Assessing leaf nitrogen concentration of winter oilseed rape with canopy hyperspectral technique considering a non-uniform vertical nitrogen distribution

Abstract Timely estimation of the vertical heterogeneity of leaf nitrogen concentration (LNC) from canopy reflectance using hyperspectral sensing is important for precision N management during winter oilseed rape productivity. However, current research pays little attention to LNC assessments by only taking LNC’s vertical distribution into consideration, leading to limited accuracy and reduced applied value of the results. The main goal of this work was to quantitatively define the contributions of LNC in different layers to winter oilseed rape canopy raw (R) hyperspectra and to its transformation technique (i.e., first derivative reflectance, FDR), and develop a monitoring model considering the vertical LNC gradient using spectral data. Two field experiments were conducted for two consecutive years (2015–2017) with different N rates, cultivars and growth stages. At seedling and budding stage, canopy hyperspectral reflectance and LNC were measured in situ. Canopies of each treatment were divided into three layers of equal vertical (upper, middle, lower). Partial least square (PLS), lambda-lambda r2 (LL r2) and support vector machine (SVM) models were used to analyze the relationships between LNC in different layers and the hyperspectral reflectance measured from above the canopy. Field sampling revealed that a vertical distribution pattern of LNC existed, presenting an evident decline from the upper to lower layer. The FDR-PLS model for LNC prediction in different layers yielded a relatively higher accuracy compared to the R-PLS based on the full range hyperspectra, the coefficient of determination (r2 val) was 0.872 for LNC in the upper layer, 0.903 in the middle layer, and 0.837 in the lower layer, with a relative percent deviation (RPD val) of 2.794, 3.052, and 2.328, respectively. Finally, seven (437, 565, 667, 724, 993, 1084 and 1189 nm), six (423, 570, 598, 659, 725 and 877 nm), and five bands (420, 573, 597, 667 and 718 nm) were identified as effective wavelengths for assessing the vertical LNC distribution in the upper, middle and lower layer, respectively. The newly-developed SVM-FDR regression model using the effective wavelengths also performed well for upper (r2 val = 0.828, RPD val = 2.358), middle (r2 val = 0.844, RPD val = 2.556), and lower (r2 val = 0.781, RPD val = 2.029) layer LNC prediction. Our results indicate that estimation of LNC using hyperspectral reflectance data is most effective for the upper and middle layers of oilseed rape canopies. Moreover, the calibration model developed in this study has great potential to assess the N status of the whole oilseed rape canopy.