|Hana, Maha - NC STATE UNIVERSITY|
|Mcclure, W - NC STATE UNIVERSITY|
Submitted to: Korean Society of Near Infrared Spectroscopy
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: October 3, 2000
Publication Date: May 15, 2001
Citation: HANA, M., MCCLURE, W.F., WHITAKER, T.B. TUNING BACKPROPAGATION NETWORKS FOR ANALYZING NIR DATA. KOREAN SOCIETY OF NEAR INFRARED SPECTROSCOPY. 2001. v. 2. p. 15-23. Interpretive Summary: It is expensive and time-consuming to use analytical chemistry methods to measure quality attributes (such as sugars and proteins) of agricultural commodities. Near infrared (NIR) technology, a method that measures either reflected or transmitted energy over a range of wavelengths, has an advantage over wet chemistry because it is a rapid, non-destructive, and low-cost method of measuring quality attributes. However, calibration equations have to be developed that correlate NIR energy at specific wavelengths to the concentration of a specific quality attribute. Calibration equations can be difficult to develop and can be imprecise. A new objective method, using artificial neural networks, was developed that gave more precise estimates of quality attributes than conventional methods.
Technical Abstract: Designing and training back-propagation (BP) neural networks for analyzing NIR data can be an arduous and time-consuming task. A BP network may be trained by randomly dividing the data set into two parts, training the network with one part and checking its performance with the other part. However, this procedure is plagued with a lack of objective information about network characteristics such as the required number of nodes in the hidden layers and the number of epochs needed to train for optimal performance. In this study, an objective method was developed to optimize a BP network tuning procedure. The tuning procedure involved randomly dividing a data set into five groups. Each of the five groups was randomly subdivided into two groups with 80% in a training set and 20% in a tuning set. Training was interrupted periodically after every 1000 epochs. During each interruption, network performance was checked against the tuning set and recording the mean-square error (MSE) and the number of epochs. This procedure was continued until the MSE was minimized. The number of epochs required to reach a minimum MSE was also recorded. Once tuned, the performance of the optimized network was determined by testing the network with all available data. Four different NIR data sets were used in this study to compare the objective method of optimizing the network to conventional methods of analyzing NIR data. Two data sets were used to determine the concentration of nicotine in tobacco samples. The remaining two data sets were used to determine the type (Burley or flue-cured) tobacco and the country of origin. Results indicate improved performance using MSE in the tuning procedure when compared to conventional methods.