Evaluation of WORLDVIEW-1 Stereo Scenes

Abstract

This paper describes the investigations carried out at ETH Zurich on the very-high resolution optical satellite sensor WorldView-1.The dataset includes a stereopair acquired in November 2007 over Morrison, Colorado, USA, consisting of a quasi-nadir scenewithmean GSD close to 51cm and an off-nadir scene, with GSD up to 73cm. The radiometry of the images was evaluated through theestimation of the noise level (standard deviation of thedigital number) innon-homogeneous areas. For the orientation of the scenes,we used ground control points to estimate the parameters of an affine transformation and improve the orientation related to theRational Polynomial Coefficients (RPC) provided in the metadata files. The next step included the automatic DSM generationwithquality assessment. In all our tests the matching and the final DSM generation was realized with our in-house software package SAT-PP (Satellite ImagePrecisionProcessing) which has already been successfully used for other satellite sensors like IKONOS,ALOS/PRISM, Cartosat-1 and airborne linear and frame cameras. For the evaluation of the potential of a WorldView-1 stereo imagepair for DSM generation, reference data are fundamental. As no DSM with a suitable resolution was available in this dataset, wegenerated the reference DSM from 1:15’000 scale infrared aerial images purchased from USGS. The aerial images were acquired inApril 2002 with an analogue camera and scanned with a resolution of 14μm. No GCPs were available for the area covered by theaerial images. Therefore we defined a sufficient number of well-defined texture points in the oriented stereo pair of WorldView-1 asGCPs and determined their 3D coordinates using the orientation parameter of the satellite stereo pair. These points were used for theimage orientation of the aerial images, which leads to a good relative orientation between the two different image sets. By this weinvestigated the quality of the matching process with WorldView-1 stereo image data mostly independent of the exterior orientationparameters. We will present here our accuracy assessment by visual and quantitative analysis (3D residuals distribution andstatistics).The difference between the acquisition times of the two data sets (about 2.5 years) caused changes in the land cover, e.g.in correspondence of vegetation and buildings. In order to avoid the influence of temporal differences in our analysis, we comparedthe surface models in areas where no significant time depending changes could be expected.In addition, we evaluated the potential ofthe scenes for 3D building model extraction. The results are presented and discussed.