Improving Sugarcane Extraneous Matter Prediction Using Transformed Near-Infrared Spectral Data and Functional Regression

Near-infrared (NIR) spectra provide rich functional information for predicting sugarcane extraneous matter, but effective modeling is hindered by the functional nature of the predictors and the bounded, percentage-based response. We propose a Transformed Partial Functional Linear Model that integrates smooth functional predictors, categorical covariates, and a data-driven monotone transformation of the response, with all tuning parameters selected through a leave-group-out cross-validation scheme designed to ensure robustness to previously unseen sugarcane varieties and field conditions. The approach yields an interpretable spectral coefficient function while preserving inference on the original percentage scale via the inverse transformation. Extensive simulations demonstrate improved stability over partial least squares regression, generalized functional linear model, and modern machine learning methods, and the analysis of real sugarcane NIR data shows that explicitly accounting for both the functional structure of spectra and the bounded nature of the response leads to substantially improved predictive accuracy. Supplementary materials accompanying this paper appear online.