Evaluation of the Relationship Between Different Hyperspectral Indices and Leaf N in Wheat Using Mixed Model Theory

Hyperspectral (HS) data of vegetation provide a wealth of detailed information to identify plant nutritional status, but data analysis is fundamental for exploiting their full potential. The discrete waveband approach (Heege, 2013), through computation of narrowbands vegetation indices (VIs), is the most straightforward strategy used to synthetize information from plant spectral signatures. However, saturation effects are known to occur for some VIs over specific LAI values; in addition, bands capturing most of information of crops characteristics may vary across growing cycle modifying VI efficacy. Conversely, full spectrum approach, through multivariate analysis, could synthetize whole plant spectral response. In any case, the statistical function is still depending on growth stage and plant health. In the study of the relationship between HS data and biochemical parameters, ordinary least squares models (OLS) are commonly employed but their use requires important assumptions to be satisfied. However, residuals are often spatially correlated and, when spatial dependence is not taken into account, type I error tends to increase leading to results misinterpretation and improper management decisions. Generalized least squares models (GLS) with correlated errors allow spatial correlation components to be assessed and filtered from the residual term (Rodrigues et al., 2013). The aim of this study was to investigate the efficacy of different approaches for HS data analysis in estimating leaf N in wheat. To reach this aim, models accounting for spatial correlation were used and compared to OLS models.