Research On Multi-scale High-precision 3D Scene Reconstruction Method Based On Synthetic Aperture Radar Data

Synthetic Aperture Radar(SAR)is an active microwave imaging technology with all-weather,all-day,high-resolution Earth observation capabilities,and has become an important mean of ground mapping.At present,multi-scale high-precision SAR 3D reconstruction(from large-scale terrain reconstruction to local point cloud fine reconstruction)is an important method for the construction of national spatial data infrastructure,which is widely used in water conservancy,transportation,national land,agriculture and forestry,environment,urban and rural planning,national defense and many other fields.In the process of 3D scene reconstruction of large-scale SAR data,various systematic and random errors seriously affect the positioning accuracy of SAR images and the elevation accuracy of the generated digital Surface Model(DSM)data.It is necessary to further merge multi-source data(SAR images,laser elevation control,reference DSM,etc.)to improve the positioning accuracy and 3D reconstruction accuracy of large-scale SAR images.In the process of locally 3D reconstruction,with the 3D interpretation capability towards the layover terrain in complex scenes of the tomography and array SAR system,it is necessary to further perform studies on multistrip SAR point cloud registration and complex targets(taking buildings as the obejct)3D reconstruction.Therefore,on the basis of summarizing and analyzing the existing research on 3D reconstruction by using SAR data,this paper further conducts research on methods ranging from large-scale 3D scene reconstruction of SAR images to locally fine 3D reconstruction of complex targets with SAR point clouds.We focus on large-scale SAR image block adjustment with reference DSM,terrain scene 3D reconstruction and subsequently block adjustment optimization with the constraint of laser point,local multi-strip SAR point cloud registration and 3D reconstruction for complex bulidngs.The main work and innovations of this paper include:(1)For the 3D terrain reconstruction of large-scale SAR images,a large-scale regional block adjustment and reconstruction method that integrates SAR images,laser elevation control,and reference DSM is established,which is mainly divided into two aspects:1)A large-scale SAR image block adjustment method is proposed,which is assisted by reference DSM,combined with optimal selection method of tie-points and selection weight iteration for error elimination.Aiming at the problem of low forward intersection accuracy caused by the weak intersection of SAR images,the reference DSM is introduced and read in chunks to assist the forward intersection solution,avoiding the problem of exceeded memory usage.Aiming at the problem that the mismatch tie-points affect the accuracy of the adjustment solution,a recursive iterative tie-points selection method is proposed,which breaks through the difficulty that the conventional ransac method is insufficient in accuracy when the mismatch points occupy a large proportion.In order to further improve the adjustment accuracy,a truncation value optimization method based on residual error histogram statistics is proposed to solve the problem that the error is easily affected by the accuracy of tiepoints during the iteration.The weighting strategy is optimized by counting the highest frequency of the residual value,by which the block adjustment accuracy achieve the sub-pixel level.Finally,the verification experiments are performed by using domestic and foreign spaceborne SAR iamges data,and then the application of RPC parameters,orthorectification and mosaic processing are carried out on the adjusted SAR images.In the large-scale block adjustment experiment of sentinel images with a resolution of20 m,the RMSE of the back-projection error after the initial block adjustment was6.030 pixels.The accuracy of the adjustment result was improved to 3.067 pixels with an increase of 49% after using the recursive iteration to select the tie-points.The accuracy was improved to 0.957 pixels after further optimizing the calculation method of the truncation value in the selection weight iteration.In the block adjustment experiment of Gaofen-3 images in Shanghai-Taihu area,the accuracy of the optimized block adjustment method was improved by 80%.Finally,the sentinel orthophotos of the large-scale scene(the absolute positioning accuracy was increased by 75%compared with that before the block adjustment)and the local Gaofen-3 orthophotos(the absolute positioning accuracy was increased by 50% compared with that before the block adjustment)are generated,the block adjustment significantly improved the relative and absolute positioning accuracy of SAR images.2)A large-scale SAR DSM data block adjustment method constrained by laser altimetry data is proposed.Interferometric synthetic aperture radar and stereo measurement methods are affected by various error sources in the process of 3D reconstruction.Existing studies have used high-precision control data to optimize the regional block adjustment of SAR DSM,but did not consider the initial positioning error of DSM.In response to this problem,this paper proposes a block adjustment method that combines with laser altimetry data and considers the initial position error of the DSM.By constraining the geographic location consistency and elevation consistency of the tie-points in the overlapping regions of the DSM,the plane and elevation block adjustment are successively performed,and finally improves the 3D accuracy of DSM.Aiming at the problem that the laser altimetry data and terrain are difficult to registration when there is no footprint image for reference,a bidirectional selection method combined with DSM terrain constraints is proposed to avoid false constraints caused by initial position offset,to achieve accurate extraction of laser control points.Finally,in order to verify the effectiveness of the proposed method,this paper conducts a block adjustment experiment on large-scale sentinel DSMs.The RMSE of the DSMs elevation error was decreased from 37.878 m to 9.619 m after the block adjustment,and the relative plane accuracy is also improved to sub-pixel level.At the same time,a joint block adjustment experiment of multi-source DSM is introduced,and the initial elevation error of the collected DSM is analyzed and modeled through laser altimetry data.After adjustment,the elevation accuracy of DSM was improved by more than 50%,which effectively promoted the elevation consistency of the overlapping area of multi-source DSM,and verified the effectiveness of the block adjustment method in application of multi-source DSM fusion.(2)For the 3D reconstruction of SAR point clouds of complex targets in local scenes,a multi-strip SAR point cloud registration and fine 3D reconstruction methods for building targets are established,which are mainly divided into two aspects:1)A coarse-to-fine adaptive registration method for urban SAR point clouds with low overlap rate is proposed.Due to the systematic errors in the processing of the data acquisition,the obtained multi-strip SAR point cloud data needs to be registered.And the SAR point cloud has the characteristics of low overlap rate,multiple noise points,unevenness,etc.So,the traditional point cloud registration method are difficult to apply directly.Aiming at this problem,this paper proposes an adaptive registration method for SAR point clouds in urban scenes based on iterative closest point method,completes the derivation of the improved optimization equation,and verifies the feasibility of the method theoretically.Aiming at the problem that building points are difficult to accurately extract in the case of multiple noise points,a method combining point density projection and histogram statistics is proposed,which breaks through the problem of extracting building standing points from SAR point clouds.And then rough registration is achieved through the extracted building facade points by using the correlation coefficient matching.Aiming at the problem of scarcity of conjugate features in the case of multi-strip SAR point clouds with low overlap rate,an automatic extraction and location constraint method for the concave-convex structure of building point clouds is proposed,which breaks through the problem of feature extraction of the same name in low-overlap rate SAR point clouds in urban scenes.Then,the rotation error correction and fine displacement compensation of the point cloud are realized.Finally,the effectiveness of the method is verified by simulation data and real SAR point clouds data.In the simulation data experiment,the conventional iterative neighbor method combined with fine displacement compensation requires conjugate features,and the registration shows instability,and the offset error after registration was more than 2m.The average rotation error was about 0.01°,and the offset translation error was less than 0.8 m.In the real data registration experiment,the results of the method proposed in this paper showed robustness in processing SAR point clouds of different real scenes.The angle deviation after registration was 0.4°,and the distance shift error between the target facade and the source facade was less than 0.8m,which was also better than the traditional iterative closest point method(angle error 1°,offset error more than 2m).2)An adaptive 3D reconstruction method of SAR point cloud based on optimized contour extraction is proposed for urban building targets.The basic process of 3D reconstruction includes roof point extraction and contour line extraction.In view of the characteristics of SAR point cloud with many noise points and uneven distribution,a method for extracting building roof points based on point density projection is proposed,and the morphological method is introduced to effectively realize the roof points extraction and hole filling.Aiming at the influence of noise on the edge points of the building roof point cloud,an adaptive contour extraction and optimization algorithm is proposed.The Alpha-Shape method is firstly used to extract the initial contour,and then the adaptive edge point serialization combined with the orthogonal constraints are used to realize the extraction of building outlines.Aiming at effective display of polygonal meshes of 3D models,the triangulation method of arbitrary polygons is introduced,which realizes the construction and correct visualization of the building base model(Level of Development 2,LOD2).Finally,in order to verify the effectiveness of the method proposed in this paper,two SAR point cloud datasets,scatter-corrected and uncorrected,in urban scenes were collected to carry out 3D reconstruction.The LOD2 building models were obtained through reconstruction,and the elevation accuracy of which is better than 1m verified by the high-precision Li DAR point cloud collected in the field.Finally,the research work of the paper is summarized,and the prospect of further work is proposed.In the finality,the problems requiring further studies are discussed.