Research On Semantic Segmentation Of Dual Source Remote Sensing Images Based On U-Net And Attention Mechanism

Remote sensing image semantic segmentation is one of the key research issues in the field of remote sensing image processing.Its main connotation is to introduce high-level semantic information of the image in the remote sensing image segmentation process,and assign a specific semantic label to each pixel,so as to accurately label the category of each pixel.High-resolution remote sensing images record detailed information such as the shape,geometry,texture,and other features of ground objects,and are widely used in various fields of remote sensing,such as land use and sea segmentation.With the development of remote sensing images from low resolution to high,high-resolution remote sensing images,while providing high-quality information,are characterized by large intra-class differences and relatively lacking spectral information for efficient and accurate remote sensing image semantic segmentation presents new challenges.Although the method based on deep learning greatly improves the performance of semantic segmentation,there are still some problems:(1)the data sources of high-resolution remote sensing images are diverse,the single source has low segmentation accuracy and the low utilization rate of dual source data;(2)the remote sensing image has a small target scale and indistinguishable texture details,which leads to the ambiguity of the semantic segmentation of the remote sensing image.And the additional detailed information brings considerable interference to the classification learning process,this reduces the accuracy and leads to blurred object boundaries.Therefore,in order to make better use of the data information of high-resolution remote sensing images,solve the problem of small-scale target segmentation of remote sensing images and improve the positioning ability of target boundaries,this paper proposed two semantic segmentation methods of dual-source remote sensing images.The two methods are described below:(1)Considering the multi-source of remote sensing image data,and the different feature information contained in different source of high-resolution remote sensing image,in order to improve the utilization of dual-source remote sensing image and realize the joint use of different source data,one semantic segmentation method of dual-source remote sensing image based on U-Net,named DU-Net,is proposed.First,considering the advantages of the encoder-decoder model and the residual structure in semantic segmentation,RU-Net based on the residual structure is proposed.Second,digital surface model(DSM)is introduced at the encoder stage,combined with optical remote sensing image and its corresponding digital surface model data,a dual source remote sensing image semantic segmentation model DUNet is proposed.The model is mainly to fuse each intermediate layer feature of the optical remote sensing image and its DSM data to obtain a more differentiated dual-source fusion feature,and then gradually upsample the fusion feature combined with shallow features,to effectively complete the processing of dual-source image data to achieve the purpose of further improving the accuracy of semantic segmentation.Secondly,using the transfer learning method,Res Net101 is used as a pre-training model to initialize network model parameters.Finally,a comparative experiment is designed on the ISPRS dataset to verify the effectiveness of the proposed method.(2)In the semantic segmentation of remote sensing images,the problem of small-scale target segmentation and the problem of fuzzy object boundaries caused by additional details,a semantic segmentation method for dual-source remote sensing images based on the DU-Net model,named ADU-Net,is established.The model ADU-Net fully considers the channel relationship between pixels.First,design a pyramid pooling module based on the channel attention mechanism to extract context information at different scales,in order to add the multi-scale features to the model.At the same time apply this module to each level of the structure to extract the multi-level features of the image,making full use of the multi-scale features and multi-level features of the remote sensing images.Second,an attention refinement module is designed to collect redundant channel information.And this model comprehensively utilizes the shallow spatial detail information and high-level semantic information,to achieve the goal of refining the semantic object segmentation of remote sensing images.Finally,a comparative experiment is designed on the ISPRS dataset to verify the effectiveness of the proposed method.