Rice leaf nitrogen content estimation based on hysperspectral imagery of UAV in Yellow River diversion irrigation district

Nitrogen is essential for the improvement of photosynthesis and productivity of plants. However, nitrogen fertilizer is also a significant non-point source of water and atmospheric pollution. Therefore, a timely and accurate assessment of leaf nitrogen content (LNC) in crops is critical for crop growth diagnosis and precision management, eventually promoting crop yield and quality while minimizing environmental costs. The aim of this study was to determine the most suitable algorithm, based on hyperspectral reflectance data, for the regional assessment of LNC at critical growth stages of paddy rice. In this study rice experiments with different nitrogen levels and growth stages were conducted at different sites of Ningxia irrigation zone. Ground-based hyperspectral datasets were obtained from the stem elongation stage to the dough grain stage at plot and field scales. The plot and field datasets were used for model calibration and validation, respectively. A hyperspectral imagery was obtained over the field region at milk grain stage using UHD 185 carried by an unmanned serial vehicle (UAV). On the basis of a comprehensive analysis of the hyperspectral data, significant spectral indices (SIs) such as the normalized difference spectral index (NDSI) and ratio spectral index (RSI) were derived for an accurate and robust assessment of the LNC. Spectral indices representing a complete combination of the spectral bands between 450 nm to 950 nm were calculated using the NDSI and RSI formulations. The contour map of coefficient of determination (R ²) between LNC and the combinations of 2 separate wavelengths in the hyperspectrum was used to evaluate the new SIs through comparing the predictions with plot-experiment measurements and determine which one produce the higher prediction accuracy over the others. Then the predictions of the SIs were validated by independent datasets collected at field experiments. The capability of multivariable regression approaches such as partial least-squares regression (PLSR) was examined. R ², root mean square error (RMSE), relative error (RE) and relative prediction deviation (RPD) were employed to assess the model performance. The results showed that the reflectance spectra showed a positive response to the LNC in the near-infrared wavelength region and a negative response in the red region. The RSI using derivative values at around 738 to 522 nm was the superior SI in terms of its accuracy, simplicity, and applicability. The best estimation model of LNC was built. The model R ² and RMSE were 0.763 and 0.369 for calibration and 0.673 and 0.329 for validation, and the RPD was 2.02. These indicated that the model produced an acceptable result. We explored the relationship between the first derivative of reflectance at 738 and 522 nm as they were affected by the LNC. The first derivative of reflectance at 738 and 522 nm had different spectral responses to the change of LNC. The first derivative of reflectance at 738 and 522 nm were nearly proportional to the similar LNC values. However, the first derivative of reflectance at 738 nm increased and that at 522 nm decreased with the increase of LNC. Independent validation using the UAV dataset demonstrated the robustness of the new SI. The predication of LNC was 1.28%-2.56%, which was similar with measurements (1.34%-2.49%). Our study demonstrated that hyperspectral measurements provided a robust and practical tool to diagnostic mapping of the LNC on a regional scale.