Grid computing for hyperspectral data processing

We investigate the use of a flexible grid architecture for hyperspectral image processing. Recording data in tens or hundreds of narrow contiguous spectral intervals, hyperspectral data outperform multispectral imagery by allowing the detection of relatively small differences in material composition and of targets occupying a surface smaller than the one covered by a pixel (called subpixel targets). However, with increased spatial and spectral resolution, processing such data often leads to computational costs prohibitive to regular computer systems. While distributed or parallel computing are often found as solutions, many current configurations are still unable to reach the computational complexity level that is required for exhaustive search solutions. In this environment, grid computing becomes a viable alternative. Grid computing, an emerging computing model, is based on the concept of distributing processes across a parallel infrastructure. Throughput is further increased by networking many heterogeneous resources across administrative boundaries to model a virtual computer architecture. Compared to distributed clusters or parallel machines, grid systems are often inexpensive or even free since they can consist of non-dedicated computer systems that are underutilized and have extra CPU cycles that can be spared. We present general considerations on grid architectures and discuss the current grid environment we have deployed. Next, we investigate exhaustive band search, a data processing problems that suffers from large computational requirements and present our grid based solutions for it. Our experimental results indicate a significant speedup in obtaining results and even solving of problems otherwise not tractable in regular computing environments.