Study On Spatial Prediction Of Soil Organic Matter In Complex Environmental Region

Soil Organic Matter(SOM)is an important index of soil fertility evaluation and an important content of soil ecosystem assessment.The spatial distribution of SOM is affected by many factors,such as climate,parent material,biology,terrain,time and human factors.In order to study the spatial distribution of SOM,scientific field management and improve ecosystem services evaluation system,it is of great significance to study the influence mechanism of environmental factors on SOM,and establish spatial prediction model of SOM.Classical statistics,principal component analysis and machine learning algorithm were used to select the best explanatory variables to reduce the number of variables without loss prediction power.Geostatistics was used to analyze the spatial characteristics of regression models residuals to predict the spatial distribution of SOM.The main results are as follows:1.Correlation between SOM and environmental variables was analyzed and the dimension of environmental variables was reduced.Pearson's coefficient shows that there is significant correlation between SOM and slope,elevation,surface roughness,relative elevation,SOS,SOA,and STI.Stepwise regression analysis finally chose surface roughness,SOS,elevation,and SOA as explanatory variables in the multiple linear regression equations.The principal component analysis compresses the information of the original environmental variables and extracted 76.0% of the original variables into four mutually independent principal components.Reach the goal of eliminating multicollinearity among variables and reducing dimensionality of environmental information.2.The multiple linear regression model,principal component regression model,partial least squares regression model and support vector machine regression model was established.The introduction of machine learning algorithm can improve the degree of information mining,at the same time,it reduced the waste of environmental information caused by correlation analysis and principal component analysis.The results shows that:(1)the response time of the regression model based on the machinelearning algorithm is significantly longer than that of the traditional linear regression model,and the response time of the neural network reaches 7200 s.(2)the multiple linear regression model as reference,except the principal component regression model,the other regression models had better fitness.(3)from the AIC index,the neural network model and the support vector machine model had better fitting results than the multivariate linear regression model,the principal component regression model and the partial least squares regression model.3.The regression model didn't take the spatial autocorrelation of SOM into account,so it was important to analyze the spatial variation of the residual parts that can not be explained by the regression model,and establishes the Original Kriging model.Combined the Original Kriging model with regression models,the spatial prediction of SOM in the study area is carried out.The results show that the final prediction map of different models are similar.The SOM presents a differential distribution in the different parts of region.The content in the southwest is high and the northeast is low.The distribution characteristics are consistent with the terrain characteristics of the study area from the northeast plain to the southwest hilly area.The spatial distribution pattern of MLRK prediction results is similar to the spatial distribution of environmental variables in the model.In the PCRK model,the fragments are merged under the condition that the number of reserved patches is abundant.The number of spots in PLSRK model was further reduced,and the transition curves were obvious.The ANNK model can accurately predict the difference of low value area of SOM,indicating that it is better than other models to extract terrain information in flat terrain.The SVMK model predicts that the transition between high value area and low value area is blurred,and the number of spots is large,which improves enrichment of information in prediction chart.