Predicting medicinal phytochemicals of Moringa oleifera using hyperspectral reflectance of tree canopies

Remote sensing has the potential to complement and perhaps replace some wet analytical chemistry measurements of some plant traits, a mechanism that can facilitate quick and nondestructive estimation. The present study reports on the use of hyperspectral data in estimating total phenolic and flavonoid concentration in Moringa oleifera, a plant with great nutraceutical properties. A completely randomized design (CBD) was used in the planting of four M. oleifera cultivars in a greenhouse setup. A hand-held spectroradiometer was used to collect canopy reflectance on two-month-old plants. This was followed by the analytical quantification of total phenolic and flavonoid concentrations in the leaf extracts. Various pre-processing techniques were carried out on the reflectance spectra such as Savitzky-Golay filter and continuum removal (convex and segmented upper hull). Analysis of variance (ANOVA) and random forest (RF) regression algorithms were then conducted to analyse the data. The ANOVA indicated that the spectral features of the cultivars differed significantly (p <= 0.01) from each other in some sections of the visible, near-infrared (NIR) and shortwave-infrared (SWIR) regions of the electromagnetic spectrum (EMS). The band depth values created from the segmented hull were used in the selection of phytochemicals absorption features. The results indicated that phytochemical concentration can be accurately estimated for individual cultivars as shown by validation models having R-2 values of between 0.58 and 0.79 and relative root mean square error (RMSErel) of between 5% and 17%. However, when data were aggregated across cultivars, the RMSErel ranged from 16% to 21%. The most important wavebands for predicting the concentration of these phytochemicals were found in the visible and SWIR regions of the EMS. Overall, accurately estimating the concentration of the phytochemicals demonstrates an insight in the potential of hyperspectral data to be extrapolated to landscape scale using hyperspectral reflectance sensing technique.