Structure and Motion Analysis of Satellite Imagery on the GPU

Scott Grauer-Gray and Chandra Kambhamettu

The accurate calculation of cloud structure and motion from satellite images is critical in applications ranging from weather predictions to understanding the structure of severe storms such as hurricanes. In this research, which was partially funded by a NASA grant, we looked at methods to improve the processing of stereo images as well as image sequences in terms of the running time and the accuracy of the desired information about the structure/motion present in the images (which given in the form of a disparity map/set of motion vectors). We focused on extracting data from sequential sequences of cloud images with an eye toward real-time analysis of satellite imagery.

In particular, we developed a belief propagation implementation that took advantage of the parallel processing power of GPUs using the CUDA architecture from nVidia in order to more quickly retrieve the output disparity map or set of motion vectors from corresponding images without losing accuracy. Our publication entitled "GPU Implementation of Belief Propagation Using CUDA for Cloud Tracking and Reconstruction" which was presented at the 5th IAPR Workshop on Pattern Recognition in Remote Sensing in 2008 was the first published implementation of belief propagation for stereo vision that took advantage of the CUDA architecture; the work also describes how we extended the implementation to estimate 2D motion vectors between sequential images. The source code used for the stereo processing portion of this publication is available below (see Software).

Fig. 1 on the left shows a stereo sequence of reference satellite images taken of Hurricane Luis on September 6, 1995. Fig. 2 on the right shows the corresponding disparity maps which were calculated using the reference images and corresponding test images.

Fig. 3 on the left shows a sequence of images of Hurricane Luis from a single satellite. Fig. 4 on the right shows the same sequence overlaid with the computed motion vectors.


Related publications


  • This work was partially supported by a U.S National Aeronautics and Space Administration award NASA NNX08AD80G under the ROSES Applied Information Systems Research program.