Comparison And Analysis Of Techniques And Methods For Predicting Topsoil Organic Matter Content In Agricultural Land At County Scale

Soil is the medium for the growth of green plants and one of the most important natural resources on which human life depends.Soil organic matter,as an important component of soil,is one of the important indicators of soil traits,and it is often chosen as a factor for evaluating soil fertility and soil quality.SOM is the basis for estimating soil organic carbon reserves,and has an important impact on the global carbon cycle and climate change.Therefore,accurately grasping the spatio-temporal variability of SOM based on limited sample data has received more and more attention.The content of OM in topsoil of cultivated land is closely related to the growth of crops.It is of great significance to master the spatial distribution of SOM to guide the cultivation of soil fertility and agricultural production.Different soil types often have certain differences in their properties,so the same explanatory variables may have different effects on the variation of SOM.For this research,Biyang county was selected as a case study and the SOM information of 2172 soil profiles as basic data were collected from special project of cultivated land fertility evaluation between 2009 and 2011?The second national soil survey database,the second national land use survey database,Biyang county statistical yearbook,and remote sensing images were collected and compiled.Natural environmental factors(elevation,aspect,slope,NDVI),soil property factors(Soil texture)and farmland management factors(irrigation level,drainage capacity)were selected as explanatory variables.The soil types in Biyang county include yellow cinnamon soil,yellow brown soil,alluvial soil and lime calcic black soil,etc.).About 70%of soil samples were collected as training data set from different soil types separately,using stepwise linear regression,spatial auto-regression,geographically weighted regression,and random forest to model the relationship between variables.The most suitable model was selected to predict the organic matter in different soil types of agricultural land,and a spatial distribution map of SOM in agricultural land in the study area was obtained.The results showed that:(1)The average value of SOM in the study area was14.26 g/kg,while average of lime calcic black soil was highest,16.44 g/kg.The statistical characteristics of samples'SOM in alluvial soil and fragmental soil are similar,and it is suitable to build one model as a data set.Yellow cinnamon soil and yellow-brown soil are suitable as a class,lime calcic black soil as one.(2)There was a significant spatial autocorrelation in the organic matter content of the samples.There were different degrees of local spatial aggregation in organic matter of different soil types.(3)The accuracy of back-generation verification of the training set by regression models was superior to the cross-validation prediction,while the RF model did not have the problem of over-fitting.(4)Regardless of indicators such as goodness-of-fit R~2 or Pearson's correlation coefficient with actual values,the performance of GWR models was better than that of the other two regression models,comparable to RF models.The prediction effect of RF in alluvial soil and fragmental soil was the best.The Pearson's correlation between predicted values of all samples and actual values reaches 0.875,but the residuals still have some spatial autocorrelation.The results of other two groups predicted by GWR models were optimal,and the correlations with the actual values are 0.653(yellow cinnamon soil and yellow brown soil)and 0.617(lime calcic black soil),respectively,and the autocorrelation of the residuals almost disappeared.(5)For alluvial soil and fragmental soil agricultural land,the factors that affected the variation of SOM were aspect and irrigation level.Elevation was the most significant variable under yellow cinnamon soil and yellow brown soil,as same as lime calcic black soil.(6)The spatial distribution prediction map of the study area showed that SOM of agricultural land in the northwestern corner of Biyang County was at a relatively high level,corresponding to the distribution area of lime calcic black soil,and other high values were distributed discretely in the northern areas while SOM of agricultural land in the southwest side was generally relatively low.This paper studied the variation characteristics of organic matter content of agricultural land in different soil types.Through the comparison between the methods,it revealed the effectiveness and application potential of RF model and GWR in the spatial prediction of soil properties and digital mapping.It was proved that the causes of variation in organic matter content in agricultural land under different soil types are different.In order to obtain more accurate prediction results,a corresponding appropriate model needs to be constructed.