Predictive modelling of chlorophyll in Mombaça grass leaves by hyperspectral reflectance data and machine learning

Abstract

Chlorophyll (Chl) concentration is one of the factors that affects crop productivity. This study investigated the prediction of chlorophyll concentrations in Mombaça grass' leaves using hyperspectral data and machine learning techniques. Chlorophyll variations were induced by different levels of nitrogen fertilization (104, 208, 312, and 416 kg ha−1). Spectral signatures (400–2500 nm) and chlorophyll contents of the leaves were obtained in October, November, and December 2017, and January 2018. Models were generated using Partial Least Square Regression (PLSR), Random Forest (RF), and Support Vector Regression (SVR). Two validation techniques were employed: holdout, dividing the data into training (75%) and testing (25%) sets; and leave‐one‐date‐out cross‐validation (LOOCV), in which one date was omitted during model training and used to predict the omitted date's value. Chlorophyll concentrations varied according to N doses, with the highest concentrations observed in October and December. In these months, there were greater variations in spectral reflectance in the green and red bands (530–680 nm). December was identified as the ideal period for chlorophyll quantification, for both holdout and LOOCV validation techniques. The SVR technique performed best (R2 = 0.71, RMSE = 0.23 mg g−1, dr = 0.72) compared to RF (R2 = 0.63, RMSE = 0.27 mg g−1, dr = 0.66) and PLSR (R2 = 0.60, RMSE = 0.27 mg g−1, dr = 0.67). Therefore, the prediction of chlorophyll in Mombaça grass using spectroradiometry is promising and applicable across different cultivation periods.