A spectral decomposition method for estimating the leaf nitrogen status of maize by UAV-based hyperspectral imaging

Abstract

Leaf nitrogen status plays a crucial role in characterizing maize nutrient activity, which ultimately affects both the photosynthetic efficiency and yield formation of maize. For this reason, hyperspectral imaging technology based on unmanned aerial vehicle (UAV) has emerged as a popular tool for estimating crop phenotypic traits. Canopy structure and leaf nutrition together determine the crop canopy spectrum. Efficient separation of the spectral information sensitive to the leaf nitrogen status from the canopy spectrum is considered of utmost importance for improving the estimation accuracy of maize leaf nitrogen status. Along these lines, the main goal of this work was to develop a canopy spectral decomposition method to reduce the interference of other traits on leaf nitrogen concentration (LNC) and leaf nitrogen density (LND) estimation using UAV-based hyperspectral images. First, the weights of the leaf nitrogen status, aboveground biomass (AGB), and leaf area index (LAI) contributing to the canopy spectrum were calculated by using the entropy weight method. Then, the sensitive bands of LNC and LND before and after spectral decomposition were selected by implementing the successive projections algorithm (SPA) and competitive adaptive reweighted sampling (CARS) algorithm. Finally, the reflectance of all bands and sensitive bands was compared to estimate maize LNC and LND using partial least squares regression (PLSR), Gaussian process regression (GPR), support vector regression (SVR), and random forest regression (RFR). These models were systematically tested using independent samples. From the acquired results, it was demonstrated that the correlation coefficient (r) between LNC and each band increased after spectral decomposition compared to the correlation before spectral decomposition. The r between the band reflectance in the near infrared region and LNC or LND after spectral decomposition was significantly higher than that before spectral decomposition. In addition, the sensitive bands of maize leaf nitrogen status after spectral decomposition were around 470 nm, 538 nm, 638 nm, 682 nm, 710 nm, 734 nm, and 830 nm. Regardless of using the reflectance of all bands or sensitive bands, the four estimation models of LNC and LND after spectral decomposition performed better than those before spectral decomposition. The estimation models of LNC and LND constructed by CARS-SVR can successfully reproduce the estimation accuracies of the models constructed by using all-bands-SVR, with R2 on the testing set of 0.68. Theresultshighlight that spectral decomposition is an effective method to significantly improve the estimation accuracy of maize leaf nitrogen status using UAV hyperspectral images, thus effectively reducing the interference of canopy structure traits (AGB and LAI).