Computational challenges in processing large hyperspectral images

The processing of large hyperspectral images presents challenges from both memory usage and computation points of view. Large images require proper partitioning in order to be stored in memory and to exploit the benefits of parallel implementation on high performance computing architectures. This paper analyzes several variants of reading and distributing large images and presents critical issues and some corresponding solutions in designing efficient parallel implementations of two clustering algorithms which use both spectral and spatial information. All experiments were conducted on a BlueGene/P supercomputer using up to 1024 processors.