Study On The Driving Mechanism And Prediction Of Soil Organic Carbon Component Storage In Swamp Wetland In Permafrost Area

The Daxing’anling and Xiaoxing’anling in China are densely distributed areas of permafrost wetlands.Against the backdrop of climate warming,the degradation of permafrost in this area has been significant over the past 40 years.The southern boundary of permafrost has been continuously moving northward,the wetland water level has decreased,the plant community structure and productivity have also changed.It is necessary to estimate the existing amount of wetland soil carbon pool and predict future dynamic changes.At present,the academic community’s research on predicting soil carbon storage in swamp wetlands under the background of climate warming mainly focuses on total organic carbon,while there is insufficient research on soil carbon components.The formation,transformation and stabilization mechanisms of different carbon components are different,and their responses to environmental changes are also different.Therefore,it is necessary to study the storage and driving mechanisms of carbon components in order to more accurately predict the sensitivity of soil carbon pools to disturbances such as climate change.In view of this,this study focuses on typical swamp wetlands in the permafrost regions of the Daxing’anling and Xiaoxing’anling.Through on-site sampling data analysis,the vertical and spatial distribution characteristics of total organic carbon(TOC),particulate organic carbon(POC),and Mineral-associated organic carbon(MAOC)were analyzed;An environmental covariate dataset for the STEPAWBH model and structural equation model were constructed to clarify the direct or indirect effects of terrain,climate,biology,and soil properties on TOC,POC,and MAOC.We built a soil characteristic prediction model library(PSO-ML)that integrates particle swarm optimization algorithm(PSO)and machine learning algorithm.PSO-ML can be used to predict the spatial distribution pattern of TOC,POC,and MAOC content and density,and quantitatively evaluate the contribution of environmental covariates to soil characteristic prediction through relative importance scores.The goal is to reveal the driving mechanism of soil organic carbon and its component accumulation in permafrost swamp wetlands and provide theoretical basis and scientific support for the research on the dynamic changes of soil carbon pool in swamp wetlands under the background of climate warming.The main conclusions and progress of this study are as follows:(1)A feature dataset was constructed based on Sentinel-2 multispectral images,Sentinel-1 radar data,and DEM data.Recursive feature elimination algorithm(RFE)was used for feature selection to construct a random forest classification model for feature optimization.The impact of multi-source data fusion(optical images,SAR data,and terrain data)on the classification accuracy of swamp wetlands was explored.The results show that the random forest model with feature optimization has high efficiency and stability,and is suitable for accurate extraction of information from swamp wetlands.There are significant differences in the contribution of different types of features to the classification of swamp wetlands.Spectral features,terrain features,and red edge index have a significant impact on classification accuracy,while radar features and texture features have a smaller contribution.This study provides ideas for constructing a simple and robust classification model,and provides theoretical and technical support for the classification of swamp wetlands based on multi-source data.(2)By analyzing the vertical and spatial distribution characteristics of TOC,POC,and MAOC contents in the four study areas of Mohe,Huzhong,Wuyiling,and Zhanhe,the differences and spatial heterogeneity of soil organic carbon and its carbon components in typical swamp wetlands in permafrost regions were clarified between the three soil layers of 0-10 cm,10-20 cm,and 20-30 cm.The results showed that there was a significant difference in TOC content among the three soil layers(p<0.05),and there was a trend of decreasing with increasing depth.There was no significant difference in TOC content among the same soil layers in different regions.The TOC content values of the three soil layers in each study area all belonged to moderate variation(except for 20-30 cm soil layer in Mohe),and there was a trend of increasing variation with increasing soil depth.The POC content and MAOC content both decrease with the depth of the soil profile,while the coefficient of variation of the three soil layers increases with the increase of depth;The TOC,POC,and MAOC contents of the three soil layers in each study area exhibit strong spatial heterogeneity.(3)The environmental covariate dataset was onstructed in the STEP-AWBH model based on multi-source data.The correlation between TOC,POC,MAOC and environmental covariates were analyzed by calculating Pearson correlation coefficients.Structural equation models were constructed to reveal the driving mechanism of spatial heterogeneity of soil organic carbon and its components in permafrost marsh wetlands.The results showed that the bulk density(BD)and total nitrogen(TN)in soil factors had a significant direct effect on the TOC of the three soil layers;The indirect effect of climate factors on TOC was greater than the direct effect;NPP has a significant direct effect on TOC in 10-20 cm and 20-30 cm soil layers;As the soil layer deepens,the indirect effect of remote sensing variables was higher than the direct effect,and the influence of terrain factors on TOC distribution gradually weakens.Remote sensing variables have a significant impact on the POC content in the 0-10 cm soil layer,but have a weaker impact on the 10-20 cm and 20-30 cm soil layers;The influence of soil factors on POC content in each layer is not significant;NPP has a significant negative effect on the 0-10 cm soil layer,while it has no significant impact on the POC content of other soil layers.Soil factors(BD and TN)have significant direct effects on MAOC.Climate factors mainly affect the accumulation of MAOC through indirect effects.The influence of terrain factors on MAOC gradually weakens as the soil layer deepens.(4)A soil characteristic prediction model library(PSO-ML)integrating PSO and machine learning algorithms was proposed and constructed.Based on the sample dataset and model library,the spatial distribution patterns of soil bulk density,total nitrogen,and total organic carbon were predicted.The contribution of environmental covariates to soil characteristic prediction was analyzed through relative importance scores.The results showed that the selection of the model,the type and quantity of variables had an impact on soil bulk density,total nitrogen and TOC prediction results.In the prediction of soil bulk density,remote sensing variables and soil factors were the main controlling factors.The importance of remote sensing variables weakens with the depth of the soil layer,while the importance of soil factors increases with the depth of the soil profile.In the prediction of total nitrogen,the PSO-XGBoost model outperforms other models in predicting soil total nitrogen at each layer.The contribution of climate to soil total nitrogen prediction was the highest(55.69%).As the soil layer deepens,the impact of climate on total nitrogen gradually weakens.The contribution of soil factors increases with soil depth.The importance scores of remote sensing variables for all three soil layers exceeded 20%.In TOC prediction,the preferred variable types include remote sensing variables,climate factors,terrain factors,soil factors,and terrain factors.In the threelayer soil TOC prediction,soil factors dominate,with total nitrogen being the most important explanatory variable,followed by bulk density.(5)Based on sample data from the Daxing’anling area and the PSO-ML model library,the spatial distribution pattern of POC and MAOC content and density were predicted.The contribution of environmental variables to POC and MAOC prediction were quantitatively analyzed based on importance scores.The transferability of the models were evaluated to provide reference for regional soil carbon pool prediction.The results indicate that the prediction accuracy of POC and MAOC was influenced by model selection,type and quantity of prediction variables.PSO-XGBoost and PSO-XGBRF have stronger ability to predict soil characteristics.In the prediction of POC content,remote sensing variables dominate,and climate factors have a moderate impact on POC prediction(14.62%-24.81%).The contribution of soil factors to the 10-20 cm soil layer was greater than the other two soil layers,and the influence of terrain weakens with the deepening of the soil layer.In the prediction of MAOC content and density,soil factors were the main factor,remote sensing factors were the second,followed by terrain and climate factors.