PERFORMANCE OF HYPERSPECTRAL IMAGING ALGORITHMS USING ITANIUM ARCHITECTURE Wilfredo Lugo-Beauchamp 1 , Kennie Cruz 2 , Carmen L. Carvajal-Jiménez 2 , and Wilson Rivera 2 1 Software Solutions Group Hewlett Packard Technology Center Aguadilla, Puerto Rico, USA Wilfredo.Lugo@hp.com 2 Electrical and Computer Engineering Department University of Puerto Rico, Mayagüez Campus P.O.Box 9042, Mayaguez, Puerto Rico, USA {Kennie.Cruz, Carmen.Carjaval, Wilson.Rivera}@ece.uprm.edu Abstract This paper describes the experiences and results on implementing a set of hyperspectral imaging analysis algorithms on the Itanium Processor Family. On Itanium architecture all instructions are transformed into bundles of instructions and these bundles are processed in a parallel fashion by the different functional units. Experimental results show that exploiting implicit parallelism and linking HP Mathematical LIBrary optimized for Itanium yield significant improvement in performance. Keywords: Itanium, IA64, Remote Sensing, Hyperspectral Imaging, Image Classifiers, HP-MLIB 1. Introduction Sensors based on imaging spectrometry or so called hyperspectral imagers collect high spectral resolution data over a couple of hundred of wavelengths effectively producing an image where at each pixel we get the spectral response of the object(s) in the field of view of the sensor. Hyperspectral Imaging (HSI) analysis is based on the concept of imaging spectrometry where spectral and spatial information is used to identify or detect objects, or estimate parameters of interest. As the object of interest is embedded in a complex media (i.e. coastal waters or skin), the measured signature is a distorted version of the original object signature (e.g. a coral reef or a blood vessel) mixed with clutter. By large hyperspectral imaging analysis concentrates on dimensionality reduction and classification algorithms. Dimensionality reduction algorithms reduce the data volume (dimensionality), without loss of critical information, so that it can be processed efficiently. Classification of a hyperspectral image sequence, in turn, identifies which pixels contain various spectrally distinct materials. Different classification metrics have been proposed from minimum distance, such as Euclidean, Fisher Linear Discriminant, and Malahanobis, to maximum likelihood [1] to correlation matched filter-based approaches such as spectral signature matching [2]. There are two major techniques to image classification: supervised and unsupervised. In supervised classification techniques, an analyst develops quantitative descriptions of the spectral characteristics of the various classes of interest for a particular scene. These descriptions are then used as reference spectral signatures against which every pixel in an image is compared. The pixels are classified according to the spectral signature they most closely resemble. In unsupervised classification, the algorithms do not use training data as the basis for classification. Instead, the algorithms used examine the unknown pixels in the image and aggregate them into various classes according to the clusters found in the spectral space that contains the image. We have implemented a set of hyperpectral imaging analysis algorithms. These algorithms have been tuned and ported to be run on the Itanium Processor Family. We describe in this paper the experiences and results on implementing these algorithms. The structure of this paper is as follows. Section 2 provides an overview of hyperspectral imaging algorithms. Section 3 discusses implementation issues. Section 4 presents the results and discusses related work. Finally, section 5 draws conclusions and potential future work. 2. Background A complete description of hyperspectral algorithms can be found elsewhere in [3]. In the next subsections we describe briefly each of the algorithms we implemented on Itanium. 449-199 327