Page - 87 - in Joint Austrian Computer Vision and Robotics Workshop 2020

GroundControlPointRetrievalFromSARSatellite Imagery RolandPerko,HannesRaggam,KarlheinzGutjahr JOANNEUMRESEARCHForschungsgesellschaftmbH,DIGITAL {roland.perko,hannes.raggam,karlheinz.gutjahr}@joanneum.at WolfgangKoppe, Ju¨rgenJanoth AirbusDefenceandSpace {wolfgang.koppe,juergen.janoth}@airbus.com Abstract. Formany applications, like for instance autonomous driving or geo-referencing of optical satellite data, highly accurate reference coordinates areof importance. Thisworkdemonstrates that such GroundControlPoints canautomaticallybederived frommulti-beam Synthetic Aperture Radar satellite imageswithhighaccuracy. 1. Introduction Reliable Ground Control Points (GCPs), i.e., points of known geographical coordinates, are an essential input for the precise ortho-rectification of remote sensing imagery, the exact location of tar- gets or the accurate geo-referencing of a variety of geo-datasets. AlthoughGCPscollectedby terrestrial means typically offer a high accuracy, their acquisi- tion is expensiveespeciallyonaworldwide level. Thus, a concept was formed to extract such GCPs from Synthetic Aperture Radar (SAR) satellite im- ages (e.g., [9, 11]). Recently, refined SAR-based GCP extraction emerged due to three main reasons: (1) The 2D geo-location accuracy of current SAR sensors is very high, actually at centimeter level if atmospheric effects and Earth surface displacements are taken into account [5]. (2) Metallic objects like lamp poles or traffic signs (i.e., common features in urban scenes) appear as focused points in SAR im- ages and can be detected with subpixel accuracy. (3) Using stereo acquisitions the 3D position (actually the ground mark) of these objects can be computed bymeansof radargrammetry. Therefore, this work presents an automatic work- flow, combining techniques from photogrammetric computer vision and remote sensing, that derives Figure 1. Stereo acquisition from space. Shown are two SAR satellites observing the same region on ground from twodifferentorbital directions and lookangles. highlyaccurateGCPsfromasetofmulti-beam1 high resolutionimagesfromTerraSAR-X,TanDEM-X,or PAZ satellites [3]. In contrast to [11], where persis- tentscatter interferometry(PSI) isdeployedforpoint detectionand3Dreconstruction,webuilduponcom- puter vision paradigms. Thus, the presented method canbeefficientlyappliedonsingle imageswhilePSI needsastackofmultiple imagesand iscomputation- allyverydemanding[4]. Inaddition,ourmethodcan be applied on amplitude images alone as it does not relyon thephase informationof the signal. 2.Method The proposed fully automatic workflow for GCP retrieval consistsof the followingsteps: 1The termmulti-beam is equivalent to what is calledmulti- view in computervision and stems from digital beamforming. 87