Selection of sugarcane clones via multivariate models using near-infrared (NIR) spectroscopy data

Abstract

Sugarcane production plays a fundamental role in the Brazilian economy, both for sugar production and renewable energy generation. The development of new cultivars to meet the current needs of the sugarcane industry sector requires efficient phenotyping methods, which should incorporate simplification, speed, accuracy, and consistency. In order to contribute to the development of new phenotyping strategies, this work aimed to develop multivariate regression models using Partial Least Squares (PLS) to classify sugarcane clones based on sugarcane biomass quality parameters, namely fiber (FIB) and apparent sucrose (SC) content. A NIR instrument was used to acquire the reflectance spectra of 196 sugarcane bagasse - collected in two different harvest seasons - and fresh stalk samples. The values predicted by these models allowed the construction of a vector using a confusion matrix that informs whether the clone should be selected or not. PLS models selected to predict each trait under study presented high accuracy and precision, besides small values of false-positive rate and good concordance indication by the Kappa statistic test. The results obtained indicate that the use of fresh stalk samples rather than bagasse samples for the prediction of SC and FIB is recommended as it delivered higher predictive power and is of a more straightforward usage. The utilization of NIR combined with multivariate techniques may help breeding programs in the classification of sugarcane clones based on biomass quality parameters.