3-D Landmarks Derived from TerraSAR-X Data as Ground Control Information for Optical Imagery

Abstract

Without ground control information, the uncertainty in the absolute geographic positioning of optical satellite imagery is in the order of few meters. These inaccuracies originate predominantly in uncertainties of angular measurements for the sensor attitude. In contrast, radar systems like Synthetic Aperture Radar (SAR) sensors do not measure angles but signal runtimes which are precisely convertible to spatial distances provided that all signal propagation effects are thoroughly considered. With the identification of corresponding point-like image features in SAR and optical images, the precise position measurements of such objects in SAR imagery figure as Ground Control Points (GCPs) for the precise absolute georeferencing of the optical data at submeter level.

In particular in urban areas, myriads of point-like scatterers are observable in SAR images. Many of these objects can be identified as poles of lamps or traffic signs. The detection accuracy of their subpixel position depends on the signal to clutter ratio of their point response and typically spans from decimeter down to centimeter level. Several poles detected in the SAR imagery are as well distinguishable in optical images from aerial or spaceborne sensors.

The radar backscatter geometry of a pole is a double bounce via the edge of the pole and the ground in front of it. The phase center of the signal double reflection is located at the base point of the pole faced toward the SAR sensor. Depending on the backscatter properties of the ground, poles are well detectable from the street side having the asphaltic street surface in front but rarely detectable from the rear side if there is uneven terrain.

A single SAR image just provides 2-D coordinates. We apply two techniques to add the third dimension: radargrammetry or direct geolocation. Radargrammetry makes use of the parallax between detections of the same object in at least two SAR images acquired under different viewing angles from adjacent orbit tracks. The 3-D coordinates of the pole result from least square adjustment as intersection point of the lines of sight. In contrast to radargrammetry, direct geolocation already works with just one viewing angle but requires a Digital Elevation Model (DEM) as additional auxiliary input data. The 3-D coordinates of the pole result as intersection point of the line of sight with the DEM. To attain the aspired accuracy level for the resulting 3-D coordinates, a high-precision Laser DEM is utilized.

For the detection of poles in optical imagery, the expected direction of their shadows is calculated from the known geolocation and the acquisition time of the image. Normally, it is only possible to detect the shadows of the poles but not the poles themselves. Narrow dark objects get filtered from the image using morphological filtering. Thereafter, shadow lines are identified using a directional Gabor filter, spatial clustering, and a threshold on the size of lines.

The base points of the poles are derived as endpoints of the remaining lines in direction towards the sun. Thereafter, correspondences between base points from the SAR image and base points from the optical image are found with an iterative procedure and the mean spatial shift for the correction of the image is estimated.

The improved localization of the optical images leads to an increased value of the image content, as semantic information, e.g. derived with machine learning, is connected with more precise absolute coordinates. For instance, this is crucial for mapping tasks.

The research work was partly founded by the Bavarian State Ministry of Commerce (StMWi) within the framework of the Drive-Nav project.