HR: 1340h
AN: NG43A-0445    [Abstracts]
TI: A Grid Service for Automatic Land Cover Classification Using Hyperspectral Images
AU: * Jasso, H
EM: hjasso@sdsc.edu
AF: San Diego Supercomputer Center, University of California San Diego, 9500 Gilman Drive, MC 0505, La Jolla, CA 92093-0505 United States
AU: Shin, P
EM: kg@sdsc.edu
AF: San Diego Supercomputer Center, University of California San Diego, 9500 Gilman Drive, MC 0505, La Jolla, CA 92093-0505 United States
AU: Fountain, T
EM: fountain@sdsc.edu
AF: San Diego Supercomputer Center, University of California San Diego, 9500 Gilman Drive, MC 0505, La Jolla, CA 92093-0505 United States
AU: Pennington, D
EM: penningd@lternet.edu
AF: LTER Network Office, University of New Mexico Dept. of Biology, MSC03 2020, Albuquerque, NM 87131-0001 United States
AU: Ding, L
EM: ljding@sdsc.edu
AF: San Diego Supercomputer Center, University of California San Diego, 9500 Gilman Drive, MC 0505, La Jolla, CA 92093-0505 United States
AU: Cotofana, N
EM: neil@sdsc.edu
AF: San Diego Supercomputer Center, University of California San Diego, 9500 Gilman Drive, MC 0505, La Jolla, CA 92093-0505 United States
AB: Hyperspectral images are collected using Airborne Visible/Infrared Imaging Spectrometer (Aviris) optical sensors [1]. 224 contiguous channels are measured across the spectral range, from 400 to 2500 nanometers. We present a system for the automatic classification of land cover using hyperspectral images, and propose an architecture for deploying the system in a grid environment that harnesses distributed file storage and CPU resources for the task. Originally, we ran the following data mining algorithms on a 300x300 image of a section of the Sevilleta National Wildlife Refuge in New Mexico [2]: Maximum Likelihood, Naive Bayes Classifier, Minimum Distance, and Support Vector Machine (SVM). For this, ground truth for 673 pixels was manually collected according to eight possible land covers: river, riparian, agriculture, arid upland, semi-arid upland, barren, pavement, or clouds. The classification accuracies for these algorithms were of 96.4%, 90.9%, 88.4%, and 77.6%, respectively [3]. In this study, we noticed that the slope between adjacent frequencies produces specific patterns across the whole spectrum, giving a good indication of the pixel's land cover type. Wavelet analysis makes these global patterns explicit, by breaking down the signal into variable-sized windows, where long time windows capture low-frequency information and short time windows capture high-frequency information. High frequency information translates to information among close neighbors while low frequency information displays the overall trend of the features. We pre-processed the data using different families of wavelets, resulting in an increase in the performance of the Naive Bayesian Classifier and SVM to 94.2% and 90.1%, respectively. Classification accuracy with SVM was further increased to 97.1 % by modifying the mechanism by which multi-class is achieved using basic two-class SVMs. The original winner-take-all SVM scheme was replaced with a one-against-one scheme, in which k(k-1) binary classes are trained for a k class problem . To deploy the data mining system as a grid service, we are using the Globus Toolkit to build a distributed environment over a computational grid. It is consisted of two major components: a) a service-oriented infrastructure b) a set of client tools to communicate with the service-oriented infrastructure, i.e., a web service based on the Kepler system, a visual modeling system for designing and executing scientific workflows that access distributed data and tools, and uses a semantic-mediation engine to integrate those resources [4] and the SKIDLkit data mining toolkit for high-dimensional data mining [5] to train and test new classification models. References [1] http://aviris.jpl.nasa.gov [2] http://sev.lternet.edu [3] Pennington, D., H. Jasso, P. Shin, & T. Fountain. The effect of landscape heterogeneity on classification accuracy: a comparison of classifier prediction in sub-opotimal sampling conditions. Seventh Workshop on Mining Scientific and Engineering Datasets, 2004 SIAM International Conference on Data Mining (SDM 2004), Orlando, Florida, 2004. [4] I. Altintas, C. Berkley, E. Jaeger, M. Jones, B. Lud„scher, S. Mock. Kepler: Towards a Grid-Enabled System for Scientific Workflows, In the Workflow in Grid Systems Workshop in GGF10 - The Tenth Global Grid Forum, Berlin, Germany, March 2004. [5] http://daks.sdsc.edu/skidl/skidldownloads.html
DE: 1640 Remote sensing
DE: 0933 Remote sensing
DE: 0994 Instruments and techniques
DE: 0999 General or miscellaneous
SC: Nonlinear Geophysics [NG]
MN: 2004 AGU Fall Meeting