Research On Spatial Variation And Mechanism Of Farmland Soil Organic Carbon In Plains

The farmland soil carbon pool is the most active component in the soil carbon pool.It responds rapidly with the change of cropping and management systems,and has great potential for carbon sequestration,which is helpful to realize the goal of "carbon neutralization" proposed by China.Soil organic carbon(SOC)is an important index to evaluate farmland fertility and soil quality.The content and change of SOC in farmland directly reflect the fixation or loss of soil carbon.Therefore,monitoring the spatial distribution of SOC in farmland is conducive to ensuring food security and predicting global climate change.Current digital soil mapping methods mainly rely on the relationship between SOC and environmental variables,that is,using easily available environmental variables to predict the spatial distribution of SOC.However,in low-relief farmlands,the traditional soil-landscape model that highly depends on terrain factors performs poorly for explaining SOC variation.The effective environmental variables of farmland SOC in plains are not clear,and thus the prediction accuracy of SOC is poor.Commonly used machine learning algorithms such as support vector machine and artificial neural network can describe the complex nonlinear relationship between SOC and environmental variables,but they are easy to overfit and the poor ability of model interpretation limits their accuracy.Additionally,these models assume that the relationship between SOC and environmental variables is stable in the whole area,and ignores the stratified and local heterogeneity of such relationships.To contribute to filling these knowledge gaps,a total of 242 farmland samples were collected from Chahe town of Jianghan plain.We obtained the spatial distribution of various potential influencing factors such as natural factors,cropping and management factors and landscape metrics,and determine the significant influencing factors of SOC by using single linear regression and one-way ANOVA.Stepwise linear regression(SLR),random forest(RF),Cubist,geographic weighted regression(GWR)and multiscale geographic weighted regression(MGWR)were used to determine the main controlling factors of SOC at global scale,each stratum and each local site.Combined with these variables and various prediction models,the spatial map of SOC in the study area was obtained,and the optimal prediction model was determined.The specific research content and results were as follow:(1)The spatial distribution of natural factors,cropping and management factors and landscape metrics were obtained,and the significant factors of SOC were determined.The spatial maps of cropping and management factors including crop type,multiple cropping index,residue index were obtained from the time series remote sensing image of HJ-1A/B and Landsat 8 satellites.The landscape metrics of sampling points on different scale(100-1000m)were obtained based on the land use map.The spatial distribution of natural factors such as soil types,terrain factors,climate factors and distance factors were obtained via open source.The single linear regression and one-way ANOVA were used to determine the relationship between SOC and these potential influencing factors.The results showed that elevation,slope,distance to the lake,land use types,crop types,multiple cropping index,cropping systems,normalized difference index and IJI,the percentages of pond and irrigated canals with 300 m buffer were the significant influencing factors of SOC.(2)Combined with the significant influencing factors and various data mining methods,the main control factors of SOC at global scale,each stratum and and each local site were determined.On the basis of the previous chapter,SLR and RF,Cubist,GWR and MGWR were separately used to explore the relationship between SOC and influencing factors at global scale,different strata and local sites,and the main control factors of SOC were determined according to the relative importance indexes.The results of SLR and RF showed that human activity factors such as land use types,multiple cropping index,residue index and percentage of pond were the main control factors of SOC at the global scale.The results of Cubist model showed that the relationship between SOC and environmental variables has stratified heterogeneity: the SOC of irrigated land samples was mainly affected by the distance to the lake,the percentages of pond and irrigation canals that are related with water,while the SOC of paddy field samples was affected by straw returning,terrain factors and landscape metrics.The MGWR that considers the diversity bandwidth of independent variables outperformed GWR model.The MGWR results showed that the main control factors of SOC at different sites are quite different.In general,local regression model MGWR and stratified regression model Cubist were better than global regression models such as SLR and RF,and the main controlling factors of SOC vary greatly in different strata and local sites.(3)Various estimation models combined with environmental variables were used to predict SOC,and the Cubist N9 model that considers the spatial dependence of regression residuals outperformed other models.Then we obtained the SOC map estimated by Cubist N9.A total of 16 estimation models including Kriging models,regression models and regression-Kriging models were used to predict SOC.The results of model evaluation show that the prediction accuracy of ordinary Kriging(OK)model is poor.SLR,RF,GWR and Cubist models are better than OK,and the consideration of residuals estimated by OK further improved the prediction accuracy of these regression models.Cubist N9 model that combined with the mean residual of nine neighboring points outperformed other models in the study area.The SOC map estimated by Cubist N9 exhibited that SOC is low in the middle of the study area,but high in the margin of the study area.Specifically,four high SOC areas are located in the west,northwest and southeast of the study area,and the northwest of the central town,while the three low SOC areas are located in the north,west and southeast of the central town.The spatial variation of SOC was mainly controlled by land use types and multiple cropping index,and also affected by straw returning,spatial distribution of ponds and irrigation canals,elevation and slope.These results indicate that cropping and management factors and landscape metrics are effective environmental variables of farmland SOC in plains.MGWR is an effective local regression model.Cubist model can not only mine the stratified heterogeneity of the relationship between SOC and environmental variables,but also an effective prediction model for farmland SOC in plains.The main control factors of SOC vary greatly in different strata and at different sites.Acquiring these knowledge can help optimize farmland cropping and management methods and landscape pattern,increase soil carbon storage and improve soil fertility.