Imaging spectroscopy, also known as hyperspectral remote sensing, is concerned with the measurement, analysis, and interpretation of spectra acquired from a given scene (or specific object) at a short, medium or long distance by an airborne or satellite sensor. Analysis in a timely manner of the acquired multi-dimensional images allows to develop applications with high social impact, such as urban growing monitoring, crop fields identification, target detection for military and defense/security deployment, wildland fire detection and monitoring, biological threat detection, biophysical parameter estimation, or monitoring of oil spills and other types of chemical contamination.
In this context, support vector machines (SVM) [1, 2, 3] have become one of the state-of-the-art machine learning tools for hyperspectral image classification. However, its high computational cost for large scale applications makes the use of SVM limited to off-line processing scenarios. Certainly, with the recent explosion in the amount and complexity of hyperspectral data, parallel processing has soon become a requirement in many remote sensing missions, especially with the advent of low-cost systems such as commodity clusters and distributed networks of computers In order to address this relevant issue, this chapter explores the development of two parallel versions of SVMs for remote sensing image classification.
Sequential minimal optimization is a very popular algorithm for training SVMs, but it still requires a large amount of computation time for solving large size problems. In this work, we evaluate the performance of a parallel implementation of the SVM based on the parallelization of the incomplete Cholesky factorization and present novel parallel implementations that balance the load across the available processors through standard Master-Worker decompositions. Both methodologies are theoretically analyzed in terms of scalability, computational efficiency and time response. The impact of the multi-class scheme is also analyzed. Results on real multispectral and hyperspectral datasets illustrate the performance of the methods. We finally discuss the possibility of obtaining processing results quickly enough for practical use via the Marenostrum supercomputer available at the Barcelona Supercomputing Center in Spain, and other massively parallel facilities at NASA's Goddard Space Flight Center in Maryland.