Spatial Simulation Of Soil Organic Carbon Density Using Environmental Information And Profile Features

The distribution and dynamic change of soil organic carbon(SOC)play an important role in the evaluation of land quality,ecological environment,soil material cycle and global climate change.More accuracy soil organic carbon density(SOCD)simulation is important for the assessment of regional SOC reserves and understanding of ecosystem carbon cycle.However,there is no systematic comparison of methods for spatial simulation of SOCD at different scales;the relationship between depths of different sections,vertical distribution characteristics,depth function and the impact on environmental information need to be discussed;there is few studies for whether the data transformation methods and compositional data analysis can improve the prediction accuracy.Therefore,this study combines the above issues for the spatial simulation of SOCD,Sanjiang Plain in Northeast China was selected as the research area,and the measured data of SOCD was used to carry out descriptive statistical analysis,horizontal and vertical distribution characteristics of SOCD in different profile depths of the region;and the exponential and spline function were used as depth function,generalized linear model(GLM)and random forest model(RF)were selected to verify the prediction accuracy,spatial simulation,uncertainty analysis,and storage of SOCD.Two new methods were proposed to improve the accuracy of spatial simulation of SOCD based on the method of isometric log-ratio transformation and land use type.The results showed that(1)the SOCD in Sanjiang Plain was moderate variation;the percentage of SOC data could reduce the variability and make the data follow the normal distribution;the SOCD decreased with the increase of depth in most land use and soil types,and their difference decreased gradually.The vertical distribution of SOCD can be divided into four curve types by combining the index function and its threshold value.The percentage data in the simplex spatial distribution shows that the data distribution of forest land and dry land is more concentrated than that of paddy field and wetland.(2)The importance of environmental variables to SOCD is different in different profile depth.Land use,topography,soil particle size and climate are the dominant factors in the surface layer,middle layer and bottom layer,respectively.Due to the lack of flexibility of exponential function and the diversity of curve types,the spline function was better,but spline function had the disadvantage of over fitting.Generally speaking,the strength of surface control ability leads to the difference of fitting accuracy of exponential function.(3)For prediction,the exponential function method was overestimated and its accuracy was poor.Although the spline function method was better than the independent method,the improvement effect was not obvious.The two new methods proposed in this study considered the proportion distribution of each depth,and the accuracy and interpretability of the model were better than other models.Compared with the prediction generated by RF,GLM has better skewness and worse accuracy,and its prediction was more aggregated,which does not conform to the distribution form of the original data.Therefore,it is recommended to use the RF model,especially in the areas with higher altitude.GLM is more suitable for predicting small areas of data variability.Based on the prediction and statistics of the organic carbon reserves,and the related research before Sanjiang Plain,the soil organic carbon reserves of each layer are consistent in the order of magnitude,and this study has advantages in the introduction of depth information,the selection of environmental variables,and the optimization of model methods compared with the previous research,which can provide reference for the large-scale SOCD spatial simulation research combining environmental information and profile depth functions.