Estimation model of wheat canopy nitrogen content based on sensitive bands

Abstract: Nitrogen is an important nutrition for wheat production and it affects significantly growth, yield and quality of wheat. Leaf non-destructive monitoring is the technique which is reported to provide accurate information of nutrition status of wheat. In this study, the raw hyperspectral reflectance of wheat leaf samples was measured by the standard procedure with an ASD FieldSpec3 instrument equipped with a high intensity contact probe under the laboratory conditions. Meanwhile, physical and chemical properties of these wheat leaf samples were analyzed. One hundred and ninety of 234 samples were used for building hyperspectral estimation models and the other 44 samples were used for model validation. In order to improve precision determination of wheat leaf canopy nitrogen content accuracy by hyperspectral technology, firstly, we compared the four different variables screening methods: MC-UVE，Random Frog，CARS and MWPLS. The selected sensitive bands by using partial least squares (PLS) model, and compared with full bands modeling. The number of variables and the variables selected by four methods were different. In the Random Frog and CARS methods, the variables were reduced to 10 and 39. Three hundred and eighty seven variables were selected by MC-UVE method and 425 variables were selected by MWPLS for calibration set and validation set. Based on the above analysis, CARS method and coefficient of determination error of the wavelength variables were optimal, and significantly improved the quality of modeling. For the nitrogen content estimation model based on sensitive bands selected by CARS, the coefficient of determination and error were optimal, resulting a significant improvement in the quality of modeling. Secondly, these collinear variables could contain a large number of redundant information. So, a combinatorial method named CARS-CC was proposed to select variables from 39 wavelength variables．As such, the number of wavelength variables was reduced to 30．It showed the method was effective．Finally, the BP, SVR and RBF models were developed with the selected variables by CARS-CC for leaf nitrogen of wheat. The selected wavelengths were used as the inputs of the models. The results showed that the BP model prediction coefficient of determination was 0.8247, root mean square error was 1.24; the SVR model prediction coefficient of determination and root mean square error were 0.847 and 1.248, and the RBF model prediction coefficient of determination and root mean square error were 0.9982 and 1.074e-009. They had an adequate precision and can quickly predict wheat leaf nitrogen content. For the prediction results of RBF neural network model of the optimal RBF model, the root mean square error of calibration set was 0.3699, the root mean square error of prediction was 1.074e-009, and the correction coefficient of determination and predictive correlation coefficient were 0.9832 and 0.9982. The experimental results showed that CARS-CC was a feasible and efficient algorithm for the spectral sensitive bands selection provided a theoretical basis for the applications of high spectral reflectance in non-destructive nitrogen level detection. At last, it could be concluded that the CARS-CC-RBF model for leaf nitrogen of wheat was better than CARS-CC-BP, CARS-CC-SVR models not only in full bands but also in significant bands. In the future, the CARS-CC-RBF model can be used as a reference for aerospace hyperspectral remote sensing of leaf nitrogen of wheat.