Northeast China holds huge wetland soil organic carbon storage: an estimation from 819 soil profiles and random forest algorithm

As a huge natural carbon storage, wetlands play an important role in the global carbon cycle. However, the spatial pattern and storage of soil organic carbon (SOC) in wetland ecosystems remain largely uncertain due to large spatial heterogeneity and insufficient field observations. In this study, we predict the spatial pattern of SOC density and estimated SOC storage in wetlands of Northeast China based on 819 field samples and multiple geospatial data using random forest algorithm. The SOC density of wetlands at different depths was affected differently by environmental factors and the SOC density in the surface layer (0–30 cm) was more susceptible to climatic change. The correlation coefficients (r) between the SOC density predicted by the random forest model and the measured SOC density were 0.86, 0.77 and 0.73 in 0–30, 30–60 and 60–100 cm soil depths, respectively. Our estimation showed that Northeast China holds huge wetland SOC storage in the amount of 3.40 ± 0.13 Pg C. The average wetland SOC density was 44.30 ± 1.72 kg C m−2, which decreased gradually from north to south in the study area. The wetland SOC density in Northeast China decreased with soil depth, and the influence of environmental factors on wetland SOC gradually decreased. Surface wetland SOC may be more sensitive to global climate change. Our results examined the relationship between wetland SOC and environmental factors, which benefits the understanding of the responses of wetland SOC to climate change.


Introduction
Wetlands are known as the "kidneys of the earth" and have various ecosystem functions and services, such as carbon sequestration, climate regulation, and biodiversity conservation (Costanza et al. 1997, Wang et al. 2012, Woodward & Wui 2001).As an important carbon pool in the Earth's surface system, wetlands play an important role in the global carbon cycle (Dargie et al. 2017;Gorham 1991;Maltby & Acreman 2011).Although wetland areas only account for 4∼6% of the earth's land, wetlands hold 20∼30% of the world's soil organic carbon (SOC) storage (Batjes 2014;Mitra et al. 2005).However, global wetland SOC storage estimation has an apparent uncertainty (Mitra et al. 2005;Mitsch et al. 2013), and the largest estimation was more than twice of the smallest, ranging from 202 to 535 Pg C (Adams et al. 1990;Gorham 1991).The large differences between these estimations may come from the multiplicity of data sources and discrepancies in methodologies (Mitra et al. 2005).Thus, an accurate estimation of the wetland SOC storage is crucial for updating the carbon budget and predicting the carbon-climate feedback (Wiesmeier et al. 2013).
Estimating wetland SOC density and understanding its controlling factors are getting more and more worldwide attention (Nahlik & Fennessy 2016;Xiao et al. 2019).There are apparent differences in wetland SOC density with the depth of 0-100 cm in different countries or regions.Carnell et al. (2018) reported that the wetland SOC density in southeastern Australia was 20.4 ± 0.1 kg C m −2 .Nahlik and Fennessy (2016) estimated the average wetland SOC density in the United States to be 47.8 ± 5.8 kg C m −2 .The SOC density of China's wetland was estimated to be 46.71± 4.32 kg C m −2 , which was close to the wetland SOC density in the United States (Xiao et al. 2019).However, in Russia, the estimated wetland SOC density reached 81.2 kg C m −2 (Stolbovoi 2002).The difference in wetland SOC density was mainly affected by various environmental factors, such as soil, topography, climate, hydrology, vegetation, and human interference (DeLaune et al. 2018;Nahlik & Fennessy 2016;Scharlemann et al. 2014).Soil texture and climate were the main influencing factors of tidal wetland SOC in the United States (Holmquist et al. 2018).Climate and plant biomass were the main factors controlling wetland SOC distribution (Xiao et al. 2019;Kim et al. 2022).Temperature was a significant control of SOC storage in the northern wetlands (Gorham 1991).The impacts of environmental factors on SOC density decreased with depth in the Greater Khingan Mountains, China (Man et al. 2019).Therefore, conducting more regional studies is crucial for wetland SOC reserves at national and global scales.
China has the largest wetland area in Asia, accounting for approximately 10% of the global wetland area (Mao et al. 2020).As a huge natural carbon storage in China, estimations of wetland SOC storage at depths of 0-100 cm differ among studies due to differences in the measurement methods and estimated wetland areas (Zheng et al. 2013;Xu et al. 2018;Xiao et al. 2019).For example, Zheng et al. (2013) (Grimm et al. 2008;Ren et al. 2020;Wiesmeier et al. 2011).Compared with previous use of regional statistics, inventory approaches, and traditional interpolation methods based on spatial autocorrelation, such as Kriging and inverse distance weighting, the random forest model was more accurate and reliable because of limited spatial autocorrelation of SOC density among samples and non-zonal distribution characteristics of wetlands.The random forest method can fully consider the influence of different environmental factors, avoiding the strong spatial autocorrelation of traditional interpolation methods (He et al. 2021, Li et al. 2022).
Northeast China is located in the mid-high latitudes where wetlands are widely distributed (Mao et al. 2020).The climate in this area is relatively humid and the temperature is low, which is conducive to the accumulation of SOC (Yu et al. 2007).Previous studies have analyzed wetland SOC in parts of Northeast China.For example, Ren et al. (2020) and Kang et al. (2020) used different methods to predict the spatial distribution of wetland SOC in the western Songnen Plain and the Liao River Plain.Man et al. (2019) studied the spatial and vertical variations in wetland SOC concentrations and their controlling factors in the Greater Khingan Mountains Region.However, there is a lack of accurate estimation of wetland SOC storage in Northeast China.Previous studies have evidenced that soil carbon emissions due to climate warming will reduce soil carbon storage (Li et al. 2022).Changes in land cover types caused by human activities can also affect wetland carbon storage.Accurate estimation of the wetland SOC storage and exploration of the factors affecting SOC density are still pressing issues.
In this study, we estimated the wetland SOC density and storage in Northeast China based on random forest algorithm and multisource geospatial datasets, as well as a large amount of profile data.It can provide a basis for understanding the regional carbon cycle and provide theoretical guidance for wetland management.Specifically, in this paper, we aimed to 1) examine the wetland spatial pattern of SOC density across the study area; 2) estimate wetland SOC storage in Northeast China; and 3) investigate the relationship of various environmental factors with the SOC density of wetlands at different depths.The generated dataset and related analysis in this study will help update the carbon budget and predict carbon-climate feedback.

Study area
The study area is in Northeast China (115°37′-135°50′ E and 38°43′-53°34′ N), including Liaoning, Jilin and Heilongjiang provinces and eastern Inner Mongolia with an area of 1.24 × 10 6 km 2 (Fig. 1).The climate is a north-temperate and semi-humid continental monsoon climate, with annual precipitation varying from 200 to 800 mm.More than 70% of the annual precipitation occurs in June, July, and August (Ren et al. 2022).The temperature decreases from south to north, with mean annual temperatures between 10 and -8 °C.The soil types in the study area are dominated by black soils, including Chernozem, Phaeozem, Bog soil and Meadow soil.The cropland area is approximately 20% of the whole region and

Wetland data
Wetlands referring to vegetated wetlands, including bog, fen, marsh, and swamp, are widely distributed in Northeast China, with an area of 7.67 × 10 4 km 2 .The wetland distribution data set was extracted from CAS_wetlands, which classified wetlands from Landsat images using a hybrid object-based and hierarchical classification approach (Mao et al. 2020).The data set provides reliable classification with an overall accuracy of 95.1% and thus has been applied in many studies (http:// www.geoda ta.cn/ thema ticVi ew/ wetla nd2020.html).For comparative analysis of the spatial variation of wetland SOC density and the differences in carbon storage between different regions, the study area was divided into six geographic regions (Mao et al. 2019): the Greater Khingan Mountains Region, the Sanjiang Plain, the Changbai Mountains Region, the Lesser Khingan Mountains Region, the Liao River Plain, and the Songnen Plain (Fig. 1).

Soil sampling
Before the field survey, we preset some sample sites based on the wetland map of Northeast China.We fully considered the distribution of wetlands and the accessibility of traffic to make the sampling points evenly distributed in order to represent the wetlands of the whole study area.The field survey was conducted from September to October when the soil was not yet frozen, and the soil moisture content was relatively low.Since the wetland SOC density has a small change in a short period, we don't have a large error in sampling for three consecutive years (Xiao et al. 2019).The depth of 100 cm was the commonly used soil depth for estimating SOC storage (Xiao et al. 2019;Zheng et al. 2013).Therefore, for convenience compared with other studies, we defined the field sampling depth to be 100 cm.In total, we collected 819 samples based on the spatial distribution of wetlands in the study area.(Fig. 1).Among them, 77 sampling points were acquired in 2015, 380 in 2016, and 362 in 2017.The sampling time and depth, and geographic locations were recorded while sampling.The geographic location file of all sample plots was mapped in the ArcGIS (version 10.2, Environmental Systems Research Institute, Redlands, CA, USA) by the longitudes and latitudes.We had three soil profiles within ten meters of the preset point, and the latitude and longitude of the soil profiles in the middle was used as the actual coordinates of the soil point.For each soil point with three soil profiles, the SOC content for each depth range (i.e., 0-30, 30-60, and 60-100 cm) was represented by the average of SOC values of three profiles at the sampling point (Hardisky et al. 1986).Due to the shallow soil layers or sampling difficulties in deep soil, the depth of some sampling points did not reach 60 cm or 100 cm, which resulted in the number difference of observations between various sampling depths.The soil at the center of each depth interval was collected using a standard container with a volume of 100 cm 3 for bulk density determination and tools to obtain a soil column for SOC concentration determination.Duplicates of each soil sample were measured for SOC concentration in g C kg −1 using a Shimadzu analyzer.(SSM-5000A, Kyoto, Japan).The SOC concentration was determined by the potassium dichromate external heating oxidation method (Bernal 2008).The soil sample was heated with the potassium dichromate, in order to oxidize carbon from organic matter to carbon dioxide, then dichromate ions were reduced.According to the change in the amount of dichromate ions by the oxidization of carbon, the SOC concentration can be calculated (Ren et al. 2020).The calculation formulas of SOC density (formula (1)) and SOC storage (formula (2)) are as follows: In the formula, the unit of SOC density is kg C m −2 , H is the thickness of the soil layer (m), BD is the soil bulk density (kg m −3 ), and A is the area of wetlands (m 2 ).

Datasets
The variability in SOC density could be explained by its relationships with soil-forming factors, such as climate, terrain, and vegetation (DeLaune et al. 2018;Nahlik & Fennessy 2016;Scharlemann et al. 2014).Based on the data accessibility, we retrieved (1) (2) SOC storage = A × SOC density 17 geospatial data from remote sensing data, climate and terrain (Table 1).
The climatic data including mean annual precipitation (MAP, mm), mean annual temperature (MAT, °C), annual average humidity (MAH, %), mean annual wind speed (MAWS), and average evapotranspiration (AE) from 2008 to 2017, which were calculated from meteorological records were downloaded from the National Tibetan Plateau Data Center (http:// data.tpdc.ac.cn/).
Altitude was derived from the DEM data, downloaded from the Geospatial Data Cloud site, Computer Network Information Center, Chinese Academy of Sciences (http:// www.gsclo ud.cn).At the same time, we calculated factors such as Slope, SR (Surface roughness), Curvature, VC (Vertical curvature), HC (Horizontal curvature) and TWI (Topographic wetness index) based on Altitude (Ren et al. 2020).
All the above related geospatial information was unified into the Albers equal-area conic WGS84 coordinate system.The attribute values of all grids were extracted for all sampling points and used as inputs for the spatial algorithms.

Algorithm development and predictions
The random forest algorithm is an integrated machine learning algorithm.The algorithm generates the final results by averaging the class allocation probabilities of all produced trees.The trees are created by replacing a subset of training samples and randomly selecting variables in the R environment (v.4.0.2).The random forest algorithm is one of the more accurate machine learning algorithms.It can estimate the importance of predictor variables and run efficiently on large data bases (Akpa Stephen et al. 2016).The mean decrease in accuracy (%IncMSE) was used to present the relative importance of each variable.The value of the factor indicated the contribution of the factor in the algorithm.The greater the contribution value is, the greater the impact on the SOC density in the algorithm.When constructing the random forest model, we adjust the parameters of the model (mtry = 4, ntree = 500) to achieve the best model accuracy.Standard deviations were quantified by tenfold cross-validation simulations.General data processing and analysis followed the flow chart depicted in Fig. 2.

Accuracy verification
Tenfold cross-validation is a common test method used to test the accuracy of the algorithm.It divides the dataset into ten parts and takes turns using nine parts as training data and one part as test data for experimentation.It can obtain a good error estimate through many tests.
Tenfold cross-validation was used for testing and comparing the performance of the random forest algorithms based on the root mean squared error (RMSE, formula (3)), mean absolute error (MAE, formula (4)), mean absolute percentage error (MAPE, formula (5)), and correlation coefficient (r) between the measured and predicted values.The calculation equations for RMSE, ME, and MAPE are as follows: where y ′ is the predicted value of the SOC density (kg C m −2 ), y is the measured value of the SOC density (kg C m −2 ), and MAE is a measure of the predictive lack of bias.

Descriptive statistics of wetland SOC density
The descriptive statistical values of wetland SOC density are listed in Fig. 3.The average SOC density

Control factors of wetland SOC density
The relationships between wetland SOC density and major environmental factors, such as MAT, MAP, MAH, and altitude, are shown in Fig. 4. The wetland SOC density at all three soil depths of 0-30, 30-60, and 60-100 cm decreased with the increase of MAT (Fig. 4 a1, a2, and a3).The most notable relationship between wetland SOC density and MAP was detected for the topsoil of 0-30 cm (P < 0.001, r = 0.54, Fig. 4 a1).The wetland SOC density at all depths was positively correlated with MAP and MAH (P < 0.001, Fig. 4 b1, b2, b3, c1, c2, and c3); that is, with the increase in MAP and MAH, the wetland SOC density tended to increase.As shown in Fig. 4 d1, d2, and d3, wetland SOC density at all soil depths increased linearly with increasing elevation (P < 0.001).
The random forest algorithm was used to predict the SOC density of wetlands at different depths.The results of the accuracy verification are shown in Table 2.The RMSEs at the 0-30 cm, 30-60 cm, and 60-100 cm soil depths were 6.54 kg C m −2 , 3.57 kg C m −2 , and 3.52 kg C m −2 respectively.The MAPE at the 0-30 cm, 30-60 cm, and 60-100 cm soil depths were 10.79%, 12.33%, and 23.2%, while the correlation coefficients (r) at different depths were 0.86, 0.77 and 0.73, respectively.These validations suggested that the trained random forest models are capable to predict the spatial distribution of SOC density at the different soil depths.
Figure 5 presents the contribution of each variable to the random forest model, which was calculated by the mean decrease in accuracy.At the 0-30 cm soil depth, MAT had the largest impacts on the wetland SOC density, followed by LST, MAH, altitude, and LAI.At the 30-60 cm soil depth, remote sensing factors (e.g., EVI and NDVI) and MAT had the greatest influences on SOC density, followed by altitude and LST.At 60-100 cm intervals, altitude played a dominant role in affecting SOC density, followed by MAT, MAH and NDVI.Terrain factor (e.g., TWI) was also important factors influencing wetland SOC density at this soil depth.

Spatial patterns of wetland SOC density and storage in Northeast China
Figure 6 presents the spatial patterns of wetland SOC density at different soil depths over Northeast China estimated by the random forest algorithm.At the 0-30 cm soil depth, the average SOC density was 24.09 ± 0.92 kg C m −2 .In terms of space, the SOC density in the northwest was significantly higher than that in other regions.At the 30-60 cm soil depth, the average wetland SOC density was 10.53 ± 0.39 kg C m −2 .At the 60-100 cm soil depth, the average wetland SOC density was 9.68 ± 0.41 kg C m −2 .The wetland SOC density at a 0-30 cm depth was significantly higher than that at the 30-60 and 60-100 cm soil depths.As the soil depth increased, the wetland SOC density decreased correspondingly.The average SOC density of the wetland of 0-100 cm was  44.30 ± 1.72 kg C m −2 .The SOC density at different depths was spatially lower in the south than in the north, showing a trend of gradual decrease from north to south.
Figure 7 shows the statistical SOC density of different regions.The SOC density of different regions was significantly different.At different depths, the SOC density was the highest in the Greater Khingan Mountains Region and the lowest in the Liao River Plain.At the 0-100 cm soil depth, the SOC density in the Greater Khingan Mountains Region was 56.87 ± 2.13 kg C m −2 , which was much higher than the average value in Northeast China.In different regions, the SOC density of the Lesser Khingan Mountains Region, the Changbai Mountain, the Sanjiang Plain, and the Songnen Plain gradually decreased at depths of 0-100 cm.The SOC density in the Liao River Plain was the smallest, only 13.85 ± 0.69 kg C m −2 .
According to calculations, the wetland SOC storage at a 0-100 cm soil depth in Northeast China was approximately 3.40 ± 0.13 Pg C (Table 3).Among them, the 0-30 cm, 30-60 cm, and 60-100 cm soil depths of SOC storage were 1.85 ± 0.07 Pg C, 0.81 ± 0.03 Pg C, 0.74 ± 0.03 Pg C respectively.The SOC storage of 0-30 cm accounted for the main part, which was approximately 54.49%.The SOC storage of different regions was significantly different (Fig. 7).At the 0-100 cm soil depth, the SOC storage in the Greater Khingan Mountains Region was 2461.87 ± 92.21Tg C (1 Tg C = 1 × 10 6 t C), which was much higher than other regions.In different regions, the SOC storage of the Lesser Khingan Mountains Region, Songnen Plain, Sanjiang Plain and Changbai Mountain gradually decreased at depths of 0-100 cm.The SOC storage in the Liao River Plain was the smallest, only 26.73 ± 1.33 Tg C.

Prediction method of wetland SOC density
We generated spatial estimates of wetland SOC density at a 30 m resolution in Northeast China by combining random forest algorithm with multiple geospatial data.Using a new database with 819 field sampling sites, we quantified the spatial distribution of SOC storage in Northeast China and determined the differences in different regions (Fig. 6 and Fig. 7).At the same time, we identified significant environmental predictors of SOC storage at depth intervals.
Data sources and methods were two differences of wetland SOC storage estimation between this study and the previous studies focusing on wetland SOC storage estimation.Our sampling time was more concentrated, and the number of sampling points was more than that in previous studies on wetland SOC storage estimation Vol:.( 1234567890) (Xiao et al. 2019;Zheng et al. 2013).A case in point is the work of Xiao et al. (2019), who collected data from 2005 to 2018 at around 100 points.However, the number of sampling points in this paper reached 819, and the sampling time was concentrated from 2015 to 2017.All field sampling data were obtained through a unified processing method, avoiding uncertainty from different data sources.Our wetland distribution data were derived from the up-to-date wetland distribution dataset of China (CAS_Wetland) with high classification accuracy (Mao et al. 2020).Moreover, we integrated climate, topography, and remote sensing variables characterize different features to predict the spatial pattern of the wetland SOC density in Northeast China.Because of the broad scale of Northeast China and non-zonal distribution of wetlands, estimation of wetland SOC using zonal statistical methods, inventory methods, and traditional interpolation methods based on spatial autocorrelation, such as kriging and inverse distance weighting, may lead to large errors.Therefore, the random forest algorithm was used for mapping SOC density which fully considered the different influences of environmental variables (Song et al. 2017).All these improvements enhanced the reliability and accuracy of the wetland SOC storage estimation in Northeast China, which was the novelty of this study compared to previous researches.

Spatial heterogeneity and influencing factors of wetland SOC density
We found that the wetland SOC density in Northeast China decreased with soil depth.This result is consistent with the report in other regions (Xiao et al. 2019).The wetland SOC in Northeast China was concentrated in the surface layer (0-30 cm), accounting for approximately 54% of the storage in total depth of 100 cm.This finding is also consistent with the estimation of global wetland SOC storage (Balesdent et al. 2018).In this study, the mean wetland SOC density was 44.30 ± 1.72 kg C m −2 in the upper 100 cm soil, while that of global wetlands is approximately 43.27-102.88kg C m −2 (Mitra et al. 2005).The wetland SOC density in Northeast China is at a relatively low level worldwide.However, Xiao et al (2019) estimated China's wetland SOC density to be 31.17kg C m −2 , which suggests a larger SOC density value of Northeast China than the national average.Ma et al  The wetland SOC density in Northeast China has clear spatial heterogeneity, with a general increase trend with increasing latitude.The SOC density of the Greater Khingan Mountains Region was the highest, while that of the Liao River Plain was the lowest.Many studies have reported that climatic factors, in particular temperature and precipitation, are the most important determinants of SOC distribution at large scales (Jobbagy & Jackson 2000;Wiesmeier et al. 2013).Our study showed that the main influencing factors of wetland SOC density at different depths were varied.The wetland SOC density in the surface soil layer (0-30 cm) was affected mostly by climate (such as MAP and MAT).The low temperature was conducive to the storage of SOC, while the high temperature caused the decomposition of SOC (Crowther et al. 2016).The temperature in the study area tended to decrease gradually with increasing latitude, which could clarify the difference in the spatial distribution of wetland SOC density in our study.The Greater Khingan Mountains Region had large areas of peatland with low human activity intensity and low temperature.Therefore, higher SOC density was observed in the soil profile (Wang et al. 2021).The effect of climate on SOC density gradually decreases with increasing depth.Vegetation residues were conducive to the formation of humus and the enrichment of organic carbon, and the long-term deposition gradually increased the SOC density of the middle soil layer (30-60 cm).Therefore, NDVI and EVI, characterizing vegetation conditions, had an important influence on SOC density in the middle soil layer (30-60 cm).The deeper layer (60-100 cm) was buried deep in the soil, and climate change and vegetation index had little impact on SOC density.Compared with low-altitude wetlands, wetlands in middle-or high-altitude areas such as the Greater Khingan Mountains Region still have higher SOC density in deeper soil layers (60-100 cm), so the importance of topography for the density in the deeper layers (60-100 cm) gradually increases.

Potential response of wetland SOC storage to human and climatic interference
Our study indicated that the relationship between wetland SOC density and climatic factors gradually decreases with increasing soil depth.This means that the surface SOC may be more sensitive to global climate change.High latitude regions have the fastest rate of temperature increase.Therefore, carbon loss caused by warming may occur in the wetlands of Northeast China.Northeast China is located in the middle and high latitudes, and there are large areas of permafrost.Global warming could put the total SOC storage at risk of instability.Cao et al. (1998) found that global warming may produce higher methane emissions, and increasing temperatures could melt permafrost soils and subsequently emit methane hydrates entrapped by these wetlands (Meng et al. 2017).In addition, climate warming and human activities will have an important impact on wetlands.Xue et al. (2021) study showed that 4.5-23.8% of the high-latitude wetlands in the Xing'anling Mountains would be lost following widespread thawing of permafrost under different climate warming scenarios by the end of this century.Mao et al. (2018Mao et al. ( , 2021) ) study showed that the impact of human activities led to the disappearance and degradation of wetlands.Wetland degradation will switch from being net sinks of carbon to becoming sources of greenhouse gas sources that accelerate climate change (Tan et al. 2022).Human activities may have important effects on wetland carbon sink and SOC storage.Therefore, it is important to correctly understand the carbon sink function of wetland ecosystems.

Conclusion
We achieved an accurate estimation of wetland SOC storage in Northeast China based on 819 soil profiles and random forest algorithm.The average SOC density of the wetlands in Northeast China was 44.30 ± 1.72 kg C m −2 , which decreased with soil depth, and the surface Vol.: ( 0123456789) layer (0-30 cm) SOC accounted for approximately 54% of the storage in total depth of 100 cm.SOC storage was abundant, with total storage reaching 3.40 ± 0.13 Pg C. Through spatial analysis, it was found that the SOC density generally showed a trend of more in the north and less in the south.By establishing independent algorithm at different depths, we concluded that the SOC density of wetlands at different depths was affected differently by environmental factors.The SOC density of wetlands at different depths was affected differently by environmental factors and the SOC density in the surface layer (0-30 cm) was more susceptible to climatic change.Climate change and human activities will have an important impact on wetland SOC storage.Given that this region contributes a substantial amount of SOC sequestration, proper wetland management may be very important for the accumulation of SOC substances.

Fig. 1
Fig. 1 Distribution of the wetlands and sampling points in Northeast China

Fig
Fig. 2 Flowchart for quantifying wetland SOC storage

Fig. 3
Fig. 3 Frequency distributions of SOC density at different soil depths.Note: SD: standard deviation

Fig. 7
Fig. 7 SOC density and storage statistics in different regions estimated that wetland SOC storage in China ranged from 5.04 Pg C to 6.19 Pg C based on 7799 soil profile data and inventory approach.Xu et al. (2018) obtained 3.62 ± 0.80 Pg C from a literature synthesis published from 2004 to 2014, while the estimate of Xiao et al. (2019) was 16.82 ± 2.10 Pg C using 370 sites across China and 193 research data sets.With the development of computer technology, machine learning models have been widely used, due to powerful data mining capabilities

Table 2
Random forest algorithm accuracy at different soil depths Vol.: (0123456789)

Table 3
Statistics of SOC density and storage