Accurate modeling of vertical leaf nitrogen distribution in summer maize using in situ leaf spectroscopy via CWT and PLS-based approaches

Abstract

The vertical leaf nitrogen (N) distribution in summer maize is believed to be an important adaptive reaction to crop physiology and ecosystem functions. Therefore, it is of great significance to understand the vertical characteristics of leaf N concentration (LNC) across summer maize canopies for crop growth and production. However, the effect of vertical canopy position on LNC and subsequently leaf hyperspectral characteristics of the entire spectrum (350–2500 nm) for summer maize is poorly understood. The purpose of this work was to quantitatively study the effects of the N nutrition on vertical distribution of LNC, identify the sensitive layer and effective wavelength of leaves, and establish a in situ leaf spectrum monitoring model considering vertical distribution of LNC. Six N field trials were conducted for four consecutive years (2017–2020). In addition, the data of 30 farmers' conventional farmland management in 2020 were collected to check the robustness of the constructed optimal estimation model for LNC prediction. Leaf spectral measurements together with LNC were studied at three vertical leaf positions on the crop stem: upper, middle and lower. Spectral reflectance was processed by continuous wavelet transform (CWT); partial least square (PLS) was used to analyze the relationships between LNC of different layers and spectral reflectance. Field sampling indicated that LNC had a vertical distribution pattern and an evident decline from the upper to lower layer. CWT technique can significantly increase the prediction accuracy of LNC at different layers, and the best decomposition scale is CWT-1. The CWT-1-PLS model for predicting vertical LNC distribution had achieved relatively higher accuracy than that based on the full range of the raw hyperspectral reflectance (R), the LNC determination coefficient (R2val) of the validation dataset was 0.832, 0.857 and 0.811 for the upper, middle and lower layer, and the relative percentage deviations (RPDval) were 2.444, 2.432, and 2.181, respectively. Eventually, ten bands were selected as the effective wavelengths for predicting the vertical LNC distribution in the upper, middle and lower layer, respectively. Newly developed CWT-1-PLS model using the effective wavelengths for LNC prediction in different layers also performed well (RPDval>2.0) based on the field experiments validation. Moreover, the validation at the farmers’ fields also showed fine precision for upper (RPDval=2.012), middle (RPDval=2.137) and lower (RPDval=1.881) layer LNC prediction. These results are of great significance for the study of summer maize leaf reflectance modelling, especially for the studies of integrating hyperspectral measurements and leaf traits data.