Design And Verification Of Soil Nutrient Content Inversion Model Based On Improved SVM Algorithm

Spectroscopy for soil attribute information,combined with machine learning algorithms regression analysis was carried out on the soil nutrients become a research hotspot,small sample forecast analysis often USES the SVM algorithm,but the SVM model based on radial basis kernel function while largely improve the efficiency of prediction,but accuracy is low,how to improve the model prediction accuracy become a research hotspot.The mulberry field of the test station of the South Campus of Shandong Agricultural University in Tai’an,Shandong Province was taken as the research area.Soil samples with a large geographical span were taken according to the five-point sampling principle.The air-dried and sieved soil samples were subjected to visible-near infrared spectroscopy data collection and soil nutrient content determination.Subsequently,Monte Carlo Simulation was used to eliminate abnormal samples,and different preprocessing was performed on the spectral data after the abnormal samples were eliminated,and the optimal preprocessing method was selected.In order to remove redundant information and increase the speed of calculation,five algorithms such as CARS,IRIV and CARS-IRIV algorithm were adopted Feature variable selection.Then,the regression prediction model is used to predict the soil nutrients.However,because of the limitations of traditional prediction models when applied,the traditional models must be integrated and optimized.Therefore,this paper adopts the hybrid kernel function support vector machine regression model,and introduces the Whale optimization algorithm(Whale optimization algorithm,WOA)to analyze the g(kernel function parameter),c(penalty factor)and k-rbf(weight coefficient)in the hybrid kernel support vector machine.Parameter optimization,build a hybrid kernel function support vector machine(RBF-poly-SVM)soil nutrient prediction model based on the whale optimization algorithm.The specific research results are as follows:(1)Standard normal transformation(SNV),multivariate scattering correction(MSC),SG smoothing,first derivative,square root,SG-square-FD and other methods were used to preprocess the spectral data.Combined with PLSR to select the optimal pretreatment method,the results show that the optimal pretreatment method for soil total nitrogen is SG-square-FD,the optimal pretreatment method for total phosphorus is square-FD,and the optimal pretreatment method for total potassium is Square root,the best pretreatment method for available nitrogen is SNV-DT,and the best pretreatment method for organic matter is square root.(2)Due to the limitation of single prediction model,this paper used the mixed kernel function composed of radial basis function kernel and polynomial kernel function to establish the mixed kernel SVM model,and established the radial basis function kernel SVM model for comparison.In order to improve the prediction performance of the two prediction models,WOA was used to optimize the kernel function parameters.The results show that the mixed kernel SVM model was superior to the radial basis kernel SVM model in predicting soil nutrients.(3)Prediction of soil total nitrogen,total phosphorus,and total potassium content based on improved SVM algorithm.After preprocessing,CARS,IRIV,SPA,CARS-IRIV and VISSA-IRIV algorithms were used for characteristic variables selection.And the selected characteristic variable were used to predict soil nutrients by RBF-poly-SVM and RBF-SVM models.The results show that the RBF-poly-SVM model based on the characteristic variables extracted by the IRIV algorithm has the best prediction results for total soil nitrogen.The correction set determination coefficient is Rc2=0.925,the prediction set determination coefficient is Rp2=0.918,and the RPD value is 3.198.The RBF-poly-SVM prediction model established by the characteristic variables extracted by the IRIV algorithm for soil total phosphorus is the best.The accuracy of the correction set is Rc2=0.999,the accuracy of the prediction set is Rp2=0.937,and the RPD value is 3.177.The RBF-poly-SVM prediction model established by the feature variables extracted by the VISSA-IRIV algorithm for total potassium in soil has the best results.The accuracy of the correction set is Rc2=0.955,the accuracy of the prediction set is Rp2=0.854,and the RPD value is 2.608.Compared with the RBF-SVM model,the RBF-poly-SVM prediction model has a better matching effect and improved accuracy.(4)Based on the detection of soil available nitrogen and organic matter content by improved SVM,the prediction models of RBF-poly-SVM and RBF-SVM were established respectively after spectral reflectance pretreatment and feature variable selection.The results show that the RBF-poly-SVM prediction model based on the feature variable extracted by IRIV algorithm for soil available nitrogen samples is the best,the coefficient of determination of correction set is Rc2=0.913,and the coefficient of determination of prediction sets is Rp2=0.845,RPD=2.268.The results show that the RBF-poly-SVM prediction model based on the characteristic variables extracted by IRIV algorithm for soil organic matter samples is the best,the determination coefficient of correction set is Rc2=0.962,the determination coefficient of prediction set is Rp2=0.912,and the RPD is 3.075.