Distributed Computing Approach for Remote Sensing Data
Gregg M. Petrie, C. Dippold, George I. Fann, Donald R. Jones, Elizabeth R. Jurrus, Brian D. Moon · 2003
Processing image data generated by new remote sensing systems can severely tax the computational limits of the classic single processor systems that are normally available to the remote sensing practitioner. Operating on these large data sets with a single computer system sometimes means that simplifying approximations are used that can limit the precision of the final results. For instance, in supervised classification it is often necessary to assume a Gaussian structure for the data. While this assumption has the advantage of greatly reducing the amount of pixels that must be processed this abstraction can also mask important structures in the raw data. Recent work at Pacific Northwest National Laboratory strongly suggests that a distributed network of inexpensive PCs can be designed that is optimal to deal with the type of computationally intensive problems encountered in processing remotely sensed images. Under the assumption that this new type of distributed computing will remove computational constraints, new image processing algorithms for remote sensed images are now being considered. A specific example will be presented, where consideration of the entire training set, instead of abstracting of the training set to a few representative parameters, can significantly improve classification algorithms.