Using Hyperspectral Spectrometry and Functional Models to Characterize Vine-Leaf Composition

Concentrations of certain chemical elements in plants need to be controlled to ensure good crop quality and yield. However, laboratory analyses are usually time-consuming and expensive. Although indirect methods based on leaf reflectance are both faster and less expensive, most are based on indexes that only take into account certain wavelengths and fail to take full advantage of the entire reflectance curve depicted by modern hyperspectral sensors. This paper applies two functional prediction models, i.e., functional linear regression and functional nonparametric methods, to the prediction of the chemical characteristics (moisture, dry mass, and concentrations of nitrogen, phosphorus, potassium, calcium, iron, and magnesium) of vine leaves, using electromagnetic reflectance between 350 and 2500 nm as the input. Cross-validation was used to obtain optimal parameters for the models, which were tested using samples reflecting 5% and 10% of the sample size. The results obtained showed different levels of correlation between reflectance and the predicted data. The nonparametric methods yielded better results as they produced smaller prediction errors than functional regression. Moisture
$(R^{2} = 0.96)$
and nitrogen
$(R^{2} = 0.95)$
were the best predicted components, whereas magnesium content was the worst predicted component
$(R^{2} = 0.77)$
.