Hyperspectral imaging for precision nitrogen management: A comparative exploration of two methodological approaches to estimate optimal nitrogen rate in processing tomato

Abstract

Hyperspectral imaging is widespread in crop nitrogen (N) monitoring for precision agriculture, although approaches that address the agronomical recommendation of the optimal N rate are still lacking. Here, two approaches are explored in defining the optimal N rate to be supplied in fertigated processing tomatoes through hyperspectral imaging. The first one, called the N uptake approach, focuses on the virtual reproduction of the critical N uptake curve through the estimation of both aboveground biomass and crop N uptake. The estimated biomass is used to derive the critical N uptake, and the optimal N rate is computed as the difference between the critical N uptake and the estimated actual N uptake. The second approach focuses on the monitoring of the Nitrogen Nutrition Index (NNI) and biomass. Again, the biomass is used to calculate the critical N uptake, which, when combined with the estimated NNI, resolves the equation to retrieve the actual crop N uptake. A modeling stage was included to estimate the N-related variables from crop canopy reflectance across the full spectrum (400–1000 nm). Canopy reflectance was measured by using an unmanned aerial vehicle at five growth stages of processing tomatoes grown under experimental plot conditions with different N rates. Three nonparametric algorithms were trained, i.e., Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Partial Least Square Regression (PLSR). Multicollinearity of spectral bands was prevented with a principal component analysis, and models were 5-fold cross-validated. Considering the pivotal role of biomass in the selected N rate estimation approaches, two distinct biomass estimation methods were explored. The direct biomass retrieval from spectral data was compared with the indirect biomass retrieval from the remotely sensed LAI applying empirical regressions. PLSR outperformed the other algorithms in estimating N uptake (Relative Root Mean Square Error, RRMSE=21.8 %), while SVR better estimated NNI (RRMSE=10.2 %) and direct biomass (RRMSE=19.4 %). The indirect estimation of biomass outperformed the direct approach when GPR is used (RRMSE 18.2 % vs. 21.4 %), although the influence of soil background at early growth stages determines an unreliable biomass estimation for both methods. The NNI approach outperformed the N uptake approach in estimating the optimal N rate, especially when the biomass is directly retrieved from GPR. The promising estimation performances in N rate estimation (R2=0.88 and RRMSE=36 %) revealed the effectiveness of hyperspectral imaging in entering the agronomical scheduling of precision N management.

Keywords

• Precision nitrogen management
• Hyperspectral imaging
• Machine learning
• Nitrogen nutrition index
• Nitrogen rate