Study On The Influence Factors And Simulation Models Of Soil Organic Carbon In Alpine Grassland

Soil organic carbon?SOC?is a sensitive indicator for climate change and also an important evaluation index for grassland ecosystem health.Its content significantly affects the productivity of grassland ecosystems.The high carbon density of high altitude ecosystems and their potential response to climate change have attracted widespread attention.Understanding the main environmental factors affecting the SOC of alpine grassland and its mechanism of action,so as to the accurately simulation of the spatial distribution and reserves of SOC in alpine grassland are of great significance for the sustainable management and utilization of grassland resources in alpine and high altitude areas,and even the study on the response of alpine grassland to global climate change.Maduo County of Qinghai Province is located in the hinterland of the Qinghai-Tibet Plateau and the Sanjiangyuan National Nature Reserve.The source of the“Yellow River”of the Huaxia Mother River is sensitive to climate change and fragile.Alpine grassland is not only the main type of land cover in the region,but also the basis for local herders to survive.However,due to the lack of field sampling data and huge spatial heterogeneity,the understanding of SOC at high altitudes regions is not sufficient.There is also a lack of understanding of environmental factors affecting SOC reserves and distribution.In order to understand the mechanism of environmental factors on the SOC of alpine grassland of the study area,to understand the spatial distribution of SOC,and to simulate SOC efficiently and accurately in data-deficient areas,this study used various variables to analyze the relationship and mechanism of action between environmental factors and SOC.Besides,Boruta and Greedy Forward?GF?,Greedy Backward?GB?,Hill Climbing?HC?,Simulated Annealing?SA?and Genetic Algorithm?GA?strategies were employed to construct all relevant variable sets and minimum optimized variable sets based on different data sources.On this basis,machine learning techniques of partial least squares regression?PLSR?,support vector machine?SVM?,Cubist,and random forest?RF?were used to build a parsimonious simulation model.Based on the above research content,this study has the following main conclusions.?1?Conducted by STEP-AWBH which is an environment-soil-landscape model,this study constructed a database of comprehensive environmental variables in the study area.The data pool covers remote sensing,soil spectroscopy and ground-measured data.This database involves soil physical and chemical factors?S?,topographic factors?T?,meteorological factors?A?and biological factors?B?in the STEP-AWBH model.In terms of the influence of environmental variables on SOC,the effect value of soil moisture content is the highest?effect value is 0.683?,followed by total nitrogen?effect value is 0.275?.When distinguishing grassland type,the effect value of water content is the highest in alpine meadow?effect value is 0.788?,while in alpine steppe is total nitrogen?effect value is 0.703?.In terms of the correlation between environmental variables and SOC,among soil physical and chemical factors,soil moisture content has the strongest correlation?r=0.889?;among biological factors,underground biomass has the strongest correlation?r=0.808?.Ground temperature in summer,coverage and aboveground biomass inhibit the accumulation of SOC,while other variables show a promotion effect.In terms of the path environmental variables impacting SOC,meteorological factors mainly directly affect surface vegetation and indirectly affect SOC,while vegetation biomass and soil physical and chemical factors directly affect SOC.?2?The Boruta variable selsction strategy eliminates all topographical factors and meteorological factors.The soil moisture content has the highest factor importance among all relevant variables.According to the variable selection results of the minimum optimization strategies,total nitrogen is the most critical variable for constructing a parsimonious model.In terms of variable data sources,the ratio vegetation index?LsRVI_w?calculated by Landsat 8 data in winter is the most important factor among variables generated by remote sensing data source,and 540-700 nm for the Visible-near infrared spectroscopy data?VNIR?.The key variables for constructing the parsimonious model are the normalized soil index?LsNDSI₈?and 560-660 nm calculated for remote sensing variables and VNIR.?3?In the simulation results based on the selected variables,the combination of the GF variable set constructed using all the environmental variables and the Cubist model gained the best result?R_v²=0.97,RMSE_v=4.72,RPIQ_v=4.52,RPD_v=6.17?.The use of variable selection strategies in this study not only greatly shortens the time–consuming on model calibration,but also improves the prediction accuracy of the machine learning techniques.In terms of the simulation using the variables based on remote sensing data,PLSR has the best performance when using unfiltered variables?R_v²=0.51,RMSE_v=21.76,RPIQ_v=0.98,RPD_v=1.34?,while using the minimum optimized variable set,the combination of GB strategy and Cubist model gained the best result?R_v²=0.54,RMSE_v=19.2,RPIQ_v=1.08,RPD_v=1.48?.In terms of using full-band VNIR,PLSR had the best performance?R_v²=0.92,RMSE_v=8.62,RPIQ_v=2.71,RPD_v=3.42?,and the combination of HC strategy and PLSR model has the best simulation effect when using the minimum optimized bands?R_v²=0.92,RMSE_v=8.63,RPIQ_v=2.71,RPD_v=3.42?.Based on the above results,it can be seen that although the VNIR data alone can obtain satisfactory simulation accuracy,if other soil physical and chemical factors can be added,the simulation accuracy of SOC will be further improved.?4?Although the spatial distributions of SOC content and density of alpine grassland by the combination of ordinary least square and ordinary kriging?OLS-OK?,cubist and ordinary kriging?C-OK?are different,both showed a trend of decreasing from southeast to northwest.The average SOC contents of alpine m in the study area using OLS-OK and C-OK were 20.07 g/kg and 36.9 g/kg,respectively.The SOC contents of alpine meadow were 26.49 g/kg?OLS-OK?and 46.92 g/kg?C-OK?,respectively.And the SOC contents of alpine steppe were 12.28 g/kg?OLS-OK?and 22g/kg?C-OK?,respectively.The average SOC densities were 2.48 kgC/m²?OLS-OK?and 4.05 kgC/m²?C-OK?,respectively.The average SOC densities of alpine meadow were 3.1 kgC/m²?OLS-OK?and 4.67 kgC/m²?C-OK?,respectively.And the average SOC densities of alpine steppe were 1.89 kgC/m²?OLS-OK?and 3.12 kgC/m²?C-OK?,respectively.Based on the two methods,the SOC stocks of alpine grassland in the study area are 1.13×10⁷ TC?OLS-OK?and 1.76×10⁷ TC?C-OK?,respectively.The SOC stocks of alpine meadow are 7.92×10⁶ TC?OLS-OK?and 1.21×10⁷ TC?C-OK?,respectively.The SOC stocks of alpine steppe are 3.33×10⁶ TC?OLS-OK?and 5.55×10⁶ TC?C-OK?,respectively.Guided by STEP-AWBH conceptual model,this study constructs a database of comprehensive environmental variables in the study area based on multiple data sources,which is helpful to comprehensively depict the environment of the study area.These variables also can be important indicators for predicting soil properties in data-deficient areas.In addition,the variable analysis enhances the understanding of the main role of environmental variables in the influence of SOC.On this basis,the use of strategic variable selection methods reduces the redundancy of variables,simplifies the model structure and improves the readability of the simulation models.This study provides reference for the quantitative simulation and digital mapping of soil properties in data-deficient areas,high-altitude and high-altitude areas.It is of great significance for the dynamic monitoring of soil properties of alpine grassland and the scientific management of alpine grassland.