Skip to main content
ARS Home » Research » Publications at this Location » Publication #165924
Title: DETECTING ERRORS IN COTTON GRADING DATA USING BACK-PROPAGATION NEURAL NETWORKS

Author
item BOGGESS, J - MISSISSIPPI STATE UNIV
item XINTONG, BI - MISSISSIPPI STATE UNIV
item Sassenrath, Gretchen

Submitted to: Meeting Proceedings
Publication Type: Book / Chapter
Publication Acceptance Date: 8/3/2004
Publication Date: 11/7/2004
Citation: Boggess, J.E., Xintong, B., Sassenrath, G.F. 2004. Detecting errors in cotton grading data using back-propagation neural networks. In: Dagli, C. H., Buxazk, A.L., Enke, D.L., Embrechts, M.J., Ersoy, O., editors. Smart Engineering System Design: Neural Networks, Fuzzy Logic, Evolutionary Programming, Complex Systems and Artificial Life. Intelligent Engineering Systems Through Artificial Neural Networks, Volume 14. New York, NY: ASME Press. p. 621-626.

Interpretive Summary: When determining prices for cotton, the discount or premium values of the fiber are determined by a process of calculation. This calculation is based on the values of a number of features of cotton, such as the color and staple characteristics of, and the leaf fragments contained within, a particular cotton sample. However, because of different harvesting methods, ginning techniques and typographical errors, the recorded numerical values of cotton features can be erroneous. Research approaches often employ methods of harvesting and ginning that vary from that used by producers, resulting in different measures of cotton properties. If we consider the producer results to be the desired values, then the researcher values are incorrect due to differences in harvesting and ginning methods. This study explores the use of a neural network to detect these errors in cotton data. Phase 1 of this project, which is reported in this paper, is designed to detect erroneous entries in files containing cotton fiber data. When an erroneous entry is detected, the program reports the location of the error so that an expert in cotton quality measurement can examine that entry in detail to verify the correctness of its features. If systematic data entries are detected it may be possible to deduce their cause and to correct the conditions that cause the errors. In this way, a grading of cotton quality more closely aligned with producer results may be achieved.

Technical Abstract: Back-propagation neural networks are powerful tools that have been applied to a variety of real-world problems, such as classification and categorization. Error detection can be regarded as classification and thus is also one of its applications. This project involves using back-propagation neural networks to detect erroneous entries in cotton fiber grading data from the USDA-ARS at Stoneville, Mississippi. The best results obtained yielded 98.89% accuracy on training, and about 94% accuracy in predicting erroneous entries in the testing data set.