Skip to main content
ARS Home » Research » Publications at this Location » Publication #229017
Title: Novel Search Algorithims for a Mid-Infrared Spectral Libary of Cotton Contaminants

Author
item LOUDERMILK, J - UGA
item Himmelsbach, David
item Barton Ii, Franklin
item HASETH, JAMES - UGA

Submitted to: Applied Spectroscopy
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 2/26/2008
Publication Date: 6/15/2008
Citation: Loudermilk, J.B., Himmelsbach, D.S., Barton II, F.E., de Haseth, J.A. 2008. Novel Search Algorithims for a Mid-Infrared Spectral Libary of Cotton Contaminants. Applied Spectroscopy. 62: 661-670 (2008)

Interpretive Summary: Contamination of cotton yarns has a great effect on the quality and value of the yard and thus fabrics produced from then. Recently, an infrared library database has been developed identify these contaminants. However, the infrared spectra of cotton plant components that become contaminants in cotton are still difficult to distinguish from each other due to their similarity yet environmental variability. Thus, the individual standard spectral library search routines often have difficulty in providing the proper identification of these components. It was found that no single search routine is suitable for all of the types of these contaminants. In this work four categories of plant based contaminants were considered: leaf, stem, seed coat, and hull. The spectral searches to identify them were run against 6 types of search routines. The results from the all 6 routines were pooled in an automated voting scheme to provide the overall best identification. This provided a more reliable method of spectral database searching and gave better results than using any single routine. It also avoids the errors of human selection. Using this group approach makes this type of database more useful and should have application to other spectral libraries.

Technical Abstract: During harvest, a variety of plant based contaminants are collected along with cotton lint. The USDA previously created a mid-infrared, attenuated total reflection (ATR), Fourier transform infrared (FT-IR) spectral library of cotton contaminants for contaminant identification as the contaminants have negative impacts on yarn quality. This library has shown impressive identification rates for extremely similar cellulose based contaminants in cases where the library was representative of the samples searched. When spectra of contaminant samples from crops grown in different geographic locations, seasons, and conditions and measured with a different spectrometer and accessories were searched, identification rates for standard search algorithms decreased significantly. Six standard algorithms were examined: dot product, correlation, sum of absolute values of differences, sum of the square root of the absolute values of differences, sum of absolute values of differences of derivatives, and sum of squared differences of derivatives. Four categories of contaminants derived from cotton plants were considered: leaf, stem, seed coat, and hull. Experiments revealed that the performance of the standard search algorithms depended upon the category of sample being searched and that different algorithms provided complementary information about sample identity. These results indicated that choosing a single standard algorithm to search the library was not possible. Three voting scheme algorithms based on result frequency, result rank, category frequency, or a combination of these factors for the results returned by the standard algorithms were developed and tested for their capability to overcome the unpredictability of the standard algorithms' performances. The group voting scheme search was based on the number of spectra from each category of samples represented in the library returned in the top ten results of the standard algorithms. This group algorithm was able to identify correctly as many test spectra as the best standard algorithm without relying on human choice to select a standard algorithm to perform the searches.