Combining mid-infrared spectroscopy and machine learning to estimate nutrient content in plant tissues of yerba mate (Ilex paraguariensis A. St. Hil.)

Abstract

Mid-infrared diffuse reflectance spectroscopy with machine learning methods has been studied worldwide for estimating nutrient content in plant tissue of numerous plant species. Therefore, this study aimed to evaluate the best combination of spectral preprocessing and machine learning methods for estimating the concentration of chemical elements in leaf tissues of Ilex paraguariensis. From 2019 to 2021, 111 samples were collected from farms throughout Rio Grande do Sul State in southern Brazil. The total concentration of 11 nutrients (N, P, K, Ca, Mg, B, S, Mn, Fe, Cu, and Zn) and the spectra recorded in the mid-infrared range (2500– 18,000 nm) were obtained. The original spectrum underwent detrending, standard normal variate, and Savitzky-Golay derivative (SGD) spectral preprocessing. Multivariate partial least squares regression and support vector machine models were used to calibrate the spectra based on the nutrient concentrations obtained from the conventional analytical method. The models’ performance was evaluated using the coefficient of determination, the ratio of performance to the interquartile range, and the root mean square error of prediction. The support vector machine model combined with SGD preprocessing resulted in a more accurate prediction of the 11 elements studied in leaf tissue samples.