multi-class change detection tasks.This thesis has 44 figures,11 tables and 110 references.