Resumen de High performance computing for geocorrection of hyperspectral imagery
In this dissertation we present a method for real-time geocorrection and an evolutionary method for Ground Control Point-based nonlinear registration of images from airborne pushbroom sensors, using the hardware acceleration and parallel computing characteristics of modern Graphics Processing Units.
The geocorrection method is based on a projective texture technique originally developed for fast shadow rendering. By combining the image data with inertial navigation and positioning ancillary data, we correct the image to a Digital Terrain Model (DTM) and produce a geocorrected and georeferenced image. The method works independently of the number of channels of the sensor. Results with an ultralight hyperspectral system show that the speed achieved with standard hardware for a 1-m grid DTM gives better than real-time performance, with processing speed achieved between 356 and 511 lines per second, x11 to x17 acquisition time. This allows in-flight inspection of the geocorrected image during acquisition, for more efficient coverage of large target areas. In regard to accuracy, differences with standard ray tracing direct geocorrection remained subpixel. GPU computing allows very fast geocorrection without accuracy loss with respect to traditional direct methods, with very little computational load for the central processor.
The registration method is an evolutionary method based on an implementation of Particle Swarm Optimization (PSO). Using the geocorrection method by the fitness function during the optimization process, the speed achieved allows using evolutionary methods in feasible time, enabling hundreds of repeated approximations during rectification, in contrast to classical geocorrection methods. In our approach, taking advantage of the speed and parallelization of Graphic Processing Units (GPU) by means of CUDA, PSO is used to find the best match between the projected pixels and a number of Ground Control Points, compensating any systematic errors in the navigation data used for the generation of the orthoimage.
With the combination of these two methods, we achieve a method to orthorectify image data and compensate any systematic errors in the navigation data used for the generation of the orthoimage in a finite time. This result contributes to an optimized operational cost of orthoimages generation processes.
