Using Leaf Based Hyperspectral Models for Monitorin g Biochemical Constituents and Plant Phenotyping in M aize

Abstract

The aim of this study was to develop and validate q ualitative and quantitative models to discriminate different types of maize and also e stimate biochemical constituents. Spectral data were taken from the central leaf of r andomly-chosen plants grown in field trials in 2011 and 2012. Leaf chlorophyll and prote in content and stalk protein content were determined in the same plants. Four different Support Vector Machine (SVM) models were generated and validated in this study. In qualitative models, maize type was designated as dependent variable while Full Spectra l (FS) data (400-1,000 nm) and Spectral Indices (SI) data (34 indices/bands) were independent variables. In the two quantitative models (SVMR-FS and SVMR-SI), independ ent variables were the same, whereas dependent variables were assigned as the qu antitatively measured traits. Results showed the qualitative models to be a robust method of classification for distinguishing different maize types, such as High Oil Maize (HOM) , High Protein Maize (HPM) and standard (NORMAL) maize genotypes. The SVMC-FS mode l was superior to SVMC-SI in terms of the genotypic classification of maize p lants. Quantitative models with full spectral data gave more robust prediction than the others. The best prediction result (RMSEC= 222.4 μg g -1 , R 2 for Cal= 0.739, SEP= 213.3 μg g -1 ; RPD= 2.04 and r= 0.877) was obtained from the SVMR-FS model developed for c hlorophyll content. Indirect estimation models, based on relationships between l eaf-based spectral measurements and leaf and stalk protein content, were less satisfact ory. Keywords : Genotypic classification, Support Vector Machine, Zea mays.