New approach for predicting nitrogen and pigments in maize from hyperspectral data and machine learning models

Abstract

Fast diagnostics from hyperspectral data and machine learning (ML) models to predict nitrogen (N) and pigment content in maize crops is challenging to optimize nitrogen fertilization. This research assessed the efficiency of the five ML algorithms, the best phenological stage, and the sensitivity of the 90 spectra to estimate N and pigment content. Therefore, this field research proposes as a novelty to test which of the five ML algorithms accurately estimates nitrogen, chlorophyll, and carotenoid content in maize leaves at different phenological stages using hyperspectral band data. The treatments were arranged in a factorial scheme with four N doses (0, 54, 108, and 216 kg ha−1) combined with five leaf collection seasons at phenological stages V6 to V14. The ML models tested were artificial neural networks – ANN, decision tree adapted for prediction problems – M5P, REPTree decision tree, random forest - RF, polynomial support vector machine – PSVM, and ZeroR - ZR (control). Spectral bands 530–560 nm and 690–750 nm are effective wavelengths because the visible region with lower reflectance (530–560 nm) affects N uptake and chlorophyll and carotenoid content, while the red-edge and near-infrared region affects N and chlorophyll content. The random forest (RF) model performed better with higher correlation (r) and mean absolute error (MAE) between predicted and observed values for all variables, with the correlation coefficient (r) value being around 0.6 and the MAE below 0.5 for the prediction of chlorophyll a+b. For the prediction of flavonoids, the r was around 0.6 and the error was 0.07. Support vector machine (SVM) and RF efficiently predicted nitrogen content, in predicting of NF, the r values for both algorithms were above 0.35 and the error was below 2.75.