Winter wheat grain protein content monitoring model driven by hyperspectral remote sensing images at different growth stages

Hyperspectral remote sensing data have strong band continuity, high spectral resolution and rich spectrum information. It can rapidly and nondestructively acquire vegetation information and it is an reliable real-time technology applicable in the monitoring and management of crop growth. Grain protein content (GPC) is an important indicator for wheat quality. Early detection of GPC of wheat using hyperspectral remote sensing data can enhance effective decision-making to support the acquisition and processing of high quality wheat. The objectives of this study were to establish GPC estimation model based on winter wheat canopy hyperspectral reflectance at different growth stages with different rates of nitrogen or phosphorus applications. The overall goal was to improve forecast precision of GPC estimation model at different growth stages of winter wheat. Thus experiments were carried out in 2009-2014 at Northwest A & F University, Shaanxi Province. The treatments included different winter wheat varieties with various drought resistances under five nitrogen fertilizer application rates (0, 75, 150, 225 and 300 kg·hm^-2 pure nitrogen) and four phosphorus application rates (0, 60, 120 and 180 kg·hm^-2 P₂O₅). Plant nitrogen content (PNC) and canopy hyperspectral reflectance of different wheat cultivars were measured at jointing, booting, heading, filling and maturity stages. Also GPC was measured at maturity stage. The relationship among PNC, canopy hyperspectral reflectance and GPC was explored using correlation analysis, regression analysis, grey relation analysis or partial least squares. The GPC monitoring model was built according to the relation of "vegetation index based on canopy hyperspectral reflectance (Ⅵ)—PNC—GPC" with PNC as the linking point. The results showed a higher GPC prediction accuracy by GPC monitoring model based on PNC at jointing, booting, heading, grain-filling and maturity stages. The monitoring models of PNC at jointing, booting, heading, filling, maturity stages of winter wheat respectively based on modified chlorophyll absorption reflectance index (MCARI₁), normalized difference chlorophyll index (NDCI), modified normalized difference vegetation index (mNDⅥ), MCARI₁ and NDCI had better estimations of PNC, with determination coefficients (R²) of 0.826, 0.854, 0.867, 0.859 and 0.819, accordingly. When linked with PNC, by using the "Ⅵ—PNC—GP" method, the GPC monitoring models for the maturity stage consisted of combinations of Ⅵ and PNC at jointing, booting, heading, filling, maturity stages had the determination coefficients (R²) of 0.935, 0.972, 0.990, 0.979 and 0.936, respectively. Then validatation of the models with measured values was conducted to verify the reliability and applicability of the models. The results showed that the relative errors (RE) between the measured and predicted values for the five vegetation indices were 11.26%, 10.74%, 8.41%, 10.25% and 11.36%, respectively. Then the corresponding root mean square errors (RMSE) were 2.221 g·kg^-1, 1.825 g·kg^-1, 1.214 g·kg^-1, 1.767 g·kg^-1 and 2.137 g·kg^-1. It therefore suggested that MCARI₁, NDCI, mNDVI, MCARI₁ and NDCI vegetation indices were the most suitable model for monitoring winter wheat GPC at jointing, booting, heading, filling and maturity stages, respectively. There was higher prediction precision with different vegetation indices at different growth stages monitoring winter wheat GPC under different N and P rates. Furthermore, the monitoring model based on different vegetation indices at different growth stages had a higher prediction accuracy. This results provided technical support for GPC monitoring of winter wheat at different fertilization and different growth stages.