Nitrogen nutrition diagnosis of winter wheat based on ASD Field Spec3

Abstract: For crop’s prediction model of nitrogen nutrition index (NNI), how to select the pretreatment and modeling method is unclear as well as different pretreatments and their influence degrees on prediction accuracy. So it is of great significance to take more systematic related research for building crop nitrogen nutrition diagnosis rapidly and accurately, which can provide important technical support for precision agriculture management, and realize high yield with high efficiency of nitrogen utilization. Taking Nanxia Village, Laoling City in North China Plain as the research area, based on ASD Field Spec3, the prediction model of winter-wheat nitrogen nutrition index was established with hyperspectral technology in this study. PLSR combined with different pretreatments was applied to establish an prediction model of winter-wheat nitrogen nutrition index, whose average value of model-set model decision coefficient R2 was 0.683, with the maximum one 0.789, the minimum root mean square error (RMSE) 0.142, and the minimum of relative medium error (RE) 12.3%. The prediction-set model’s mean value of decision coefficient R2 is 0.588, with the maximum one 0.717, the minimum value of root mean square error (RMSE) 0.150, and the minimum of relative medium error (RE) 12.8%. The comparison shows that the pretreatment methods with SG(Savitzky-Golay), SNV(standard normal variate transformation), SG+SNV and SG+BC(baseline correction) are effective when partial least square method is used to build the model, especially SG smoothing is the optimal one as mentioned above with the R2 of 0.789, the RMSE of 0.142, the RE of 12.3%, and the R2 of the prediction accuracy of 0.717. Meanwhile, BP neural network method combined with different pretreatments was used to establish an prediction model of nitrogen nutrition index, whose average value of model-set model decision coefficient R2 was 0.834, with the maximum one 0.861, the minimum RMSE 0.115, and the minimum of RE 9.8%. The prediction-set model’s mean value of decision coefficient R2 was 0.714, with the maximum one 0.780, the minimum value of RMSE 0.133, and the minimum of RE 12.3%. It can be known that the regression model constructed by BP neural network is significantly more accurate than the one constructed by PLSR model. The decision coefficient of all the models pretreated by BP neural network was above 0.8, while the average one was increased to 0.834 from 0.683 under PLSR model. All pretreatment modeling accuracy combined with SG, SG+MSC(multiple scatter correction), SG+SNV and SNV+D(De-trending) reached above 0.85, when the predictive effect reached above 0.7 except BC, MSC and SNV. SG pretreatment R2 with the best modeling effect reached 0.861, with mean square root error 0.115, relative error 9.8% and predicted effect R2 0.780 as mentioned above. The NNI estimation model RF algorithm combined with different pretreatments was used to establish an prediction model of NNI, whose average value of model-set model decision coefficient R2 was 0.945, with the maximum one 0.959, the minimum RMSE 0.061, and the minimum of RE 5.3%. The prediction-set model’s mean value of decision coefficient R2 is 0.742, with the maximum 0.795, the minimum value of RMSE 0.125, and the minimum of RE 11.7%. It can be known that the regression model constructed by RF is significantly improved compared with PLSR model and BP neural network. The decision coefficient of all the models by RF are all above 0.9, and the prediction model accuracy R2 is above 0.7. After 10 spectral pretreatment methods and 3 modeling ones have been comprehensively compared in this study, it is found that different pretreatment and modeling methods have great impacts on modeling precision. The optimal hyperspectral modeling method is RF(random forest) algorithm. The average value of decision coefficient R2 for the prediction-set model obtained through RF algorithm was higher than the biased least squares model and BP(back-propagation) neural network respectively, with lower RMSE and RE. Therefore, it can be seen that 10 pretreatment methods combined with RF model have higher prediction accuracy, which is a robust modeling method to invert nitrogen nutrition index. From above, to preprocess the spectrum for winter wheat can improve modeling accuracy. The best spectral pretreatment method in this experiment is SG convolution smoothing. Therefore, by comparing three models, the average R2 increase range of SG convolution smoothing compared with the other nine pre-processed prediction set models is 0.054~0.121, with average RMSE decrease range 0.016~0.032 and average RE decrease range 1.7~3.1 percentage points. In terms of single preprocessing, other preprocessing transformations can be carried out after combining with SG convolution smoothing, which can also improve the prediction accuracy of its single preprocessing model.