Spatial distributions, driving factors, and future changes of soil organic carbon in China: arid regions vs. humid regions Authors: Bin Yang, Shihang Zhang, Xiaoguo Wang Published: 2025 ABSTRACT AbstractSoil carbon sequestration is of great significance for achieving China’s 2060 carbon neutrality goal. However, differences in carbon sequestration between arid and humid regions remain unclear. Here, based on the Chinese terrestrial ecosystems carbon density dataset, this study employed the random forest (RF) model to map soil organic carbon density (SOCD) (1 km × 1 km) in arid and humid regions, and assessed the spatial uncertainty. The results indicated: (1) The RF model can explain about 79%-88% of SOCD variation in 2020. (2) The SOCD in arid regions (4.03 kg C m−2 in 0–20 cm soil depth, and 9.84 kg C m−2 in 0–100 cm), were significantly lower than those in humid regions (5.54 kg C m−2, and 12.91 kg C m−2), and had greater spatial heterogeneity. (3) In arid regions, SOCD was mainly driven by mean annual precipitation (MAP), mean annual temperature (MAT), normalized difference vegetation index (NDVI), soil moisture (SM), and land use and land cover (LULC). In humid regions, it was driven by MAT, MAP, elevation, human footprint, and LULC. (4) SOCD exhibited a decreasing trend in arid regions, but exhibited an increasing trend in humid regions under different climate scenarios. The results of this study are of great significance for exploring the response of the SOC pool to climate change in arid and humid regions, identifying vulnerable areas and carbon sinks, and scientifically formulating regional carbon management policies to achieve the goal of carbon neutrality.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-32482-0.Keywords: Soil organic carbon, Spatial distribution, Driving factors, Digital soil mapping, Arid region, Humid regionSubject terms: Climate sciences, Ecology, Ecology, Environmental sciences Introduction Soil organic carbon (SOC) represents the largest carbon reservoir in terrestrial ecosystems1–3, storing approximately 2047 Pg (1 Pg = 10¹⁵ g) of carbon4, which is equivalent to two to three times the atmospheric carbon stock5,6. Numerous studies have indicated that even minor changes in SOC stocks can significantly alter atmospheric carbon dioxide (CO₂) concentrations3, leading to substantial fluctuations in the global carbon balance and potentially driving climate change7. In recent years, the carbon sequestration capacity of different terrestrial ecosystem types and their variations have attracted widespread attention, with efforts made to uncover the underlying driving mechanisms8. For instance, extensive research has revealed SOC sequestration patterns in forests9,10, wetlands11,12, and grasslands13, among others. These studies have encompassed most terrestrial ecosystems and indicate that SOC dynamics are influenced by climate14–16, land use and land cover changes (LULCC)17,18, soil physicochemical conditions19, and hydrological conditions20. Concurrently, numerous studies have highlighted that the balance between carbon inputs and decomposition outputs is regulated differently in humid versus arid environments. For example, in humid climates, both carbon input and decomposition output increase with rising temperatures, but the increase in output is more pronounced21–23. In contrast, in arid climates, precipitation constrains both plant production and decomposition, with a stronger limitation on plant production24–26. Although the role of climate in terrestrial carbon cycles has been extensively discussed, the heterogeneity of SOC between arid and humid regions remains inadequately characterized. Moreover, future climate warming and changes in precipitation patterns are expected to exacerbate the asymmetry in SOC stocks between these regions, yet current research has not quantitatively assessed this disparity. Thus, elucidating the differences in SOC between arid and humid regions is crucial for optimizing regional carbon management strategies and enhancing soil carbon sequestration efforts. It is well established that SOC stocks depend on the dynamic balance between inputs via plant residues, root exudates, and external materials, and outputs via decomposition, erosion, and hydrological leaching2,27. The spatiotemporal variation of SOC is profoundly influenced by complex interactions among multiple factors, including climate, biota, and soil conditions28–30. For example, Jenny (1941)31 and Carvalhais et al. (2014)32 emphasized that climate, particularly temperature and precipitation, serves as a key controller of SOC turnover and resource distribution, regulating sequestration through processes such as plant productivity and microbial activity. Thus, in ecosystems under precipitation or temperature stress, the carbon cycle exhibits high sensitivity to climate change. Accounting for this asymmetric climate response is essential for accurately estimating future carbon stocks. For nearly a century, soil microorganisms have been primarily regarded as decomposers of SOC, releasing carbon back to the atmosphere as CO₂33,34. However, recent advances, including the concept of the soil microbial carbon pump, have demonstrated that microorganisms also play an active role in SOC stabilization35–37. Soil environment further modulates SOC dynamics, with physical protection mechanisms such as clay adsorption and aggregation being critical for SOC stabilization38–40. Additionally, soil physicochemical properties, such as bulk density (BD), moisture (SM), temperature, and pH, govern microbial processes involved in the carbon cycle41. Moreover, profound changes in land use and land cover (LULC), driven by both human activities and natural processes, can lead to significant restructuring of SOC spatial distribution. Previous studies have documented substantial SOC variations under different LULC conditions18. A global meta-analysis on LULCC further indicated that conversions of forests and grasslands to croplands, as well as forests to grasslands, generally reduce SOC, whereas reverse processes tend to increase it17. Based on the aridity index (AI, the ratio of precipitation to potential evapotranspiration), terrestrial ecosystems are classified into arid (AI < 0.65) and humid (AI ≥ 0.65) regions42. Previous studies have pointed out significant differences in SOC dynamics between these regions43. In arid regions, insufficient water supply greatly limits primary productivity44, thereby reducing plant-derived carbon inputs45,46. A meta-analysis on precipitation effects indicated that increased precipitation in arid regions has a stronger positive effect on carbon sequestration than in humid regions47. Water limitation also indirectly influences carbon dynamics by suppressing microbial activity45. In contrast, humid regions benefit from abundant water and heat, supporting high net primary productivity (NPP) and substantial carbon inputs. However, warming in these regions can accelerate carbon loss due to enhanced microbial decomposition48. High soil moisture in humid regions may also limit oxygen diffusion45 and lower soil pH, both of which can suppress microbial activity and thus reduce SOC decomposition49. Future climate warming and precipitation changes are projected to further amplify SOC asymmetry between arid and humid regions, though this disparity remains unquantified. Numerous methods have been developed for mapping soil attributes, among which digital soil mapping (DSM) is widely adopted for its efficiency and accuracy in representing the spatial distribution of soil properties50. Based on the theory of soil-forming factors and the first law of geography50, DSM establishes quantitative relationships between soil properties and environmental covariates to derive spatiotemporal patterns of soil attributes15. For example, Luo et al. (2020)51 applied five interpolation methods to map the spatial distribution of SOC across China. In recent years, machine learning approaches have gained widespread recognition and application in DSM52. These data-driven models rely on soil-environment relationships, particularly the interactive effects of environmental covariates such as climate, topography, and vegetation on SOC53–55. For instance, Yang et al. (2016)56 used Random Forest (RF) and Boosted Regression Trees (BRT) to map SOC content on the northeastern edge of the Tibetan Plateau. Li et al.57 used RF models to accurately predict the spatial distribution of SOC, while Keskin et al. (2019)58 employed machine learning to analyze SOC distribution under varying land use, soil, and climate gradients. Data-driven models have proven effective in predicting the spatiotemporal distribution of SOC based on diverse environmental covariates16,59,60, and have been successfully applied at regional8 and national scales61. Numerous studies confirm that SOC dynamics are highly sensitive to climate change14–16. For example, mean annual temperature (MAT) and mean annual precipitation (MAP) influence the input, output, and turnover of SOC among different pools9,14,41,62. In the 21st century, temperature and precipitation are projected to undergo dramatic changes63, with asymmetric precipitation shifts between arid and humid regions likely leading to asymmetric SOC responses47. Therefore, revealing SOC changes in arid and humid regions and projecting future trends are of great significance for guiding regional carbon management and exploring carbon sequestration potential. Future climate warming and changes in precipitation patterns are expected to further exacerbate SOC asymmetry between arid and humid regions, yet current research has not quantified this disparity. This study utilizes the Chinese terrestrial ecosystem carbon density dataset (2010s) compiled by Xu et al. (2019)64, which includes 4536 topsoil (0–20 cm) and 3147 profile (0–100 cm) SOC samples, thereby providing a robust basis for our analysis. The objectives of this study are: (1) to reveal differences in the spatial distribution patterns of SOC between arid and humid regions; (2) to identify the driving factors underlying these spatial patterns; and (3) to project and quantify the future changes of SOC under different climate scenarios. By elucidating the distinct regulatory mechanisms of SOC in arid and humid regions of China, this study aims to provide a scientific basis for optimizing regional soil carbon management and formulating climate adaptation strategies. Materials and methods Data collection SOC density dataset This study focuses on SOC dynamics at two depths: 0–20 cm and 0–100 cm. The former represents the plow layer most intensely affected by environmental changes and management practices, holding practical management significance. The latter depth standard is most commonly used to assess total ecosystem carbon storage. They are both crucial for understanding regional carbon cycling. The SOCD (kg C m−2) (Soil organic carbon density, the total mass of SOC contained within a unit area of soil at a specific soil depth. This indicator facilitates the estimation of carbon stocks.) data are derived from a publicly available SOC dataset64, which collected 4536 samples from the 0–20 cm soil layer and 3147 samples from the 0–100 cm soil layer in China between 2004 and 2014 (Fig. 1). Each sample recorded the geospatial location of the sampling point (latitude, longitude, and elevation), climate (MAT, MAP), soil physicochemical properties (clay content, sand content, soil moisture (SM), and soil pH), normalized difference vegetation index (NDVI), land use and land cover (LULC), and human footprint (HFP). Missing environmental data were obtained from the following sources: climate (http://www.gis5g.com/), elevation (https://www.ncdc.ac.cn/portal/?lang=zh), soil physicochemical properties (https://www.ncdc.ac.cn/portal/?lang=zh), LULC(http://www.gis5g.com/), NDVI (http://www.gis5g.com/), and HFP (https://data.casearth.cn/?tdsourcetag=s_pcqq_aiomsg). And missing values were imputed based on latitude and longitude information. Spatial distribution maps of arid and humid regions and sampling points in China. (a) The spatial distribution of arid and humid regions, where yellow regions represent arid regions and green regions represent humid regions. The orange dot indicates the location of the sampling points (0–20 cm) in arid regions, and the purple pentagon indicates the location of the sampling points (0–100 cm) in arid regions; The green dot indicates the location of the sampling points (0–20 cm) in humid regions, and the red pentagon indicates the location of the sampling points (0–100 cm) in humid regions. (b) The inset shows the frequency statistics of the SOCD of sample points, where the red vertical line represents the mean value of the sample points. Environmental covariates DSM can simulate the spatial distribution patterns of SOC by establishing relationships between environmental variables and soil properties65. In this study, a set of environmental covariates raster data was used to predict the spatial distribution of SOC, including MAT, MAP, elevation, clay, sand, pH, SM, HFP, LULC, and NDVI (Table 1). To ensure spatial consistency, all raster layers used the WGS84 coordinate system with a spatial resolution of approximately 0.0083333° (~ 1 km). Table 1Geospatial data for predicting the Spatial distribution of SOC in arid and humid regions of China.TypeVariableUnitScaleResourceaSoil propertiesClay%1 kmNCDC (https://www.ncdc.ac.cn/portal/?lang=zh)Sand%1 kmNCDC (https://www.ncdc.ac.cn/portal/?lang=zh)Soil Moisture%1 kmNCDC (https://www.ncdc.ac.cn/portal/?lang=zh)pH–1 kmNCDC (https://www.ncdc.ac.cn/portal/?lang=zh)ElevationElevationm1 kmNCDC (https://www.ncdc.ac.cn/portal/?lang=zh)ClimateMAT°C1 kmGRDC (http://www.gis5g.com/)MAPmm1 kmGRDC (http://www.gis5g.com/)Land use and land coverLULC–30 mGRDC (http://www.gis5g.com/)Normalized difference vegetation indexNDVI–30 mGRDC (http://www.gis5g.com/)Human footprintHFP–1 kmDSSP (https://data.casearth.cn/?tdsourcetag=s_pcqq_aiomsg)Arid indexAI–1 kmGlobal-AI_PET_v2 (https://cgiarcsi.community/) Geospatial data for predicting the Spatial distribution of SOC in arid and humid regions of China. Future climate scenarios dataset To reveal the impact of climate change on SOC dynamics, this study used future climate downscaled temperature and precipitation data generated by the EC-Earth3-Veg model66, which was developed based on the Coupled Model Intercomparison Project Phase 6 (CMIP6), and is available in the WorldClim database (https://worldclim.org/). Under different shared socio-economic pathways (SSPs) and climate trajectories, the following three representative scenarios were used: SSP119 (SSP1, sustainable development pathway; RCP1.9, low emissions), SSP245 (SSP2, intermediate pathway; RCP4.5, medium emissions), and SSP585 (SSP5, fossil fuel development pathway; RCP8.5, high emissions). For analysis at the same spatial scale, all raster layers used the WGS84 coordinate system, with a spatial resolution of approximately 0.0083333° (~ 1 km). Data analysis and modeling In this study, we compared different machine learning models to predict the spatial distribution of SOC. Firstiy, arid and humid regions are classified based on the aridity index (AI)42. Subsequently, multicollinearity was tested using the variance inflation factor (VIF), and variables with VIF values greater than 5 were excluded67. As a result, there was no severe multicollinearity in MAT, MAP, Elevation, LULC, Clay, Sand, pH, SM, NDVI, and HFP (Table S1). Next, different machine learning models were constructed and evaluated, and the optimal model was selected for mapping the distribution of SOC. In addition, path analysis (PLS-PM) was used to reveal the influence mechanism of the driving factors. Finally, we predicted the SOC changes in arid and humid regions under different future climate scenarios. The SOC estimation process is shown in Fig. 2. SOC estimation flowchart. RF, Random forest model; BRT, Boosted regression trees; R2, R squared; RMSE, Root mean square error; MAE, Mean absolute error; PLS-PM, Partial least squares path modeling; GOF, Goodness of fit; SSP119, low emissions, sustainable development pathways 1-1.9; SSP245, medium emissions, intermediate pathways, shared socio-economic pathways 2-4.5; SSP585, fossil fuel pathways, shared socio-economic pathways 5-8.5. Spatial distribution prediction of SOC Machine learning has been extensively studied in predicting the spatial distribution of SOC. The Random Forest (RF) model consists of a collection of random regression trees, which constructs numerous regression trees and aggregates their results to produce the final prediction. During training, each random regression tree is constructed from a random subset of the training set’s original observations, selected via bootstrapping. In the RF modeling process, three training parameters require user definition: (i) The number of trees in the forest (ntree), which affects the model’s robustness and prediction variance. The more trees used, the more robust the model becomes, but computational demands increase accordingly. Beyond a certain number of trees, model performance levels off. We start with ntree = 500 and incrementally increase the number of trees by 100 until reaching 2000 trees, selecting the optimal ntree value that balances overall performance and computational efficiency. (ii) The number of features randomly selected at each node (mtry), where the mtry parameter increases the randomness among different random trees. Increasing mtry enhances tree strength and reduces inter-tree correlation, potentially increasing overfitting risk. Decreasing mtry reduces tree strength and increases inter-tree correlation, effectively improving model generalization capability68. Splitting trees based on the mtry parameter effectively reduces the correlation between tree outputs, thereby decreasing the variance of their average values69. We set the initial input value for mtry to one-third of the number of features selected by the model, with the mtry test value range defined as [p/2, p/4]. (iii) The minimum number of terminal nodes (nodesize) controls the tree depth, affecting model complexity and generalization capability. For regression problems, the default input value is 5. We set gradient values [1, 3, 5, 7, 10] as inputs. The optimal input parameters were selected as model parameters based on the results of 10-fold cross-validation. The final hyperparameter combination is: mtry = 5, ntree = 500, nodesize = 7. Boosted Regression Trees (BRT) models combine the characteristics of gradient boosting and regression trees to enhance the predictive performance of a single model. During the modeling process, four parameters can be defined by the user: learning rate (shrinkage), tree complexity (interaction.depth), number of trees (n_tree), and bagging score (n.minobsinnode). The parameter “shrinkage” represents each tree’s contribution to the final fitted model. Lower values indicate better model generalization. Typically start debugging from 0.01, with a search range set to [0.005, 0.01, 0.02, 0.05, 0.1]. The parameter “interaction.depth” controls tree size and feature correlations, with a gradient search set to [1, 3, 5, 7, 9] for grid search. The parameter “n_tree” is determined by the combination of shrinkage and interaction.depth. The parameter “n.minobsinnode” controls the proportion of samples used for training each tree, enhancing model robustness, with search range [0.7, 0.75, 0.8]. We tested several combinations of parameter values (shrinkage, interaction.depth, n_tree, and n.minobsinnode). The optimal input parameters were selected as model parameters based on 10-fold cross-validation results. The final parameter combination is: n_tree = 1200, interaction.depth = 9, n.minobsinnode = 0.75, and shrinkage = 0.02. Support Vector Machine (SVM) is a nonlinear model based on statistical learning theory70. Due to its applicability and robustness, SVM is also widely used in soil science. When employing the SVM algorithm for SOC prediction, we selected the radial basis function (RBF) kernel, tested different values of the parameter “cost”, and determined the optimal model parameters that minimize the root mean square error (RMSE). The parameter “cost” was set to 1 as the model hyperparameter input. The model performance was evaluated using a 10-fold cross-validation procedure that involved comparisons between the predicted and observed SOC values. Three validation measurements were calculated: coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). The relative importance of features (%IncMSE) was estimated using the built-in function of the RF model. The higher the %IncMSE, the more important the feature71,72. Finally, the model with the best performance was selected for mapping. The calculation formula is as follows: where Pi and Oi are the predicted and observed SOC values at site i, respectively; n is the number of samples. MAE measures the average prediction bias, and RMSE represents the overall quality of the prediction. Predictions become increasingly optimal as MAE and RMSE approach zero. where is the j-th feature, refers to the mean squared error of the model after the feature has been randomly permuted, is the mean square error of the model when the feature is not displaced. The spatial uncertainty of SOCD at depths of 0–20 cm and 0–100 cm in arid and humid regions was estimated using the 95th, 50th, and 5th percentiles of multiple outputs from the optimal prediction model. The calculation equation was as follows: where UNC represents the uncertainty index; P95, P50, and P5 represent 95%, 50%, and 5% quantiles of the predicted results, respectively. Soil organic carbon density (SOCD) represents the amount of SOC per unit area at a specific soil depth, while soil organic carbon stock (SOCS) indicates the total amount of SOC at a given depth across a country or region. Both are subject to dynamic changes. By quantifying variations in SOCD and SOCS (SOCDChange and SOCSChange), the spatiotemporal dynamics of a country’s or region’s soil organic carbon can be clearly observed. These changes can be calculated using the following formula: where SOCDChange represents the change of SOC density, SOCDFuture represents SOCD under the future climate scenario, SOCDNow represents the current SOCD, SOCS represents the SOC stock, S represents the area, SOCSChange represents the change of SOC stock, SOCSFuture represents the SOCS under the future climate scenario, and SOCSNow represents the current SOCS. Future predictions Based on EC-Earth3-Veg model data from CMIP6, climate conditions (average temperature and average precipitation) in 2060 and 2100 were simulated under the SSP119 (sustainable pathway), SSP245 (intermediate pathway), and SSP585 (fossil fuel pathway) scenarios. We utilized pre-trained machine learning models to predict spatial distribution patterns of SOC under different climate scenarios. The core objective of this study’s future projections is to assess the effects of climate change on SOC dynamics. Therefore, we established a baseline assumption: all environmental covariates (such as NDVI, LULC, and soil properties) remain unchanged at their current (approximately 2020) state, while climate variables are derived from CMIP6 models. This provides a baseline reference for evaluating climate-driven SOC dynamics. However, this assumption undeniably carries limitations. Therefore, we utilized future LULC projections for 2100 under the SSP245 scenario (https://worldclim.org/) to drive our trained optimal model, generating a new SOCD prediction map. This new map was then compared with the predictions from the baseline scenario (where LULC remains unchanged) (Fig. S14, Table S2). Statistical analysis In this study, Pearson correlation analysis was used to assess the linear correlation between all environmental covariates and SOC (Fig. S1), the Pearson correlation coefficient (r) reflected the correlation between all environmental covariates and SOC. Variance inflation factor (VIF) tests (VIF < 5) were performed for each covariate to eliminate multicollinearity when constructing the model. The results showed that the 10 environmental covariates (MAT, MAP, Elevation, LULC, pH, Sand content, Clay content, SM, NDVI, and HFP) did not exhibit severe multicollinearity (Table S1). Subsequently, each environmental covariate was fitted with SOCD to explore linear relationships (Fig. S2 and S3). Group analysis was performed based on environmental variable gradients (Fig. S4), and the mean differences in SOCD between different groups were compared using one-way analysis of variance (ANOVA) combined with Tukey’s HSD test (α = 0.05). The PLS-PM was used to study the relationship between observed variables and latent variables in multivariate interaction systems73, and can effectively distinguish between direct and indirect effects73,74. This study used the “plspm” package in R to construct a PLS-PM model, dividing environmental variables into six latent variables (climate (MAP and MAT), soil (pH, clay, sand, SM), elevation, LULC, NDVI, and HFP). The significance of path coefficients was calculated using the bootstrap method (1000 iterations), and a goodness-of-fit index (GOF > 0.3) was considered acceptable. Results Prediction performances of the models The predictive performance of the RF, BRT, and SVM models is shown below (Table 2; Fig. 3). In general, the R2 range of the RF model ranged from 0.79 to 0.88, which was higher than the R2 of the BRT and SVM models, indicating that the RF model can reveal about 79–88% of SOCD variation. The ranges of RMSE and MAE of RF were 0.35–0.45 kg C m− 2 and 0.22–0.28 kg C m− 2, respectively, which were also lower than those of the BRT and SVM models, indicating that the deviation of the RF model was smaller than that of the BRT and SVM models. Therefore, the RF model had the best prediction performance. Table 2RF, BRT, and SVM models predict SOC performance evaluation (Unit: kg C m-2). “China” is a combination of “arid” and “humid” regions.Depth(cm)RegionModel R 2 RMSEMAE0–20AridRF0.880.350.22BRT0.800.450.32SVM0.650.590.38HumidRF0.800.450.28BRT0.690.560.39SVM0.460.740.47ChinaRF0.850.390.24BRT0.710.530.38SVM0.550.670.430-100AridRF0.850.390.23BRT0.670.580.39SVM0.500.710.41HumidRF0.790.450.29BRT0.720.530.38SVM0.270.850.49ChinaRF0.840.400.24BRT0.700.550.38SVM0.400.770.45 RF, BRT, and SVM models predict SOC performance evaluation (Unit: kg C m-2). “China” is a combination of “arid” and “humid” regions. SOCD of predictions and observations. In the top 0–20 cm soil in the (a) arid region, (b) humid region, and (c) in China, and in 0–100 cm in the (d) arid region, (e) humid region, and (f) in China. “Total” is a combination of “arid” and “humid” regions.The red dotted line represents the 1:1 line, and the green dotted line represents the regression line between the predicted and observed SOCD. Descriptive statistics of SOCD The descriptive statistics for soil organic carbon are shown below (Table 3; Fig. 4). In general, the SOCD were 3.67 kg C m− 2 in 0–20 cm and 8.23 kg C m− 2 in 0-100 cm in arid regions, which were significantly lower than that of SOCD in humid regions (5.14 kg C m− 2 and 11.44 kg C m− 2, respectively). The coefficient of variation (C.V.) for all SOCD values was moderate (0.1 < C.V. < 1). In humid regions, the C.V. of SOCD was 0.65 in 0–20 cm and 0.59 in 0–100 cm, which were lower than the values of 0.82 and 0.78 in arid regions, respectively. Therefore, the variation degree of SOCD in humid regions was smaller. The SOCD in 0–20 cm and 0–100 cm showed a positive skew distribution in arid and humid regions (Skewness > 0). Except for the SOCD in 0–20 cm in arid regions (Excess < 0), SOCD showed a peak distribution in other conditions. Table 3Descriptive statistics of SOCD (kg C m− 2) at the depths of 0–20 cm and 0–100 cm in arid and humid regions. Min., Max., S.D., and C.V. Represent the minimum value, maximum value, standard deviation, and coefficient of variation, respectively.Depth(cm)AI N Min.Max.MeanS.D.SkewnessExcessC.V.0–20< 0.03130.321.280.860.23-0.46-1.810.270.03–0.22550.1712.042.221.812.072.460.820.2–0.518560.2214.053.362.671.34-1.250.800.5–0.654940.4119.555.333.761.18-2.150.71< 0.6526520.1719.553.673.031.49-0.680.82≥ 0.6518360.6723.755.143.341.821.820.650-100< 0.03111.585.103.041.250.43-4.440.410.03–0.22520.3934.125.254.353.1611.980.830.2–0.513960.4857.908.417.282.586.800.870.5–0.653170.5774.6811.508.093.4316.360.70< 0.6519360.3974.688.236.382.409.330.78≥ 0.659790.9673.3011.446.752.8113.140.59 Descriptive statistics of SOCD (kg C m− 2) at the depths of 0–20 cm and 0–100 cm in arid and humid regions. Min., Max., S.D., and C.V. Represent the minimum value, maximum value, standard deviation, and coefficient of variation, respectively. SOCD in arid and humid regions in the top 0–20 cm soil and 0–100 cm. (a) Arid (AI < 0.65) and humid (AI ≥ 0.65). (b) Extreme arid (AI < 0.03), arid (0.03 ≤ AI < 0.2), semi-arid (0.2 ≤ AI < 0.5), semi-arid and semi-humid (0.5 ≤ AI < 0.65), and humid (AI ≥ 0.65). The middle horizontal line of the box represents the average value, the upper and lower horizontal lines represent the third (Q3) and the first quartile (Q1), respectively, the height of the box represents the interquartile range (IQR, Q3- Q1), the upper whisker line represents Q3 + 1.5IQR, and the lower whisker line represents Q1-1.5IQR. The right half of the violin plot on the right shows the densities of the data point distribution, the wider the width, the denser the data point distribution. Different uppercase and lowercase letters indicate significant differences, with significance α = 0.05. Spatial distribution of SOC Spatial distribution and uncertainty of SOCD This study mapped the spatial distribution of soil organic carbon in arid and humid regions (Fig. 5). Overall, the SOCD in 0–20 cm and 0–100 cm in arid regions had similar distribution patterns, and the same was true in humid regions. In general, in arid regions, the spatial distribution pattern predictions were high in the northern part of Northeast China, the southeastern part of the Tibetan Plateau, and the Tarim Basin. However, the low-level SOCD were distributed in the Loess Plateau, Northwest China, the central and western parts of Inner Mongolia, the North China Plain, the northwestern part of the Tibetan Plateau, and certain areas of Xinjiang. In humid regions, the spatial distribution pattern predictions were high in Northeast China, the southeastern part of the Tibetan Plateau, and Southwest China. Yet, other regions had relatively low SOCD levels. The predicted SOCD were 4.04 kg C m− 2 (in 0–20 cm) and 9.84 kg C m− 2 (in 0–100 cm) on average in arid regions, and 5.54 kg C m− 2 (in 0–20 cm) 12.91 kg C m− 2 (in 0–100 cm) on average in humid regions. Spatial distribution of soil organic carbon density (SOCD) in the top 0–20 cm soil and 0–100 cm in arid and humid regions based on the random forest (RF) model. SOCD in 0–20 cm in arid (a), humid (c) regions, and China (e). And SOCD in 0–100 cm in arid (b), humid (d) regions, and China (f). “China” is a combination of “arid” and “humid” regions. The uncertainty in predicting the spatial distribution of soil organic carbon is shown below (Fig. 6). Overall, the uncertainty values were high in 0–20 cm and 0–100 cm in arid and humid regions (UNC > 0.33). In arid regions, UNC was 1.12 on average, which was similar to that (UNC = 1.13) in humid regions in 0–20 cm, but the uncertainty of the SOCD predictions in arid regions (UNC = 2.38) was higher than that in humid regions (UNC = 1.26) in 0–100 cm, indicating that the uncertainty of the SOCD predictions of the top 0–100 cm soil in arid regions was higher. In general, in arid regions, the SOCD prediction uncertainty increased with the soil depth from 0 to 20 cm to 0–100 cm. The high uncertainties were mainly in the Tibetan Plateau and Xinjiang. In humid regions, the prediction uncertainty distributions of SOCD predictions in 0–20 cm and 0–100 cm were similar, and were high in Southwest China and Taiwan province. Uncertainty distribution of SOCD (UNC SOCD) prediction in 0–20 cm and 0–100 cm in arid and humid regions based on the Random Forest (RF) model. SOCD in 0–20 cm in arid (a), humid (c) regions, and China (e). And SOCD in 0–100 cm in arid (b), humid (d) regions, and China (f). “China” is a combination of “arid” and “humid” regions. Differences of SOC under different land use types The statistics for soil organic carbon under different land use types are shown below (Fig. 7). Overall, the total SOCS in arid and humid regions were 25.83 Pg C and 17.33 Pg C in 0–20 cm (Fig. 7c), respectively. And the total SOCS in arid and humid regions were 63.03 Pg C and 40.35 Pg C in 0–100 cm (Fig. 7d), respectively. In terms of average predicted SOCD for different land use types (Fig. 7a, b), wetland had the highest total SOCD of the two regions, about 7.04 kg C m− 2 in 0–20 cm and 17.94 kg C m− 2 in 0-100 cm, respectively. Cropland had the lowest total SOCD, about 4.02 kg C m− 2 in 0–20 cm and 10.30 kg C m− 2 in 0-100 cm, respectively. The statistical results for the two regions respectively indicated that, in arid regions, forest had the highest predicted SOCD (6.81 kg C m− 2), followed by wetland (5.70 kg C m− 2), and the lowest was other land use types (3.45 kg C m− 2) in 0–20 cm. In the top 0–100 cm soil, wetland had the highest predicted SOCD (15.26 kg C m− 2), followed by forest (14.57 kg C m− 2), and the lowest was grassland (8.65 kg C m− 2). In humid regions, wetland had the highest predicted SOCD (9.77 kg C m− 2 in 0–20 cm, and 23.41 kg C m− 2 in 0-100 cm, respectively.), followed by grassland (8.45 kg C m− 2 and 21.57 kg C m− 2), and the lowest were cropland (4.65 kg C m− 2 and 11.29 kg C m− 2). In terms of total SOCS for different land use types, grassland had the highest total SOCS (9.99 Pg C in 0–20 cm and 22.26 Pg C in 0–100 cm, respectively.) in arid regions, while forest had the highest total SOCS (10.09 Pg C in 0–20 cm and 22.19 Pg C in 0–100 cm, respectively.) in humid regions. Average SOCD and SOCS in the top 0–20 cm soil and 0–100 cm of main land use types in arid and humid regions. SOCD in the top 0–20 cm (a) and 0–100 cm (b), and SOCS in the top 0–20 cm (c) and 0–100 cm (d). The bars represent SOCD and SOCS, respectively. The lines on the bars represent standard errors. Orange, green, and blue represent statistical data for arid regions, humid regions, and the entire Chinese territory, respectively. A t-test (p < 0.01) was conducted to determine whether significant differences existed in the mean SOCD between arid and humid regions under different land use types (a and b). “Total” is a combination of “arid” and “humid” regions. Driving factors of SOCD change The relative importance of environmental covariates The relative importance of environmental covariates was assessed using random forest and multiple linear regression models (Figs. 8 and 9). Overall, the climatic factors and soil physicochemical properties were the most important environmental covariates affecting the spatial distribution of SOCD in both arid and humid regions. In general, MAP, MAT, NDVI, and soil physicochemical properties were the most important environmental covariates in the arid regions. However, elevation, MAP, HFP, and soil physicochemical properties were the most important environmental covariates in humid regions. Relative importance of the environmental covariates of SOCD prediction models based on the RF model in 0–20 cm and 0–100 cm in arid and humid regions of China, and China. In the top 0–20 cm soil, in (a) arid, (b) humid regions, and (c) China, and in 0–100 cm, in (d) arid, (e) humid regions, and (f) China. *p < 0.05, **p < 0.01, ***p < 0.001. “Total” is a combination of “arid” and “humid” regions. Relative importance of the environmental covariates of SOCD prediction models based on the multiple linear regression model (MLR) in 0–20 cm and 0–100 cm in arid and humid regions of China, and China (scatter plot), and contribution proportion of environmental covariates (pie chart). In the top 0–20 cm soil, in (a) arid, (b) humid regions, and (c) China, and in 0–100 cm, in (d) arid, (e) humid regions, and (f) China. *p < 0.05, **p < 0.01, ***p < 0.001. Driving pathway analysis of SOCD change Figure 10 plots the PLS-PM analysis revealed the effects of the environmental covariates of SOCD change through direct or indirect pathways. Overall, the PLS-PM model explained 43% and 30% of the SOCD variation in 0–20 cm and 0–100 cm in arid regions, respectively, and 23% and 12% of the SOCD variation in humid regions, respectively. In general, in arid regions, soil physicochemical properties (i.e., Clay, Sand, SM, and pH), land use and land cover (LULC), and NDVI were significantly and positively associated with SOCD (Fig. 10a, b). In the top 0–20 cm soil, climate (i.e., MAT and MAP) was significantly and positively associated with SOCD, and at the same time, NDVI and soil physicochemical properties were positively associated with SOCD via positive association with climate. However, climate was significantly and negatively associated with SOCD in 0–100 cm. In humid regions, climate was significantly and negatively associated with SOCD, while positively correlated with LULC and elevation. In the top 0–20 cm soil, NDVI and soil physicochemical properties had no significant association with SOCD, while the HFP was significantly and negatively associated with SOCD. In 0–100 cm, soil physicochemical properties were significantly and negatively associated with SOCD. Partial least squares path modeling. Pathway analysis in arid regions in (a) 0–20 cm and (b) 0–100 cm, and in humid regions in (c) 0–20 cm and (d) 0–100 cm. The arrows in the path indicate the direction of influence, the thickness indicates the magnitude of the path load, and the dotted line indicates the indirect influence. Red indicates positive impact, and blue indicates negative. The trends of SOC under future climate scenarios The trends in SOC under different climate scenarios are shown in Fig. 11. In arid regions, SOCD exhibited a decreasing trend under all other scenarios, except for an increasing trend (0.75 kg C ha⁻¹ yr⁻¹ in 0–20 cm, and 19.40 kg C ha⁻¹ yr⁻¹ in 0–100 cm) under the low-emission scenario SSP119 by 2060. However, in humid regions, SOCD exhibited an increasing trend under all other scenarios, except for a decreasing trend (-23.31 kg C ha⁻¹ yr⁻¹ in 0–20 cm) under the high-emission scenario SSP585 by 2100. The trends of SOCS changes were consistent with those of SOCD (Fig. 11c, d). Trends of SOC from 2020 to 2060 and 2100 under different climate change scenarios (SSP119, SSP245, and SSP585) in 0–20 cm and 0–100 cm in arid and humid regions. The changes of SOCD (kg C ha−1 yr−1) in (a) 0–20 cm and (b) 0–100 cm. And the changes of SOC stocks (SOCS) (Gg C yr−1) in (a) 0–20 cm and (b) 0–100 cm. “Total” is a combination of “arid” and “humid” regions. Figures 12 and 13 plot the spatial distribution maps of the trends of the SOCD change. Overall, SOCD showed the most drastic changes under the high-emission scenario SSP585, with the most evident decreasing trends. Followed by SSP245, a medium-emission scenario, and the most slight changes under the low-emission scenario SSP119. In general, under the three emission scenarios (SSP119, SSP245, SSP585), Northeast China and the Tibetan Plateau were the main carbon sources. Moreover, with the change of emission scenarios from low to high emissions, the area of regions transforming into carbon sources continued to expand, and the magnitude of the SOCD reduction increased. Spatial distribution of the average changes of SOC (kg C m−2) from 2020 to 2060 under different climate change scenarios (SSP119, SSP245, and SSP585) in 0–20 cm and 0–100 cm in arid and humid regions. Spatial distribution of the average changes of SOC (kg C m−2) from 2020 to 2100 under different climate change scenarios (SSP119, SSP245, and SSP585) at 0–20 cm and 0–100 cm in arid and humid regions. Discussion Spatial patterns of SOC in the arid and humid regions The results indicated that SOC shows a gradual increase trend from extremely arid, arid, semi-arid to semi-arid and sub-humid regions in arid regions (Table 3), which is consistent with the findings of Ren et al. (2024)75 on the impact of aridification on soil organic carbon in the northwestern arid regions. However, no similar pattern was observed in humid regions. This is mainly due to the abundant moisture and heat conditions in humid regions, which are conducive to the accumulation of SOC from plant sources76. In terms of spatial structure, the distribution of SOC in arid regions exhibits significant spatial characteristics77. For example, the SOCD is relatively high along mountain ranges. Such as the Qilian Mountains, Bayankala Mountains, and Greater Khingan Range in the Tibetan Plateau; the Tianshan and Altai Mountains in northern Xinjiang; the Tarim Basin in central Xinjiang; and the Changbai Mountains in northeastern China. However, similar spatial distribution characteristics were not observed in humid regions. Arid regions in China have abundant SOC stocks, but the SOC here is more fragile78. Therefore, more scientific management measures need to be taken to maintain existing carbon storage capacity and explore its carbon sequestration potential. Driving mechanisms of SOC in arid and humid regions The results indicated that SOCD in arid regions of China was significantly lower than in humid regions (Fig. 5), and spatial heterogeneity was more pronounced (Table 3). This is mainly driven by the heterogeneity of hydrothermal conditions and soil physical and chemical properties. As previous studies have revealed, climate, soil physical and chemical properties, vegetation, spatial location, and human activities collectively determine the spatial distribution and spatial heterogeneity of SOC22,31,79. Precipitation and temperature play an important role in regulating SOC dynamics2,80,81. Arid regions are constrained by water availability, and SOC is particularly sensitive to changes in precipitation. As is well known, vegetation growth is influenced by water supply76,82,83. However, water limitations in arid regions (MAP < 400 mm) lead to reduced primary productivity, which in turn reduces the input of plant-derived carbon into the soil84,85. Temperature has an adverse effect on SOC accumulation by exacerbating aridification75 and reducing vegetation photosynthetic efficiency86. With climate change, SOC in the Tibetan Plateau and Northeast China will decline sharply. This is mainly attributed to rising temperatures, which increase enzyme activities in cold, humid regions, exacerbating soil respiration and carbon emissions84,87,88. In humid regions, where moisture conditions are abundant, soil organic carbon exhibits higher temperature sensitivity. As revealed by Liu et al. (2016)89 and Ge et al. (2017)90, temperature regulates SOC dynamics by affecting soil temperature, moisture, vegetation growth, and microbial activities. Compared with low-temperature and low-humidity environments, warm and humid conditions can more fully release microbial activities, thereby accelerating the decomposition of SOC30,91. In addition to climatic factors, soil physical and chemical properties (such as clay content, sand content, soil moisture content, and soil pH) also play an important role in the dynamic changes of SOC. Firstly, soil moisture content affects SOC dynamics in arid and humid regions through two different pathways. Low soil moisture content in arid regions not only adversely affects NPP, but also inhibits microbial activities due to insufficient water availability, thereby limiting carbon accumulation30. Conversely, high soil moisture content in humid regions may also inhibit microbial activity through oxygen supply limitations45, leading to SOC changes. Secondly, soils in arid regions have high sand content and poor aggregate stability, making it difficult to form stable mineral-associated organic carbon (MAOC)92, which makes SOC more susceptible to erosion and mineralization93. In addition, soil pH values in arid regions are typically higher, creating a more alkaline environment that promotes the mineralization of organic matter94, thereby exacerbating its loss. Conversely, the higher silt and clay content in humid regions favors the formation of stable components through physical adsorption and chemical bonding with carbon95, which is considered an important mechanism for SOC stabilization96. However, our path analysis results indicate that within the 0–100 cm soil depth in humid regions, clay content exhibits a negative correlation with SOCD (Fig. 10). This finding contradicts prevailing assumptions and may stem from the influence of other environmental factors in humid regions (e.g. precipitation) on soil aggregates. Further research is required to validate this observation. In addition, microorganisms are inhibited in low pH environments28,97,98, which also slows down carbon decomposition and mineralization processes, thereby maintaining SOC stability. In addition, the effects ofelevation, vegetation, and human activities on SOC cannot be ignored. The results of the relative importance analysis of the RF model indicated that elevation plays a significant role in the spatial heterogeneity of SOC in both arid and humid regions. In particular, the Tibetan Plateau is known as the roof of the world99. Here, vegetation types such as alpine meadows and marshes provide a large input of SOC, while the unique permafrost environment inhibits decomposition and mineralization processes, thus maintaining long-term stability of SOC. Previous studies have also pointed out that elevation increases lead to lower temperatures, which also inhibit the microbial decomposition and mineralization processes of SOC100. Human activities have influenced the SOC dynamic by disturbing natural soils101,102. Dynamics of SOC under climate change This study revealed the dynamics of SOC under climate change (Figs. 11, 12 and 13, Fig. S6-S13). The driving mechanisms originate from the interactions between rising temperatures, changes in precipitation patterns, and ecosystem productivity. In humid regions, SOC showed an upward trend, which is consistent with the effects of increased precipitation47,103 and warming104 on NPP. Because the addition of plant debris, dead branches, and fallen leaves leads to the accumulation of more fresh SOC. This supports the hypothesis that climate-induced increases in carbon inputs may partially offset losses from SOC decomposition27. However, in arid regions, SOC showed a downward trend under climate change, highlighting the vulnerability of SOC in arid regions104. This is because warming accelerates the mineralization and decomposition rates of SOC32,105. Regional differences in SOC dynamics further reveal the specificity of terrestrial ecosystems’ responses to climate change. Although the model takes into account climate-carbon feedback effects, uncertainties remain regarding nutrient cycling constraints and land use interactions106. They emphasize the highly nonlinear characteristics of SOC in response to environmental factors, as well as the urgency of protecting SOC. A core assumption underlying our future projections is that all environmental variables other than climate remain constant, allowing us to focus our forecasts on purely climate-driven factors. However, our sensitivity analysis reveals that future LULC changes represent a significant source of uncertainty in SOCD. Assuming static LULC in 2020 led to our overestimation of future SOCD (Fig. S14, Table S2). Under the medium emissions scenario, socioeconomic development and the expansion of industrial manufacturing and agricultural cultivation have driven significant growth in farmland and built-up areas, resulting in substantial carbon emissions107. Nevertheless, the implementation of ecological conservation policies has played a significant role in enhancing carbon storage capacity while reducing emissions. Consequently, under scenarios of intense LULC change, projected future SOCD in China is only slightly lower than projections under static LULC scenarios. Spatial patterns of SOCD and NPP We compared our baseline SOCD distribution in 2020 with China’s regional NPP distribution in 2020. We found that in humid regions, areas with high SOCD distribution showed good agreement with high NPP distribution, whereas this correlation was absent in arid regions. Generally, high NPP increases soil carbon input, exerting a positive driving force for enhancing soil carbon storage108,109. However, increased plant growth may also intensify the extraction of soil nutrients and stimulate the decomposition and loss of soil carbon109. This finding provides significant insight into why SOCD remains low in many regions despite the prevalence of high NPP in moist zones. Within these ecosystems, nitrogen limitation may drive increased plant growth, which in turn accelerates soil carbon decomposition to extract nutrients for sustaining this additional growth110. In arid regions, the spatial decoupling relationship between NPP and soil carbon storage indicates that sustained soil carbon inputs and low outputs are the primary factors governing the stability of SOC pools. Therefore, carbon input and carbon retention time must be considered simultaneously in carbon cycle models for arid zones. Limitations and uncertainties This study has the following limitations. Firstly, due to the limited number of sampling points and their uneven distribution, prediction accuracy is poor in conditions such as complex terrain and extreme cold. Secondly, the environmental covariates used in the modeling are very limited, and there is a lack of other equally important variables, such as above-ground biomass, root biomass, microbial biomass, and microbial activities. Finally, when predicting SOC under climate change, only changes in MAP and MAT were considered, while other input variables were assumed to remain consistent with their current states. But, this assumption lacks credibility, as all variables are subject to dynamic change. Therefore, both the availability of data and the inherent uncertainty of the model indicate that there is still significant room for improvement in this field of research in the future. Conclusion In this study, we utilized the SOCD dataset of Chinese terrestrial ecosystems (2010s), employing methods such as RF to map the spatial distribution of SOC. We analyzed the impact of environmental factors on the SOCD in 0–20 cm and 0–100 cm in arid and humid regions, quantified the direct and indirect effects of driving factors, and explored the trends of SOC under the SSP119, SSP245, and SSP585 scenarios. The results indicated: (1) The RF model could better simulate the spatial distribution patterns of SOC both in 0–20 cm and 0–100 cm in arid and humid regions of China. (2) SOCD in arid regions were significantly lower than that in humid regions, and exhibited a greater spatial heterogeneity. Overall, Northeast China, the Tibetan Plateau, Southwest China, and the Tarim Basin had relatively high values of SOCD. However, the low-level SOCD were distributed in the Loess Plateau, Northwest China, Inner Mongolia, the North China Plain, and certain areas of Xinjiang. (3) The variations of SOCD in arid regions were primarily driven by MAT, NDVI, MAP, SM, and LULC. The variations of SOCD in humid regions were primarily driven by MAP, MAT, elevation, LULC, and HFP. (4) Under the SSP119, SSP245, and SSP585 scenarios, SOCD exhibited a decreasing trend under all other scenarios, except for a slight increase by 2060 under the low-emission scenario SSP119 in arid regions. However, in humid regions, SOCD exhibited an increasing trend under all other scenarios, except for a decreasing trend by 2100 under the high-emission scenario SSP585 in 0–20 cm. The results of this study are of great significance for understanding the response of the SOC pool to climate change in the arid and humid regions of China, identifying vulnerable areas and carbon sinks, and scientifically formulating regional carbon management policies. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Acknowledgements The authors gratefully acknowledge the financial support provided by the National Natural Science Foundation of China (42171067), and the National Key Research and Development Program of China (2023YFD1901204). Author contributions B.Y.: Writing—review & editing, Writing—original draft, Visualization, Validation, Methodology, Supervision, Formal analysis, Conceptualization. S.Z.: Writing—review & editing, Writing—original draft, Validation, Supervision, Data curation, Conceptualization. X.W.: Writing—review & editing, Supervision, Funding acquisition. Funding Declaration. This work was supported by the National Natural Science Foundation of China (42171067), and the National Key Research and Development Program of China (2023YFD1901204). Data availability Data collected and analyzed in this study are available from the corresponding author upon request. Declarations Competing interests The authors declare no competing interests. Ethical approval There was no ethical approval required for this study. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Data Availability Statement Data collected and analyzed in this study are available from the corresponding author upon request.