Multi-year hyperspectral remote sensing of a comprehensive set of crop foliar nutrients in cranberries

Abstract

Hyperspectral remote sensing has emerged as an efficient tool to quantify the spatial and temporal variations in crop foliar nutrients, thus reducing the burden on in-situ tissue sampling and traditional chemical assays. However, the physical mechanism of hyperspectral remote sensing of foliar nutrients is under-explored, especially for those lacking absorption features. Using four-year data collected from a cranberry farm, we demonstrate the capacity of leaf and imaging spectroscopy to quantify a comprehensive set of crop foliar nutrients, including seven macronutrients (N, P, K, Mg, Ca, S, Na) and five micronutrients (Fe, Mn, B, Cu, Zn). Specifically, we: 1) compared the performance of four data-driven approaches to estimate foliar nutrients at both leaf and canopy scales, including partial least square regression (PLSR), support vector regression (SVR), Gaussian process regression (GPR) and random forest regression (RFR); and 2) explored the physical basis of hyperspectral remote sensing of foliar nutrients. Our results showed that: 1) at leaf scales linear approaches PLSR and SVR performed best for nine nutrients (P, Mg, Ca, S, Na, Fe, B, Cu and Zn), whereas nonlinear approaches GPR and RFR performed best only for three nutrients (N, K and Mn); 2) at canopy scales no data-driven approach significantly outperformed others; 3) the best modelling accuracy varied with foliar nutrients (leaf scales: R2 from 0.30 to 0.93 and RRMSE from 9 to 51%; canopy scales: R2 from 0.15 to 0.81 and RRMSE from 7 to 37%). The physical basis of hyperspectral remote sensing of foliar nutrients was mainly attributed to their strong correlations with leaf compounds that have apparent absorption features. More specifically, at leaf scales the correlation between foliar nutrients and LMA (leaf mass per area) was leveraged by models to predict foliar nutrients from leaf spectra; at canopy scales the correlation of foliar nutrients with leaf chlorophyll and canopy LAI (Leaf area index) was leveraged by models to predict foliar nutrients from canopy spectra. This study revealed the importance of trait correlations in predicting foliar nutrients, and improved our understanding of the physical mechanisms in hyperspectral remotes sensing of foliar nutrients.