Phenology-based discrimination of maize (Zea mays L.) varieties using multitemporal hyperspectral data

Abstract

Discriminating maize varieties is crucial for crop monitoring, high-throughput phenotyping, and yield forecasting. Crop experts discriminate maize varieties using morphological and biophysical characteristics. However, visual classification suffers from inconsistency, low-throughput, and is only applicable in small scale. A cost-effective and accurate in situ varietal discrimination using multitemporal hyperspectral data and multivariate techniques is explored with threefold objectives, namely, to (1) discriminate maize varieties, (2) define suitable spectral bands, and (3) determine the optimum phenological stage(s) for varietal discrimination. Spectral data in 0.5-nm discrete narrowbands between 400- and 900-nm range are taken from 25 varieties measured using Apogee spectrometer at five phenological stages. Prior to discrimination analysis using partial least squares-discriminant analysis (PLS-DA), three preprocessing transformations are performed: autoscaling, Savitzky–Golay smoothing, and generalized least squares weighting. Ten optimal bands are identified for maize varietal discrimination across the visible and near-infrared section of the wavelength. The significant bands are located in the blue (400 and 455 nm), green (545 nm), red and red-edge (625, 680, 705, and 720 nm), and near-infrared (765, 840, and 895 nm) ranges of the spectrum. Flowering and onset of senescence are identified to be the most ideal phenological stages for accurate maize varietal discrimination. The overall discrimination accuracy improves by 52% and 63% with autoscaling at flowering and senescence, and by 55% and 62% with generalized least squares weighting at flowering and onset of senescence, respectively, compared to no preprocessing transformation. Of the three preprocessing transformations used, autoscaling and generalized least squares weighting are the most effective. Therefore, with appropriate preprocessing transformation, hyperspectral data and PLS-DA are effective in discriminating maize varieties.