Research On The Key Technologies Of Multi-source Remote Sensing Data In Surveying And Mapping Application

Synthetic aperture Radar(SAR),due to its all-weather and day/night capability,is more effective than optical sensor in a variety of complex environment and poor visibility meteorological conditions.Then,SAR is particularly suitable for information acquisition,sudden disasters(e.g.,floods,landslides,etc.),and crustal deformation monitoring.With the successive launch of recent SAR satellites,such as ALOS/PALSAR,COSMO-SkyMed,TerraSAR-X and Radarsat-2,the ground resolution of SAR images is raised to 1~5 meter,which is just to satisfy the requirements for image interpretation and object extraction.However,SAR image is very different from the optical image in the imaging mechanism,radiation and geometric characteristics.Therefore,it is difficult for image interpretation and to apply to surveying and mapping.This thesis explores the intelligent application of high-resolution SAR images in surveying and mapping field and focuses on automatic acquisition of ground objects,multi-source remote sensing images registration,and topographic change detection.The main contents of the thesis are as follows:(1)Intelligent extraction has been focused on typical ground objects in SAR images.For semi-automatic extraction,the threshold segmentation method based on gray level cooccurrence matrix is designed to extract the simple,homogeneous texture(e.g.water body and vegetation)and the region growing method based on Gabor transforms is proposed to extract the complex texture(e.g.residential area).For automatic extraction,the thesis proposed two kinds of extraction based on deep learning.The former one used multi-scale window and cumulative probability to solve the problem that the CNN-based segmentation method cannot determine the contour of ground objects.The latter method based on conditional generative adversarial nets(CGAN)extend the research scope from the supervised learning to unsupervised or semi-supervised learning.The experimental results show that semi-automatic extraction algorithms are easy to use and could achieve high positioning accuracy.The CGANbased method obtained high accuracy and the overall contours are more regular,which proves the ability and potential of deep learning in ground object extraction.(2)An integrative approach for multi-source remote sensing image registration and change detection based on ground objects is proposed.Based on cognitive model of human,we design algorithm flow and analyze key technologies and research emphases.The attribute and shape information of ground objects can be used for image registration.The related ideas and algorithms in computer vision field is introduced for shape matching.Considering the shape characteristics and segmentation problem of the linear and area objects,the extended optimal subsequence bijection(EOSB)and the shape context(SC)matching algorithm are proposed respectively.The ground objects are utilized as matching unit,which conform to the human visual cognition and avoid the correspondence problem between different features in multisource remote sensing images.The experimental results suggest that EOSB and SC algorithm has proven a strong robustness against noise,the difference of the extraction results,and local distortion.The integrative processing for change detection and image registration can make full use of all unchanged objects as the basis of image registration,more precise and more automation,the detection results can be used directly in topographic map revision and geographic national condition monitoring.(3)A novel multi-temporal DEM matching method based on a combination of contourderived point features and surface alignment is proposed to solve matching problem and terrain change detection in variable mountainous areas.The problem and causes of DEM matching in mountainous areas are analyzed.The the stable and invariant peaks are detected as control points using the similar characteristics of contour lines,which can solve the correspondence problem between two deformed surfaces.A least squares surface-matching method is used for optimization.The landslide is detected from the elevation difference after precise DEM matching.Compared with the actual landslide data,the veracity of DEM data and the reliability of the detection result can be verified.The experimental results indicate that the proposed method exhibits robustness to noise and artefacts caused by terrain changes and nonlinear distortion from DEMs that are derived from different systems.Two deformed DEMs can be sufficiently well aligned in coarse matching,even when georeferencing information is unavailable.The proposed method is more robust than conventional point-and surface-based methods.DEM-based approach not only can detect the region and scope of landslides,but also can detect the degree of elevation change in landslide area,which is more necessary for disaster assessment,rescue purposes and post-disaster reconstruction.