| As an important index of soil fertility,soil nitrogen plays a long-term key role in plant growth and development.Although the excessive application of nitrogen fertilizer has increased the yield of grain,it also brings a series of problems such as the decrease of biodiversity and soil acidification.In order to avoid the waste of nitrogen fertilizer,it is necessary to measure the total nitrogen content of soil.At present,the determination of soil total nitrogen mostly by traditional chemical methods,but it is time-consuming and expensive,and it also brings a series of problems such as drug residues and environmental pollution.With the development of spectral technology,the quantitative estimation model of soil total nitrogen based on mid-infrared spectroscopy can effectively estimate soil total nitrogen content accurately.In order to explore the quantitative prediction model suitable for total nitrogen in farmland soil in Inner Mongolia,this study collected 450 farmland soil samples in Hulunbuir City and Bayannur City in Inner Mongolia Autonomous Region and measured their total nitrogen content and mid-infrared spectrum information respectively.SG smoothing(SG),standard normal transformation(SNV),multiple scattering correction(MSC),normalization(Normalize)pretreatment,SG smoothing combined with SNV pretreatment(SG+SNV),SG smoothing combined with MSC pretreatment(SG+MSC),and SG smoothing combined with Nor were performed for the collected mid-infrared spectral information,respectively malize pretreatment(SG+Normalize),and combined with soil total nitrogen content,partial least square regression(PLSR),support vector machine(SVM),random forest(RF)quantitative prediction models of soil total nitrogen were established,and the effects of different pretreatment methods on model accuracy were explored through model verification results.The optimal modeling method for predicting the total nitrogen content of farmland soil in Inner Mongolia was selected.The samples of Hulunbuir and Bayannur are respectively brought into the optimal model and compared with the prediction results of separate modeling in the two regions,so as to explore whether separate modeling is needed in different regions.The results are as follows:(1)Soil mixing modeling results in Hulunbuir and Bayannur regions showed that all pretreatment methods in PLSR model and RF model could effectively improve the signal to noise ratio of infrared spectrum in soil and increase the prediction accuracy of the model.For SVM model,SNV pretreatment can effectively improve the prediction accuracy of the model,but SG smooth pretreatment can reduce the prediction accuracy of the model.The optimal preprocessing method corresponding to PLSR model is SG+Normalize(R2=0.9178,RMSE=0.1678,RPD=3.4286).The optimal pretreatment method corresponding to SVM was SNV(R2=0.9256,RMSE=0.1643,RPD=3.5019).The optimal pretreatment method corresponding to RF is SG+Normalize(R2=0.81,RMSE=0.2879,RPD=1.999).SNV-SVM method is the most suitable method for the quantitative prediction model of soil total nitrogen in the two regions.(2)Soil in Hulunbuir area and Bayannur area were respectively introduced into the mixed modeling SVM model.The results showed that the mixed modeling SVM model had a better prediction effect on Hulunbuir area,but a worse prediction effect on Bayannur area,and the model was of poor universality.Among them,the optimal effect of Normalize pretreatment in Hulunbuir was the best(R2=0.90,RMSE=0.3383,RPD=2.9415),and the optimal effect of SG smoothing pretreatment in Bayannur was the best(R2=0.52,RMSE=0.1151,RPD=1.4703).(3)Compared with the independent modeling and mixed modeling,the prediction effect of Hulunbuir(R2=0.90,RMSE=0.1622,RPD=3.6020)and Bayannur(R2=0.78,RMSE=0.1109,RPD=1.5466)was improved.The results show that SG+Normalize preprocessing has the best effect on model optimization.(4)The modeling results of the three sample diversity methods show that the accuracy of the model using SPXY algorithm diversity is better than KS algorithm diversity and random diversity.Stochastic diversity is better than KS algorithm diversity in Bayan Nur region prediction,KS algorithm diversity is better than random diversity in Hulun Buir region prediction and mixed prediction of two regions.Therefore,the results of this study show that in order to achieve high prediction accuracy,SG+Normaliz-SVM models should be established for different regions,so as to obtain reliable prediction models. |
