Espectroscopia NIR portátil e aprendizado de máquina: desenvolvimento de modelos para a quantificação de carboidratos não estruturais em macieira (Malus domestica)

Abstract

The quantification of non-structural carbohydrates (NSC) in perennial tissues of apple trees (Malus domestica) is essential for the physiological monitoring of the crop; however, reference methods such as high-performance liquid chromatography (HPLC) and enzymatic assays are time-consuming, destructive, and waste-generating. This work aims to develop a rapid analytical method using portable near-infrared (NIR) spectroscopy combined with chemometric and machine learning tools. The focus was the simultaneous quantification of seven NSC: starch, sucrose, glucose, fructose, sorbitol, total hexoses, and raffinose, in three perennial tissues (wood, bark, and apical bud). For this purpose, diffuse reflectance spectra of the freeze-dried and ground samples, in the 900 to 1700 nm region, were collected in triplicate, totaling 141 samples (423 spectra), at different stages of the plant's physiological development. Subsequently, calibration and validation models were built using reference values obtained by HPLC coupled with Refractive Index detection (HPLC-RI) for soluble sugars/sorbitol, and by the 3,5-dinitrosalicylic acid (DNS) enzymatic-colorimetric method for starch. The construction of the multivariate calibration models was carried out using linear and machine learning methods: Partial Least Squares regression (PLS), Support Vector Machines (SVM), Artificial Neural Networks (ANN), and Random Forest (RF), with the application and comparison of different spectral preprocessing techniques, including: Multiplicative Scatter Correction (MSC), Standard Normal Variate (SNV), and first and second derivatives. The models were evaluated by cross-validation and external validation, based on the values of the Root Mean Square Error of Calibration (RMSEC), Root Mean Square Error of Prediction (RMSEP), coefficient of determination (R²), and Ratio of Performance to Deviation (RPD). The best models were obtained for fructose (MSC + PLS; R²val = 0.83; RPD = 4.84 and SVM + MSC; R²val = 0.808; RPD = 4.747), with satisfactory results also obtained for hexoses (SG2 + SVM; R²val = 0.669; RPD = 1.693) and starch (SG1 + PLS; R²val = 0.782; RPD = 2.053), demonstrating the potential of portable NIR spectroscopy combined with machine learning for the screening of NSC in apple tissues.