Object Change Detection Based On Deep Learning And Vectorization

With the popularization of high-resolution remote sensing image,the information contained in the image is more and more large.At the same time,due to vegetation occlusion and weather,there are a lot of interference information in the image,which brings great challenges to the change detection of typical target elements（road,building,water,vegetation,farmland,etc.）.Compared with raster data,vector data has strict data structure and less redundancy,which can more accurately express geographical location,facilitate network analysis and spatial query,and also facilitate attribute expression.But the existing vectorization algorithms ignore the expression of attribute information.To resolve the above problems,this paper transforms the change detection of typical ground object elements into multi classification problem,through the end-to-end deep learning network model,the change region and change type information are directly obtained from the input image with high-resolution remote sensing image.At the same time,based on the vectorization of linear target,the vectorization algorithm of multilateral shape target is proposed.The change detection result graph is vectorized,and the change type information is expressed.Finally,the vector data with attribute information is generated.The main work of this paper is listed below:（1）The preprocessing of remote sensing image and multi-objects change detection label annotation.The remote sensing data of the experimental area were calibrated by radiation and registration.The labels of typical surface target elements（road,building,water,vegetation and farmland）are obtained by manual labeling method.Then,the two periods of image label data are processed by difference operation.Based on the difference results,the change type of the corresponding area is judged,and the change detection label data is obtained.（2）Multi-objects change detection algorithm based on deep learning.For multi-objects change detection,it is transformed into a multi classification problem.The two period remote sensing images are combined into six channel data as the input of the model.The model is trained end-to-end to directly obtain the change area of the input image,and the change type of the change area is obtained according to the classification results.In order to solve the problem of low accuracy of direct application of U-Net model,this paper uses the residual module with stronger learning ability to replace the coding module of U-Net to build Res-Unet model.The final experimental results show that the accuracy of ResUnet is about 2% higher than that of U-Net,FWIo U is about 2.4%,and F1 is about 2.8%.（3）Vectorization of raster data.Aiming at the road raster data,a vectorization algorithm based on image thinning is proposed,which can extract the center line of the road accurately and quickly,and vectorize it with the attribute information to form the vector file.In the change detection result graph,the change area is composed of polygons,and the change types are represented by different gray values.When vectoring,we need to extract the boundary lines of each change area,then generate corresponding surface vector elements according to the boundary lines,and write the corresponding change information into the attribute information to form the vector file.（4）Software module design and implementation.The image basic operation function module,remote sensing image change detection function module and vectorization module are designed to realize the change detection and vectorization operation conveniently and quickly.