New approaches for field analyses of cotton quality by means of near infrared spectroscopy supported by chemometrics

Authors: VOGT, FRANK - University Of Tennessee; LUTTRELL, ROBERT - Salisbury University; Rodgers Iii, James

Submitted to: Analytical Letters
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 11/1/2010
Publication Date: 10/6/2011
Citation: Vogt, F., Luttrell, R.D., Rodgers III, J.E. 2011. New approaches for field analyses of cotton quality by means of near infrared spectroscopy supported by chemometrics. Analytical Letters. 44(15):2466-2477.

Interpretive Summary: An important quality measure of cotton is “micronaire”, a physical parameter that is indicative of the fiber maturity and its fineness. Maturity and fineness are two important properties that determine the resulting products’ quality and hence there is a strong economic interest in routine measurements of cotton quality. Currently, micronaire can only be measured in a few laboratories with equipment that prohibits routine field analyses. Goal of this study is a proof-of-principle that cotton quality is also correlated to properties that can be determined by means of near infrared (NIR) reflection spectroscopy. Principal Component Regression (PCR) and Partial Least-Squares (PLS) have been utilized for spectrometer calibration and micronaire prediction. The long-term goal is to develop a spectroscopic, field-capable technique for measurement of micronaire values as a complementary technique to the standard laboratory based analyses. In this application, overfitting of the regression was found to have a strong impact, requiring automated methods for selecting the number of components of the chemometric calibration model. Two approaches have been utilized: standard crossvalidation and a novel algorithm that adapts the calibration model to each unknown spectrum individually. The latter method has been designed such that the best model is derived that can model the spectroscopic features and at the same time minimizes overfitting. In conjunction with input from cotton producer and cotton processing industries, it was defined that such a spectroscopic measurement is a complementary and promising alternative to the standard laboratory measurements, if at least 70% of the micronaire values obtained spectroscopically fall inside ±0.3 micronaire units (m.u.) of the laboratory based technique. A set of 191 cotton samples was acquired from three different upland cotton varieties. This set was split into a calibration and an independent test set. The above requirement was met for >90% of the test samples, and thus a promising new approach for field analyses is on the horizon. Further, the acceptance range could be reduced to ±0.2 m.u. and still =70% of the samples would be inside the restricted acceptance range. * corresponding author; 1University of Tennessee, Knoxville, TN; 2Salisbury University, Salisbury, MD; 3Southern Regional Research Center (SRRC)-Agricultural Research Service (ARS)-United States Department of Agriculture (USDA), New Orleans, LA

Technical Abstract: An important quality measure of cotton is “micronaire”, a physical parameter that is indicative of the fiber maturity and its fineness. Maturity and fineness are two important properties that determine the resulting products’ quality and hence there is a strong economic interest in routine measurements of cotton quality. Currently, micronaire can only be measured in a few laboratories with equipment that prohibits routine field analyses. Goal of this study is a proof-of-principle that cotton quality is also correlated to properties that can be determined by means of near infrared (NIR) reflection spectroscopy. Principal Component Regression (PCR) and Partial Least-Squares (PLS) have been utilized for spectrometer calibration and micronaire prediction. The long-term goal is to develop a spectroscopic, field-capable technique for measurement of micronaire values as a complementary technique to the standard laboratory based analyses. Results obtained in this study also demonstrate that the quality of micronaire predictions is strongly dependent on the dimension of the utilized PCR or PLS calibration model. In this application, overfitting was found to have a strong impact, requiring automated methods for selecting the dimension of the chemometric calibration model. Two approaches have been utilized: standard crossvalidation and a novel algorithm that adapts the calibration model to each unknown spectrum individually. The latter method has been designed such that the most parsimonious model is derived that can model the spectroscopic features and at the same time minimizes overfitting. In conjunction with input from cotton producer and cotton processing industries, it was defined that such a spectroscopic measurement is a complementary and promising alternative to the standard laboratory measurements, if at least 70% of the micronaire values obtained spectroscopically fall inside ±0.3 micronaire units (m.u.) of the laboratory based technique. A set of 191 cotton samples was acquired from three different upland cotton varieties. This set was split into a calibration and an independent test set. The above requirement was met for >90% of the test samples, and thus a promising new approach for field analyses is on the horizon. Further, the acceptance range could be reduced to ±0.2 m.u. and still =70% of the samples would be inside the restricted acceptance range. Up to 60% of the samples fell inside an acceptance range of ±0.1 m.u. * corresponding author; 1University of Tennessee, Knoxville, TN; 2Salisbury University, Salisbury, MD; 3Southern Regional Research Center (SRRC)-Agricultural Research Service (ARS)-United States Department of Agriculture (USDA), New Orleans, LA