HR: 09:45h
AN: A11E-07 [Abstracts]
TI: Using Neural Networks for Instrument Cross-Calibration
AU: * Lary, D
EM: David.Lary@umbc.edu
AF: UMBC/GEST NASA/GSFC, University of Maryland Baltimore County
5523 Research Park Drive, Suite 320, Baltimore, MD 21228, United States
AB:
Neural networks are non-linear non-parametric learning algorithms that are universal approximators. They have
proved very useful to us in a variety of applications (Lary et al., 2004, 2007a,b), from the acceleration of expensive
code elements to learning the cross-calibration between large earth observing datasets including atmospheric
composition, aerosol optical depth, and vegetation indices. We have been using a variety of networks including
feed-forward multi-layer perceptron networks trained with the Levenberg-Marquardt algorithm, and neuro-fuzzy
networks. The success of the neural networks largely depends on two factors. First, having a training dataset that
adequately spans the parameter space. Second, including the variables that explain the variance in the dataset. If
these two criteria are met then the neural networks give excellent results as they are universal approximators. We
present several examples of neural network cross-calibration.
DE: 0300 ATMOSPHERIC COMPOSITION AND STRUCTURE
DE: 0555 Neural networks, fuzzy logic, machine learning
SC: Atmospheric Sciences [A]
MN: 2007 Fall Meeting