Remote sensing prediction of winter wheat grain protein content based on nitrogen nutrition index at anthesis stage

Abstract: Grain protein content (GPC) is an important quality index for wheat to meet a variety of needs of the commodity. Advanced site-specific knowledge of GPC would provide opportunities to the classification of wheat harvest and graded storage. Areas with higher GPC can be distinguished from the rest to maximize the price premium. Advanced knowledge of grain protein of the wheat may also provide opportunities to manipulate inputs to optimize outputs. In order to select an appropriate nitrogen nutrition index as an intermediate variable to improve the inversion accuracy of wheat GPC, in this study, the GPC predicting models at anthesis with vegetation index (VI) -nitrogen nutrition index (NNI) - GPC pattern were constructed and evaluated. The NNI included 4 nitrogen nutrition index, i.e. plant nitrogen accumulation (PNA), plant nitrogen content (PNC), leaf nitrogen accumulation (LNA) and leaf nitrogen content (LNC). In previous studies, only a single nitrogen nutrition index was used as the intermediate variable to construct GPC model, and it did not indicate which of the parameters could be utilized as the intermediate variable to obtain the best result. To improve the prediction model accuracy of GPC, we chose the optimal intermediate variable to retrieve the GPC of winter wheat in this study. Field experiments of 6 winter wheat cultivars in Beijing during the growing seasons of 2008-2011 and 2012-2015 were carried out for model building. Firstly, suitable vegetation indices were selected through analyzing the correlation between vegetation indices and nitrogen nutrition index to construct the model of VI-NNI by the support vector machines (SVM) algorithm and the optimal one was selected from the 4 nitrogen nutrition index. Secondly, the measured nitrogen nutrition index and winter wheat GPC were used to construct the NNI-GPC model, and another optimal one was obtained from the 4 nitrogen nutrition index. Thirdly, the 4 nitrogen nutrition index would be used as the intermediate variables to construct and evaluate VI-NNI-GPC prediction model of winter wheat, and we could get the optimal comprehensive model to retrieve GPC of winter wheat. We used modeling determination coefficient (R2), normalized root mean squared error (nRMSE) and normalized average error (NAE) to evaluate the accuracy of models. The results showed that: 1) The selected 5 vegetation indices, MSAVI (modified soil-adjusted vegetation index), PSRI (plant senescence reflectance index), DVI (difference vegetation index), RDVI (re-normalized difference vegetation index) and GNDVI (green normalized difference vegetation index), which were used to retrieve the 4 nutrition index, produced higher correlation than the other vegetation indices. 2) In the model of VI-NNI, the LNC estimated by SVM was relatively better with the R2 of 0.820 in the modeling, and the nRMSE of 9.553% and the NAE of -1.4% in the validation. 3) The modeling precision and validation precision of PNC in constructed NNI-GPC model were relatively high, with the R2, nRMSE and NAE values of 0.653, 9.843% and -0.3%, respectively. 4) The 'VI-NNI-GPC' model with the PNC as the intermediate variable performed better than the other intermediate variables, with the R2 of 0.631 in the modeling, and the nRMSE and NAE values of 8.564% and -0.9% in the validation, respectively. It demonstrates that it is feasible to use nitrogen nutrition index as intermediate variables to retrieve GPC by remote sensing, and using PNC as the intermediate parameter achieves more accurate prediction results. This study provides a reliable reference for the accurate prediction of GPC and has a broad range of potential applications.