Geo-location correction of CBERS 2B imagery using fast template matching on a GPU

CBERS 2B satellite imagery was acquired in the period 2007-2010, at 2.5 m (panchromatic) and 20 m (multi-spectral) spatial resolution, respectively. Much of the archived data is available, free of charge, via the Brazilian space organization INPE. However, to be suitable for wide area mapping at scales 1:50,000 to 1:100,000 scales (or better), errors in geo-location, which range from several 100s of meters to several kilometers, must be corrected for each individual image frame. We demonstrate a novel method which deploys very fast template matching using GPU accelerated convolution. The method can be used with template samples from a globally consistent reference set. We demonstrate this with the use of the Terracolor LANDSAT product over Brazil and with Bing maps extracts for some urban areas in panchromatic scenes. The main advantage of the method is the possibility to use a large number of templates of a significant size (256 by 256, or 512 by 512 pixels). Processing time remains below 1 minutes per frame, which includes template selection via a WMS service. The result of the match is supplied as a list of ground control points, which we store in the VRT format, allowing the use of GDAL for transformation. We discuss implementation details and show how we can apply the method to other sensor combinations as well.