Research On Method Of Remote Sensing Image Change Detection Based On Integrated Residual Attention Unit And Depth-wise Separable Convolution Module

The research of remote sensing image change detection is an analysis of images taken of the same area at different times,resulting in the changes that occurred in the area’s land features during this time period,which is important in applications such as land cover and urban planning.With the improvement of surface observation technology,there has been a significant breakthrough in the resolution and clarity of remote sensing images,and the traditional methods are difficult to achieve good results on remote sensing images.However,the concept of deep learning techniques has led to a new way of exploring change detection research,which has also achieved good results.As the number of features in the current change detection datasets is becoming more complex and the types of changed are increasing,the problem of how to effectively extract features while avoiding the interference of pseudo-changed information in remote sensing images has become a pressing issue in the task of remote sensing image change detection research nowadays.As the wave of research into remote sensing image change detection grows,researchers are eager to achieve the goal of improved detection accuracy without considering the difficulties that their research has caused in terms of redundant network models and increasing computational costs while achieving better detection results.Therefore,there are two aspects to this paper:improving the change detection method to achieve higher detection accuracy in view of the high resolution and complex features of remote sensing images in the current datasets;and further weighing the contradiction between the high accuracy detection model build and lightweight network design to achieve the real-time demands of remote sensing image change detection.The following is the research presented in this paper.Deep learning-based change detection methods for remote sensing images can still be improved by effective acquisition of multi-scale feature and accurate detection of the edge of changed regions.This paper proposes a end-to-end change detection network,named the Multi-Scale Residual Siamese Network fusing Integrated Residual Attention(IRA-MRSNet),which adopts an encoder-decoder structure,introduces the Multi-Res block to extract multi-scale features,and uses the Attention Gates module before the skip connection to highlight the changed region features.This paper proposed an IRA unit,consisting of the Res2net~+module,the Split and Concat(Split and Concat,SPC)module,and the channel attention module,which can make the change detection results better through finer-grained multi-scale feature extraction.The experimental results show that the F1 and OA values of the network model outperform other state-of-the-art(SOTA)change detection methods on the CDD and SYSU-CD datasets,and the number of parameters and the calculated amount are reduced significantly.Taking into account the real-time demands of remote sensing image change detection,this paper proposes a lightweight remote sensing image change detection network model base on depth-wise separable convolution(Depth-wise Separable Convolution,DSC)module.The network uses the UNet architecture as the backbone,introduces DSC module for feature extraction,which significantly reduces the redundancy and memory consumption of the network model,and uses an efficient channel attention(Efficient Channel Attention,ECA)module to suppress irrelevant feature information.Meanwhile,a lightweight ASPP~+(Atrous Spatial Pyramid Pooling)module is proposed,which can effectively improve the detection capability of the network without adding additional parameters.The results show that the number of parameters in the network model is only 1.08 MB and the FLOPs are only7.14 G,and its results on CDD and SYSU-CD datasets still maintain high detection accuracy.