Predicting Cotton Fiber Properties with Fiber Length Parameters Measured by Dual-beard Fibrograph

Abstract Cotton fiber properties, although influenced tremendously by plant growing conditions, are largely dictated by cotton variety, and therefore certain associations inherently exist among the properties. Previous studies have examined cotton properties that have direct impacts on others, e.g., fiber maturity on strength, but latent associations between fiber properties, such as fiber length vs. maturity, have not been explored. This paper focuses on investigating the relationships between fiber length and other important properties (e.g., strength and maturity), and creating the regression models to predict these properties in case the testing conditions are unattainable. We firstly selected 117 cotton samples and divided them into a training set (100 samples) and a testing set (17 samples). We then measured the fiber length parameters with Dual-beard Fibrograph (DBF) that can generate accurate fiber length distributions, and the seven other fiber properties, strength, elongation, micronaire, nep, fineness, immature-fiber-content and maturity-ratio, with High Volume Instrument (HVI) and Advanced Fiber Information System (AFIS). Finally, we performed multivariate regression, Bland-Altman plot, clustering and other statistical analyses to assess the correlations, multicollinearity, agreements and clusters of the fiber properties. It was found that fiber length parameters had moderate associations (0.3<|r|<0.7) with the seven properties, and the prediction errors on the training set varied from 2.25% (for maturity-ratio) to 14.36% (for nep). The Bland-Altman analysis proved that for any of the seven properties, more than 94.9% of the predicted and actual points were within the 95% limits of agreement and without systematic biases. The regression models based on cotton clusters significantly lowered the prediction errors because the cluster centroids better represented the collective features of the fiber properties. The analyses on the testing set showed the prediction models can generate comparable results for both the training and testing sets, and demonstrated good generalization over new samples.