Analysis On Influence Facters And Model Optimization Of Soil Organic Matter Predicton Accuracy Based On Vis-NIR Spectra

Soil Organic Matter(SOM)is an important index of soil quality,which is closely related to soil fertility,soil ecosystem and agricultural development.The difference of SOM content plays an important guiding role in agricultural production activities.The VIS-NIR hyperspectral quantitative analysis technology has the advantages of high efficiency and low cost,and can be applied to quickly obtain SOM content.However,most of the researches related to hyperspectral quantitative inversion of SOM are carried out on a single small sample set.Affected by the number of sample sets and the spatial range of sampling points,the researches on the influence of potential factors such as sample size,distribution of SOM content and soil physical and chemical parameters of key modeling technologies on the accuracy of SOM inversion have certain limitations.In view of the above problems,it is necessary to conduct in-depth research on the potential factors that may affect the inversion accuracy,and optimize the quantitative inversion model of SOM from the perspective of influencing factors,so as to effectively improve the inversion accuracy of estimating SOM content using hyperspectral technology.In this paper,the Lucas large sample spectral dataset covering 23 European member countries and the water-bearing soil sample set in the black soil region are taken as the research basis.Firstly,a variety of commonly used key techniques of hyperspectral organic matter inversion,spectral preprocessing and machine learning modeling methods are implemented.Based on the characteristics of Lucas dataset,such as large number of samples,large coverage area and strong spatial heterogeneity,the influence of sample data characteristics and key technologies on the accuracy of organic matter inversion was analyzed.On this basis,the influence of the contents of nine typical physicochemical parameters on the accuracy of organic matter inversion was systematically analyzed.In view of the influence of soil component content difference on the accuracy of SOM inversion,the spectral library-spectral filtering method was used to optimize the organic matter inversion model of local sample set,and the method was further applied to the water-bearing soil samples in the black soil area to achieve the removal of parameters other than SOM and the optimization of the model.At the same time,the multi-scale CWT-PCA combination method was used to reduce the influence of soil moisture on the accuracy of SOM inversion,and the key technologies of SOM inversion were optimized to achieve a fast and accurate inversion of water-bearing SOM.The main research contents and conclusions are as follows:(1)The influence of key factors of hyperspectral SOM inversion modeling on the inversion accuracy was analyzed,including the distribution range of SOM,sample size,key techniques of inversion modeling and other factors,which provided support for further analysis of the influence of soil physical and chemical parameters and inversion modeling of SOM.cLHS sampling method to extract different number of samples set,sample size calculation,spectral resolution scales,pretreatment method and modeling method of the inversion precision of cross change 4800 groups of SOM,through MANOVA influence significance of all the factors,the method of quantitative analysis to analyze the factors on the accuracy of spectral reflectance and SOM inversion are studied.The results show that:1)The greater the distribution range of SOM,the greater the stability R~2 of the organic matter inversion model,but the greater the prediction error RMSE.2)The larger the number of samples,the more stable the model.For the sample set used in this paper,relatively stable validation set accuracy can be obtained when the SOM distribution range is 0-8%and the number of samples reaches about400.3)When the spectral resolution scale is 6-10nm,the operation time of the model can be greatly reduced and the effect on the accuracy of SOM inversion is small;4)When the number of samples is large,the machine learning method has more advantages on the SOM inversion results;When the number of samples is small,the preprocessing method and modeling method are significant factors affecting the accuracy of SOM inversion.(2)From the distribution range,spectral reflectance characteristics of SOM,etc,analysis of the impact of typical soil physical and chemical parameters of the content change,under the premise of considering inversion key technologies influence further analyzed the influence of the inversion accuracy of SOM,soil physical and chemical parameters of the complexity of the inversion precision influence of SOM differences and severity.Seven data preprocessing methods such as CR,MSC and six modeling methods such as PLS and SVR were used to analyze the spectral characteristics of nine typical physical and chemical parameters such as N,P,K,sand and p H,as well as the accuracy of SOM inversion when the content of each parameter changes.The band range and degree of influence of each parameter are analyzed comprehensively.The results showed that:1)When the soil physical parameters sand and silt were grouped according to their content,the inversion accuracy of SOM in each group was higher than that of the total sample.When soil N,clay content and soil CEC value changed,the inversion accuracy was higher than that of the total samples only in the specific content range.2)Combined pretreatment combined with PLS-VIP method can effectively analyze the band range of the SOM prediction model affected by the change of each physical and chemical parameters;3)For the test sample sets,the SVR model has good model accuracy in most sample sets.Effective pretreatment methods combined with machine learning methods can reduce the influence of soil physicochemical parameter content differences to a certain extent.(3)In view of the influence of soil component content differences on the accuracy of SOM inversion,four local sample sets with different data characteristics were taken as the research object,and the spectral library-spectral filtering method was used to optimize the SVR inversion model,and the SOM inversion model with stronger prediction ability and robustness was verified.Aiming at the problem that soil component content difference interferes with the SOM inversion model,a spectral library combined with spectral filtering method is proposed to optimize the soil component content difference filtering,and the EPO,GLSW and YGLSW spectral filters are established by selecting the reference sample set from the spectral library.The results of the SOM inversion model were compared with those of the optimization method using key techniques of inversion modeling and the optimization method using typical physical and chemical parameters as auxiliary variables.The results show that:1)there is no general preprocessing and modeling method for the sample sets with different data features.The SOM inversion model can be optimized by combining multiple preprocessing methods with modeling method.2)Due to the influence of soil physical and chemical parameters on the content and spectral characteristics of SOM,the inversion model of SOM can be optimized by taking typical physical and chemical parameters as auxiliary variables.Soil N as the auxiliary variable has the best optimization effect,followed by soil sand,clay and silt as the auxiliary variables.3)The accuracy of SVR model optimized by spectral library-spectral filtering method can be better than that of SVR model without filtering for sample sets with different data features,and EPO filtering method has stronger applicability.The SVR model optimized by the best spectral library-spectral filtering method can have higher inversion accuracy than the model optimized by the key technology of inversion model and the optimization model with sand,clay and silt content as auxiliary variables.(4)In the dry soil samples of SOM inversion precision influence factor,the influence rule and model after optimization is studied,further research on the effect of soil moisture and model optimization strategy,the soil samples in black soil area at the water cut sets the research object,the verification effectively remove soil moisture interference,optimize water inversion model of SOM.The influence of soil moisture in the black soil area was studied,and the spectral library-spectral filtering method was further applied to the water-bearing soil sample set in the black soil area,and the black soil dry sample-spectral filtering method was taken as a reference.According to the data characteristics of the water-bearing sample set in the black soil area,a multi-scale CWT-PCA combination method was proposed to reduce the influence of soil moisture,and the optimization strategy of key technologies of inversion modeling could effectively improve the performance of the water-bearing SOM inversion model.The results show that:1)spectral library-spectral filtering method can effectively remove soil moisture effect,optimize water inversion model of SOM,in spite of the use of EPO,GLSW filter method when the model performance is better than from the black soil dry samples selected benchmark sample set,but it still has a good effect,and compared with more time to save cost.2)The YGLSW spectral filtering method only needs to take the reference correction set as the input matrix,which has a better development in the study of removing the influence of soil moisture.3)The multi-scale CWT-PCA combination method can be used as a spectral preprocessing optimization method to rapidly reduce the influence of soil moisture,and provide efficient and accurate data support for precision agricultural fertilization management.