Random forest regression results in accurate assessment of potato nitrogen status based on multispectral data from different platforms and the critical concentration approach

Abstract

Remote sensing can be used for precision nutrient management to assess plant nitrogen (N) status in a spatially detailed and real-time manner. Despite recent advances in satellite- and drone technology and machine learning, neither differences between platforms nor methodological aspects for estimating plant N status have been sufficiently investigated. In this study, multispectral data obtained by ground (handheld Rapidscan), air- (unmanned aerial vehicle, UAV) and spaceborne (Sentinel-2) platforms were exploited to estimate plant N uptake (PNU), concentration (PNC) and N nutrition index (NNI). The test plant was potato grown for three years on a sandy soil in Denmark and the analysis was based on the critical N dilution curve. Parametric (PR) and non-parametric (random forest, RFR) regressions were conducted and compared in predicting mid-season PNU, PNC and NNI from band reflectances or vegetation indices (VIs) derived from each platform data. The results obtained by the UAV data had the highest accuracy, largely due to the fine spatial resolution. For both regression types, PNU and NNI correlated better than PNC to reflectance data. For the UAV data, validation Nash-Sutcliffe model efficiency (NSE) of PNU and NNI ranged between 0.64–0.95 and 0.41–0.92 respectively, with corresponding values for relative root mean square error (RRMSE) of 7.1–22% and 5.86–22%. The lower end of NSE and higher end RRMSE intervals systematically being from the PR, which demonstrates the robustness and the high accuracy of RFR in predicting plant N status. The other platforms resulted in acceptable results, with validation NSE and RRMSE for PNU and NNI of, respectively, 0.60–0.79 and 14–20%, 0.25–0.79 and 10–17% for Rapidscan, and 0.48–0.83 and 17–28%, 0.42–0.82 and 12–19% for Sentinel-2. The band reflectance and the VIs were equally suited as input predictors for the RFR algorithm. The N requirement calculated from all three datasets reflected the field observations well. The study reveals the potential of different regression methods for detailed spatial estimation of plant N status to guide in-season fertilization by matching the plant growth demands, emphasizing the strengths of the RFR. The procedure is helpful for the digital agriculture and the smart farming industry aiming to avoid excess application of N.