Leaf-proximal Hyperspectral Data and Multivariate Modelling Approaches to Estimate Phosphorus and Potassium Content of Wheat Leaves

The assessment of plant nutrient status to provide sufficient fertilization for rapid and
continuous uptake by plants has been based on visual diagnosis in the field, which is quick but
demands a lot of experience and has a low operability. Visible near-infrared spectroscopy
(VNIS) has shown to be a quick, non-destructive, accurate, and cost-effective analytical
method in precision agriculture. In this study, we assessed the potential of this technology to
predict phosphorus and potassium content in the wheat leaves using different multivariate
regression methods. The hyperspectral and the reference measurements were taken from wheat
plant leaves grown in a long-term fertilization trial under contrasted concentrations of
phosphorus and potassium. The leaf proximal and hyperspectral data were collected using an
ASD FieldSpec4 spectroradiometer operating in the spectral range from 350 to 2500 nm.
Before conducting the analysis, the leaves spectra were preprocessed with a Savitzky–Golay
smooth filter and a Standard Normal Variate normalization method. A total of 60 samples,
collected between flowering and maturity stages, combined with the preprocessed spectra were
used to develop support vector regression (SVR), random forest (RF), and principal component
regression (PCR) prediction models for estimating leaves phosphorus content (LPC) and leaves
potassium content (LKC). The entire sample set was randomly split into a training set (70%)
and a test set (30%), and the performances of the different prediction models were compared
using normalized root mean square error (NRMSE) and coefficient of determination (R2) in
both cross-validation and testing processes. The results showed that LPC prediction models
outperformed the LKC models, with high accuracies (R2
) in cross-validation in the order of
0.84, 0.85, and 0.79 for SVR, PCR, and RF, respectively. For potassium, the coefficient of
determination of cross-validation was 0.64, 0.59, and 0.54 for SVR, PCR, and RF, respectively.
The highest validation results were returned by the RF algorithm for both LPC and LKC
predictions, with moderate R2 values equal to 0.56 and 0.53, respectively. In the RF model,
phosphorus and potassium in wheat leaves can be predicted with errors of 19 and 13%,
respectively.