Hydrothermal and Vegetation-Mediated Controls on Soil Organic Carbon in an Alpine Headwater Region of the Tibetan Plateau: Implications for Sustainable Grassland Management

Abstract

Soil organic carbon (SOC) is essential for ecosystem stability and long-term carbon storage in alpine grasslands, yet the relative importance and interactions of hydrothermal and biotic controls remain poorly understood at regional scales. In this study, we quantified surface SOC (0–20 cm) across the Yellow River Source Region (YRSR) on the northeastern Tibetan Plateau, a climate-sensitive alpine headwater system characterized by strong hydrothermal gradients and freeze–thaw dynamics. Field-based SOC measurements were integrated with multi-source remote sensing and reanalysis data that describe thermal conditions, moisture processes, vegetation productivity, soil properties, topography, and human influence. A two-step screening approach was applied using Boruta and variance inflation factor filtering, followed by modeling with random forest. The model outputs were interpreted using Shapley Additive Explanations (SHAP). SOC displayed significant spatial heterogeneity across the region. Vegetation productivity, moisture availability, and thermal conditions were identified as the dominant nonlinear drivers of SOC variation. Moisture availability emerged as a central regulator of SOC, affecting it both directly and indirectly through vegetation productivity and thermal conditions. These findings underscore the importance of hydrothermal stability in sustaining soil carbon stocks and provide a quantitative basis for adaptive grassland management under a warming climate.

1. Introduction

Soil organic carbon (SOC) is a major component of terrestrial biogeochemical cycling and is widely regarded as an important indicator of soil function, ecosystem resilience, and carbon–climate interactions [1,2]. In alpine grasslands, SOC dynamics is intricately linked to hydrological regulation, vegetation productivity, and permafrost dynamics. Even slight variations in soil temperature and moisture under ongoing climate warming can significantly alter the carbon balance, making alpine SOC both a vulnerable component of the climate system and highly relevant to sustainable land management [3]. However, SOC should be interpreted within a broader soil biogeochemical framework, as soil functioning is also shaped by nutrient availability, hydrological conditions, and vegetation–soil interactions [1,4]. The Tibetan Plateau, often referred to as the “Third Pole,” contains vast alpine grasslands and permafrost systems that are highly sensitive to warming [5]. These grasslands store approximately 7.4 Pg of SOC within the upper 100 cm of soil, with significant spatial variation largely driven by moisture availability [6,7]. However, warming-induced permafrost degradation may accelerate organic matter decomposition and carbon release, whereas increases in vegetation productivity may enhance carbon inputs [8,9]. These opposing processes generate complex, spatially heterogeneous SOC responses, adding uncertainty to the assessment of regional carbon balance and alpine soil functioning.

At regional scales, SOC in alpine environments is characterized by strong spatial heterogeneity, typically exhibiting right-skewed distributions with localized high-carbon patches embedded within broader low-carbon areas [5]. These spatial patterns suggest that regional carbon stability is strongly dependent on areas characterized by favorable hydrothermal conditions, including wetlands and seasonally moist grasslands. Identifying and protecting these high-SOC areas is thus essential for sustainable grassland management and climate adaptation strategies [10]. The dynamics of SOC in these alpine systems are governed by interacting thermal, hydrological, biological, and edaphic processes [11]. Temperature influences microbial activity and freeze–thaw disturbances, moisture availability governs vegetation growth and soil redox conditions, and soil texture and mineral composition affect carbon stabilization YRSR [12,13]. These tightly coupled processes generate nonlinear and hierarchical responses, which are difficult to capture using conventional linear models [14,15,16]. Recent advances in remote sensing have enabled spatially explicit assessment of vegetation productivity, surface temperature, moisture conditions, and cryospheric processes, providing new opportunities for SOC research in remote alpine environments [17,18,19]. Although machine-learning approaches such as RF have improved SOC prediction, many studies still emphasize predictive accuracy over ecological interpretation and management relevance [20,21]. Bridging predictive modeling with process-based understanding is therefore critical for translating SOC mapping into effective land management.

Against this background, this study examines SOC variability in the YRSR. By integrating field-measured SOC data with multi-source environmental predictors, this study aims to characterize SOC spatial heterogeneity, identify its dominant hydrothermal and biotic drivers, and clarify the pathways influencing SOC distribution. These objectives are relevant for identifying vulnerable carbon pools and supporting adaptive grassland management under climate change.

2. Materials and Methods

2.1. Study Area

The YRSR, located on the northeastern Tibetan Plateau, forms the headwaters and a critical water supply area of the Yellow River. As an ecologically important region, it plays a key role in regulating regional hydrological processes, maintaining alpine ecosystem stability, and supporting downstream socioeconomic development [22,23]. Environmental changes in the YRSR therefore affect not only local ecosystem functioning but also water security and sustainable development across the broader Yellow River Basin (Figure 1). Originating from the Bayan Har Mountains, the study area ranges in elevation from approximately 2680 to 6248 m, with topography generally decreasing from northwest to southeast [24,25]. The landscape is characterized by glaciers, seasonal snow cover, lakes, and permafrost, representing a typical alpine ecosystem of the Tibetan Plateau [13,26]. The regional climate is cold and continental, with low mean annual temperatures, a short growing season, and frequent freeze–thaw cycles. Precipitation occurs mainly between May and September and shows pronounced seasonal and spatial heterogeneity [27,28]. Vegetation is dominated by alpine meadows, alpine steppes, and wetlands, with localized floodplains, marshes, and seasonally inundated depressions. Wetland and marsh areas, characterized by low temperatures and high moisture conditions, often develop peat-rich or humus-enriched soils with relatively high SOC accumulation. In contrast, surrounding alpine grasslands and wind-eroded uplands typically exhibit lower SOC contents [29,30]. The strong climatic, hydrological, and vegetation gradients across the YRSR provide a suitable environmental setting for examining spatial variability in SOC.

2.2. Data Sources

2.2.1. Field Sampling and Laboratory Analysis

Soil sampling was conducted during July–August 2023 across the YRSR on the Qinghai-Tibet Plateau. A total of 240 sampling sites were selected to capture the spatial variability of SOC (Figure 1). To ensure representative coverage of environmental gradients with a limited number of samples, conditional Latin hypercube sampling (cLHS) was used to design the sampling sites. Auxiliary variables representing topography, climate, vegetation, and hydrological conditions were resampled to a 500 m resolution. Road networks, settlements, elevation, and slope were incorporated as constraints to construct a sampling cost surface, assigning penalty weights to inaccessible areas (Figure 2). The sampling design was implemented in R (version 4.5.2; R Foundation for Statistical Computing, Vienna, Austria) [31]. Samples were collected using a handheld soil auger (5 cm diameter) from the 0–20 cm soil layer to represent surface SOC characteristics. This depth was selected because it corresponds to the biologically active surface soil, where vegetation-derived carbon input, hydrothermal fluctuations, and management effects are most pronounced, and it is also commonly used in regional assessments of surface SOC in alpine grassland (Figure 1). Potential uncertainty in the soil dataset may arise from microsite heterogeneity, disturbance during sampling, and laboratory processing error. To reduce these effects, three sub-samples were collected and composited at each site, and combined into a composite sample to reduce the influence of small-scale spatial heterogeneity. All samples were air-dried, cleared of plant residues and gravel, ground, and passed through a 0.25 mm sieve prior to analysis. SOC content was determined using the potassium dichromate–sulfuric acid external heating oxidation method following standard procedures [32,33].

2.2.2. Construction of Environmental Predictor Set

To ensure that variable grouping and pathway specification were based on ecological processes rather than purely statistical associations, we adopted the STEP–AWBH framework as a process-oriented conceptual structure for organizing environmental predictors. In this study, the framework was used to group variables into ecologically meaningful components and to guide interpretation of potential regulatory pathways, rather than as a rigid standardized protocol [34]. Specifically, environmental drivers were partitioned into eight interacting components: soil properties (S), topography (T), ecological and vegetation processes (E), precipitation (P), atmospheric conditions (A), hydrological processes (W), biotic activity (B), and human influences (H) [34]. This process-oriented framework enabled the systematic representation of thermal regulation, moisture mediation, biological productivity, and anthropogenic disturbance within a unified conceptual model. Following this framework, 37 initial candidate environmental variables were compiled from multi-source remote sensing products and reanalysis datasets (

Table 1. Assembled environmental variables representing STEP-AWBH factors (S: soil, T: topography p, E: ecology, P: precipitation, A: Atmosphere, W: Water, B: Biota, H: Human).

Variable	Relevant Variable Abbreviation	Factor	Source	Resolution	Units	Date
Soil pH	pH	S	OpenLandMap	250 m	–	Static
Soil clay content	Soil clay content	S	SoilGrids v2.0	250 m	%	Static
Soil sand content	Soil sand content	S	SoilGrids v2.0	250 m	%	Static
Soil bulk density	BD	S	OpenLandMap	250 m	g/cm3	Static
Elevation	Elevation	T	SRTM DEM	30 m	m	Static
Slope	Slope	T	SRTM DEM	30 m	°	Static
Aspect	Aspect	T	SRTM DEM	30 m	°	Static
Height above nearest drainage	HAND	T	SRTM DEM	30 m	m	Static
Upslope contributing area	UpslopeArea	T	MERIT Hydro	90 m	m2	Static
Topographic Wetness Index	TWI	T	SRTM + MERIT Hydro	30 m	–	Static
Surface shortwave radiation	SR	E	ERA5-Land	~9 km	J m−2	2023
Terra daytime LST	Terra_LsTD	E	MODIS Terra MOD11A1.061	1 km	°C	GS 1
Terra nighttime LST	Terra_LsTN	E	MODIS Terra MOD11A1.061	1 km	°C	GS 1
Terra diurnal LST amplitude (annual)	Terra_LST_amp	E	MODIS Terra MOD11A1.061	1 km	°C	GS 1
Freeze–thaw cycle days	FTD	E	MODIS Terra MOD11A1.061	1 km	days	2023
Soil thaw proxy (MaySep)	LST_amp_MaySep	E	MODIS Terra MOD11A1.061	1 km	°C	2023
Normalized Difference Snow Index	NDSI_snowMean	E\W	Sentinel-2 Level-2A	20 m	–	CS 1
Cold-season snow frequency	SnowFreq_OctApr	E\W	Sentinel-2 Level-2A	20 m	%	CS 1
Annual snow frequency	SnowFreq_year	E\W	Sentinel-2 Level-2A	20 m	%	2023
Annual precipitation	Precip_year	P	ERA5-Land (ECMWF)	~9 km	mm	2023
May–Sep precipitation	Precip_MaySep	P	ERA5-Land (ECMWF)	~9 km	mm	GS 1
Annual mean temperature	Temp_year	A	ERA5-Land (ECMWF)	~9 km	°C	2023
May–Sep temperature	Temp_MaySep	A	ERA5-Land (ECMWF)	~9 km	°C	GS 1
Aridity Index (P/PET)	AI	A	ERA5-Land (ECMWF)	10 km	–	2023
Soil moisture	SM	W	SMAP L3	~9 km	m3 m−3	GS 1
Normalized Difference Water Index	NDWI	W	Sentinel-2 MSI L2A	20 m	–	2023
Leaf Area Index	LAI	B	Sentinel-2 Level-2A biophysical product	10 m	–m2/m2	GS 1
Normalized Difference Vegetation Index	NDVI	B	Sentinel-2 Level-2A Surface	10 m	–	GS 1
Enhanced Vegetation Index	EVI	B	Sentinel-2 Level-2A Surface	10 m	–	GS 1
Soil Adjusted Vegetation Index	SAVI	B	Sentinel-2 Level-2A Surface	10 m	–	GS 1
Modified Adjusted Vegetation Index	MSAVI	B	Sentinel-2 Level-2A Surface	10 m	–	GS 1
Gross Primary Productivity (May–Sep)	GPP_MaySep	B	MODIS/061/MOD17A3HGF	500 m	g C m−2 yr−1	GS 1
Net Primary Productivity (annual)	NPP	B	MODIS/061/MOD17A3HGF	500 m	g C m−2 yr−1	2023
Gross Primary Productivity (annual)	GPP	B	MODIS/061/MOD17A3HGF	500 m	g C m−2 yr−1	2023
Land cover class (ESA WorldCover)	LC_WorldCover	H	ESA WorldCover	10 m	Class	2020
Land cover class (MODIS IGBP)	LC_IGBP	H	MODIS MCD12Q1	500 m	Class	2023
Nighttime light intensity	NightLight	H	NOAA VIIRS	500 m	nW cm−2 sr−1	2023

¹ Note: GS represents the growing season (May–September 2023), and CS represents the cold season (October 2022–April 2023).

), ensuring that variable selection reflected ecological mechanisms relevant to SOC regulation rather than purely statistical associations. Environmental predictor values were extracted at the SOC sampling coordinates from the original raster layers while retaining their native spatial resolutions. To generate the regional SOC distribution map, the predictor layers used for spatial prediction were subsequently resampled and harmonized to a common 500 m spatial resolution. This ensured consistency among variables during regional mapping while preserving the original data characteristics during point-based extraction. The microclimatic influences were not measured directly at each sampling site, but were represented indirectly through spatial proxies of near-surface hydrothermal conditions. These variables included land surface temperature, freeze–thaw cycle days, soil moisture, snow-related indicators, and moisture-sensitive indices, which together reflect local thermal buffering, moisture availability, and seasonal energy–water balance.

2.2.3. Sentinel-2 Remote Sensing Data and Cloud Masking

High-resolution environmental variables representing vegetation productivity, surface moisture status, and cryospheric dynamics were derived from Sentinel-2 Multispectral Instrument (MSI) Level-2A surface reflectance products (10–20 m spatial resolution). The red-edge and near-infrared bands were used to compute vegetation indices sensitive to canopy structure and physiological activity, which are closely linked to carbon input processes in alpine ecosystems [35,36]. To reduce the influence of frequent cloud cover and terrain-induced shadows, all imagery was subjected to strict cloud and cloud-shadow masking prior to index calculation. Variable-specific temporal compositing strategies were applied to reflect ecological seasonality. Leaf Area Index (LAI) was derived from Sentinel-2 Level-2A surface reflectance data using the ESA SNAP Biophysical Processor, which applies a neural network trained on radiative transfer simulations to retrieve biophysical variables from red-edge and shortwave infrared bands. Growing-season (May–September) LAI was composited using the median to represent peak vegetation productivity. Surface moisture-related indices were calculated as growing-season means to represent average hydrothermal conditions. Snow-related metrics, including the normalized difference snow index (NDSI), were computed for the cold season (October–April), and snow frequency was quantified as the proportion of snow-covered observations relative to the total number of valid observations during the cold season [37,38,39]. This seasonally stratified preprocessing framework minimized atmospheric noise and ensured that derived variables were ecologically consistent with SOC formation and stabilization processes. Sentinel-2 imagery and other remote sensing datasets were accessed and processed using the Google Earth Engine (GEE) cloud-computing platform (https://earthengine.google.com/).

2.3. Methods

2.3.1. Screening and Optimization of Environmental Predictors

To construct an independent and ecologically interpretable predictor set, a two-stage variable convergence strategy combining relevance screening and multicollinearity control was implemented. First, the Boruta algorithm was applied to evaluate the relevance of each environmental predictor to SOC. Boruta was selected because it is well suited to high-dimensional environmental datasets and can account for nonlinear relationships and interactions among predictors, thereby providing a robust basis for subsequent model construction and interpretation [34]. Boruta compares the importance of original predictors with that of randomly permuted shadow variables [12]. For each predictor $X_i$, its importance score is ${I(X}_i)$ compared with the maximum importance of shadow features $I_{shadow}^{max}$. A predictor was confirmed as relevant when:

$${{I(X}_i)>I }_{shadow}^{max}$$

(1)

consistently across repeated iterations. This procedure retained only predictors with statistically significant explanatory power.

Second, multicollinearity among retained predictors was assessed using the Variance Inflation Factor ${VIF}_i$ [40]. For each predictor $X_i$ ${VIF}_i$ is defined as follows:

$${VIF}_i ={\displaystyle \frac{1}{1-R_i^2}}$$

(2)

where $R_i^2$ is the coefficient of determination obtained by regressing predictor $X_i$ against all remaining predictors. A high ${R }_i^{2 }$ indicates strong linear dependence on other variables. Predictors with $VIF\ge 5$ were iteratively removed, beginning with the variable exhibiting the highest ${VIF}_i$ value, while maintaining ecological representativeness [41]. The final predictor set satisfied:

$${VIF}_i <5 \forall i$$

(3)

ensuring statistical independence and model stability for subsequent analyses.

2.3.2. Random Forest Modeling and SHAP Interpretation

The RF algorithm was chosen for modeling SOC due to its ability to capture nonlinear relationships and high-order interactions among multiple environmental predictors without assuming specific distributions [42]. RF also handles high-dimensional datasets effectively and is robust to multicollinearity among predictors, which is particularly important given the 37 candidate variables in this study. Compared to alternative methods such as support vector machines, gradient boosting, or multiple linear regression, RF provides a favorable balance between predictive accuracy and interpretability. The ensemble prediction for an observation x is expressed as follows:

$$\widehat{y_i}(x)={\displaystyle \frac{1}{T}} \sum _{t=1}^T{f}_t(x)$$

(4)

where T denotes the number of trees and $f_t\left(x\right)$ represents the prediction from the t-th tree.

Model performance was evaluated using the coefficient of determination ($R^2$) and root mean square error (RMSE) [43]. The coefficient of determination was calculated as follows:

$$R^2=1-{\displaystyle \frac{{\sum }_{i=1}^n{(y_i-\widehat{y_i})}^2}{{\sum }_{i=1}^n{(y_i-\overline{y_i})}^2}}$$

(5)

where $y_i$ denotes observed SOC, ${\widehat{y}}_i$ represents predicted SOC, and $\bar{y}$ is the mean of observed SOC values. RMSE was defined as follows:

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}\sum _{i=1}^n{\left(y_i-\widehat{y_i}\right)}^2}$$

(6)

To improve model interpretability, SHAP was applied after RF model construction. SHAP was used to quantify the contribution of each factor to SOC predictions and to interpret the direction, magnitude, and nonlinear effects of the retained environmental variables [1,12]. Feature $i$, the SHAP value ${\varphi }_i$ quantifies its marginal contribution to the prediction and is defined as follows:

$${\varphi }_i=\sum _{S\subseteq N\backslash \left\{i\right\}}{\displaystyle \frac{\left|s\right|!\left(\left|N\right|-\left|S\right|-1\right)!}{\left|N\right|!}}\left[f\left(S\cup \left\{i\right\}\right)-f\left(S\right)\right]$$

(7)

where $N$ represents the set of all predictors and $S$ denotes a subset not containing feature $i$.

Global variable importance was assessed using the mean absolute SHAP value:

$$\left|{\varphi }_{ij}\right|\left|\overline{{\mathrm{\varnothing }}_{\mathrm{i}}}\right|={\displaystyle \frac{1}{n}}\sum _{j=1}^n\left|{\varphi }_{ij}\right|$$

(8)

where $\left|{\varphi }_{ij}\right|$ denotes the SHAP value of predictor i for sample j, and n is the total number of samples. This metric reflects the average contribution magnitude of predictor i across all observations, and SHAP dependence patterns were examined to characterize nonlinear SOC responses to dominant environmental drivers.

SHAP attribution was used strictly for model interpretation and identification of influential predictors, rather than for causal inference.

2.3.3. Spatial Prediction and Distribution Analysis of SOC

Based on the final predictor set selected through the RF–SHAP framework, a point-based analytical dataset was constructed by extracting environmental variable values at the coordinates of SOC sampling locations using the Sample tool in ArcGIS 10.2 (Esri, Redlands, CA, USA). This dataset, linking observed SOC values with their corresponding environmental predictors, was used to train an RF model for predicting SOC distribution across the YRSR [9,41]. To evaluate differences among vegetation types, the predicted SOC raster was overlaid with land cover data, and zonal statistical analysis was performed to extract descriptive statistics [44,45].

3. Results

3.1. Characteristics of Soil Organic Carbon Measurements

SOC at a depth of 0–20 cm in the YRSR exhibited considerable variability, ranging from 2.96 to 169.32 g/kg (

Table 2. Descriptive statistics of soil organic carbon at 0–20 cm in the YRSR.

Variable	Unit	N	Min	Mean	Median	Max	MAD	SD	CV%	Skewness	Kurtosis
SOC	g kg−1	240	2.96	46.61	40.75	169.32	17.94	30.29	64.97	1.22	4.85

). The mean and median values were 46.61 g/kg and 40.75 g/kg, respectively, indicating moderate central tendency but substantial dispersion. The coefficient of variation (CV) was 64.97%, reflecting high relative variability across sampling sites. The SOC distribution was positively skewed (skewness = 1.22) and leptokurtic (kurtosis = 4.85), suggesting a right-skewed and heavy-tailed pattern. Most observations clustered in lower SOC ranges, while a few high-value samples extended the upper tail of the distribution. This distribution pattern indicates that localized high-SOC patches disproportionately contribute to regional variability. Spatially, elevated SOC concentrations were concentrated in discrete areas, whereas lower SOC values dominated the broader regional background. This patch-matrix configuration implies that SOC accumulation is strongly regulated by localized hydrothermal and ecological conditions, rather than by uniform regional controls. The observed non-normality and spatial concentration patterns underscore the importance of employing spatially explicit and nonlinear approaches to resolving SOC variability in alpine environments.

3.2. Identification of Key Environmental Predictors for SOC

The Boruta algorithm was applied to assess the relevance of 37 candidate environmental variables to SOC in the YRSR. After iterative comparison with randomly permuted shadow variables, 27 predictors were identified as significant contributors to SOC variability, with their importance values significantly exceeding those of the random shadow variables (Figure 3). These retained variables encompassed key environmental factors, including thermal conditions (e.g., land surface temperature), moisture-related factors (e.g., precipitation and soil moisture), vegetation productivity (e.g., Leaf Area Index (LAI), Gross Primary Productivity (GPP)), and soil properties (e.g., pH, bulk density). To control for multicollinearity among the retained predictors, the Variance Inflation Factor (VIF) was applied. Predictors with VIF values greater than 5 were iteratively removed until all remaining variables exhibited acceptable VIF values (<5), ensuring statistical independence and reducing potential model instability (

Table 3. Variance inflation factors (VIF) of predictors retained after collinearity screening.

Variable	VIF
LAI_MaySep	4.98
GPP_MaySep	4.95
NDWI	4.37
Terra_LsTN	3.96
Soil sand content	3.74
Terra_LsTD	3.67
SnowFreq_OctApr	3.66
SnowFreq_ year	3.53
AI	3.40
Precip_MaySep	3.10
FTD	2.37
pH	2.23
BD	1.55
Slope	1.40
SM	1.24
Night Light	1.13

). The final predictor set included 16 variables, which adequately represent the dominant environmental gradients shaping SOC spatial variability in the region. These variables will serve as robust inputs for subsequent nonlinear modeling, offering a comprehensive, ecologically interpretable basis for assessing SOC in the YRSR.

3.3. Nonlinear Importance Patterns of SOC Revealed by RF–SHAP

The random forest (RF) model demonstrated robust predictive performance, explaining 68% of the variance in soil organic carbon (SOC), with an RMSE of 15.59 g/kg (Figure S1). These results indicate that the optimized environmental predictors effectively captured the primary spatial gradients in the YRSR. To further interpret the RF model, SHAP was used to quantify the contribution and effect patterns of individual predictors on SOC, and the corresponding interpretation is presented in Figure 4. As shown in Figure 4, vegetation productivity indicators, particularly LAI and GPP, were the most influential predictors, with high SHAP values indicating their strong contributions to SOC accumulation. Their effects were clearly nonlinear, with SOC increasing with vegetation productivity and then tending to level off beyond certain ranges. Moisture-related variables, such as NDWI and SM, also showed substantial contributions, with SHAP results indicating that moderate to high moisture conditions generally promoted SOC accumulation, whereas drier conditions were associated with lower SOC values. Thermal variables, especially daytime land surface temperature (Terra_LsTD) and aridity index (AI), further exhibited significant and nonlinear effects on SOC variability. Lower temperatures generally corresponded to higher SOC levels, likely because reduced microbial activity slowed organic matter decomposition, whereas strong thermal fluctuations, particularly those associated with freeze–thaw processes, tended to decrease SOC stability. Overall, the RF–SHAP results demonstrated that SOC regulation in the YRSR is governed by nonlinear and context-dependent interactions among vegetation productivity, hydroclimatic conditions, and thermal factors.

3.4. Spatial Heterogeneity and Vegetation-Type Differences in SOC in the YRSR

SOC in the Yellow River Source Region exhibited pronounced spatial heterogeneity in 2023 (Figure 5). The mean SOC during the study period was 69.2 g/kg. Overall, SOC showed a clear spatial gradient, decreasing from southeast to northwest. High SOC values were primarily concentrated in the southeastern part of the region, whereas relatively low SOC levels were observed in the northwestern areas. Significant differences in SOC were identified among vegetation types. Mountain meadow exhibited the highest mean SOC (85.63 g/kg), representing the most carbon-rich ecosystem in the region. Forest land, improved grassland, lowland meadow, and wetlands also showed relatively high SOC levels, with mean values exceeding 76 g/kg. Alpine meadow and alpine steppe had intermediate SOC levels, with mean values of 70.63 g/kg and 53.12 g/kg, respectively. Temperate steppe showed a mean SOC of 58.75 g/kg, while temperate desert grassland exhibited the lowest SOC content (42.40 g/kg). In terms of variability, alpine meadow and wetland displayed relatively large SOC ranges, with standard deviations of 17.72 g/kg and 16.82 g/kg, respectively, indicating stronger spatial fluctuations. In contrast, improved grassland and lowland meadow showed smaller standard deviations, suggesting a more uniform spatial distribution of SOC. Overall, SOC in the Yellow River Source Region demonstrated clear vegetation-type differentiation and a distinct spatial gradient, reflecting substantial heterogeneity in carbon storage among different grassland ecosystems.

4. Discussion

4.1. Spatial Heterogeneity of Soil Organic Carbon in the Yellow River Source Region

Surface SOC in the YRSR exhibited pronounced spatial heterogeneity, consistent with observations from other alpine regions. SOC values ranged from 2.96 to 169.32 g/kg, with a mean of 46.61 g/kg and a high coefficient of variation (64.97%), indicating substantial spatial variability across the study area. The right-skewed and heavy-tailed distribution further suggests that a limited number of high-SOC sites contribute disproportionately to the regional SOC pattern [46,47].

This spatial structure implies that SOC accumulation in the YRSR is not governed by uniform regional controls, but by localized environmental conditions that favor carbon storage. In particular, moist grasslands, wetlands, and other hydrothermally favorable patches are likely to function as local SOC hotspots, whereas drier or more exposed upland areas tend to maintain lower SOC levels. This interpretation is consistent with previous studies showing that alpine SOC often exhibits strong patchiness in response to moisture gradients, temperature constraints, and vegetation heterogeneity [48,49]. But our results further suggest that such patchiness may be especially pronounced in alpine headwater landscapes like the YRSR, where wetland grassland mosaics and strong hydrothermal contrasts intensify local SOC differentiation [50].

These findings highlight the need to interpret regional SOC not as a spatially uniform property, but as a heterogeneous carbon pool shaped by localized hydrothermal and ecological conditions. From a management perspective, identifying and conserving high-SOC patches may be particularly important for maintaining carbon stability and ecological resilience in climate-sensitive alpine source regions.

4.2. Hydrothermal and Vegetation Controls on SOC Variability: Key Drivers and Nonlinear Interactions

Vegetation productivity, moisture availability, and thermal conditions emerged as the dominant controls on SOC variability in the YRSR [51]. In the present study, vegetation-related variables such as LAI and GPP were positively associated with SOC accumulation, likely because greater biomass production increases litter input, root-derived carbon input, and the overall supply of organic substrates to the soil [52]. Moisture-related variables, including precipitation and soil moisture, also played major roles in SOC regulation, but their effects were clearly nonlinear. In alpine headwater environments such as the YRSR, moisture influences SOC through several interconnected pathways. Adequate soil moisture promotes vegetation growth and carbon input, while also affecting soil aeration, microbial activity, and decomposition intensity [4]. Under relatively cool and moist conditions, SOC accumulation is favored because plant-derived carbon input remains sufficient while decomposition is constrained. Under dry conditions, by contrast, reduced vegetation productivity and limited organic matter input may restrict SOC formation [53,54].

Thermal conditions, especially surface temperature and freeze–thaw dynamics, were likewise important contributors to SOC variability [55]. Temperature may affect SOC directly by regulating microbial decomposition and carbon turnover, and indirectly by altering plant growth, soil moisture status, and freeze–thaw intensity [56,57]. Lower temperatures were generally associated with higher SOC concentrations, likely because microbial activity and decomposition were suppressed. However, frequent freeze–thaw cycles may destabilize soil aggregates, alter water availability, and accelerate organic matter turnover, thereby reducing SOC stability in some locations [58,59].

A notable feature of SOC regulation in the YRSR is that these hydrothermal and biotic controls operate nonlinearly rather than as simple independent effects. SOC increased with vegetation productivity only up to a certain level, after which the relationship weakened or plateaued, suggesting that carbon input alone does not guarantee continuous SOC accumulation because stabilization processes, substrate saturation, and decomposition may increasingly constrain additional SOC storage at higher productivity levels [60,61]. A similar threshold-like pattern was observed for moisture: moderate to high moisture availability generally favored SOC accumulation, whereas dry conditions were associated with lower SOC levels [62]. These results indicate that hydrothermal regulation of SOC in the YRSR is highly context-dependent. In cold and moist environments, low temperatures may suppress decomposition while sufficient moisture sustains vegetation-derived carbon input, thereby favoring SOC accumulation. In warmer or drier conditions, this balance may shift because plant inputs decline and microbial turnover increases. This difference may reflect the environmental setting of the region, where permafrost influence, strong moisture gradients, and wetland–grassland mosaics intensify the interaction between temperature, moisture, and vegetation productivity.

4.3. Ecological and Management Implications of SOC Spatial Heterogeneity for Alpine Grassland Management

The pronounced spatial heterogeneity of SOC in the YRSR has important environmental and management implications. Wetlands, moist grasslands, and other hydrothermally favorable areas are likely to function as local SOC hotspots, making them disproportionately important for regional carbon storage and ecological stability. In contrast, drier and more exposed alpine grasslands generally contain lower SOC stocks and may be more vulnerable to carbon loss under ongoing climate warming [63]. The strong dependence of SOC on moisture availability and vegetation productivity further suggests that changes in hydrological conditions, freeze–thaw activity, and plant growth may substantially alter carbon stability in alpine headwater landscapes [13,64]. In the YRSR, warming-induced shifts in soil moisture or freeze–thaw dynamics could reduce SOC persistence by weakening plant-derived carbon inputs while accelerating organic matter decomposition [65].

These findings support the need for spatially differentiated management in alpine source regions of the Tibetan Plateau. In particular, the conservation of wetlands, seasonally moist grasslands, and other high SOC areas may be especially important for maintaining both carbon stability and ecosystem resilience under climate change [8,66]. More broadly, identifying and protecting SOC-rich patches may contribute not only to soil carbon conservation, but also to the broader goals of grassland sustainability, water source protection, and climate adaptation in cold and environmentally sensitive ecosystems [19,67].

In addition, machine learning methods such as RF offer clear advantages when integrated with multi-source remote sensing data, particularly because they can capture nonlinear relationships between SOC and environmental predictors. Our results highlight the value of combining multi-sensor remote sensing data with advanced modeling approaches to better characterize SOC dynamics across regions with strong environmental gradients [3,43]. Future research should further explore the integration of high-resolution optical, thermal, and radar remote sensing data with machine learning models to improve SOC prediction in remote and data-scarce regions such as the Tibetan Plateau.

4.4. Study Limitations and Future Perspectives

While this study highlights the importance of hydrothermal conditions and vegetation productivity in regulating shallow SOC, SOC should not be viewed as the sole indicator of soil ecological functioning [51,68]. Soil biogeochemical regulation is jointly shaped by carbon, nitrogen, plant-available nutrients, macro- and microelements, hydrological processes, and vertical heterogeneity within the root zone. Another limitation of this study is that only surface soil layers (0–20 cm) were analyzed. Deeper soil horizons, which may store substantial amounts of carbon, were not included, thereby limiting our understanding of the full soil carbon profile and its long-term stability [69]. In addition, SOC data were collected during a single growing season (July–August 2023). Consequently, the findings represent only a temporal snapshot and may not fully capture seasonal or interannual variability in SOC distribution. Future research should incorporate deeper soil layers, as well as multi-year and multi-season field observations, to provide a more comprehensive understanding of SOC dynamics and to improve the generalizability of these results.

Although the RF SHAP framework was effective for identifying dominant predictors and capturing nonlinear SOC environment relationships, it remains fundamentally correlational. Variable importance may partly reflect covariance among hydrothermal, vegetation, and terrain factors, and the model cannot by itself establish causal pathways [70]. Therefore, the ecological mechanisms discussed above should be interpreted as plausible explanations supported by the observed spatial patterns, rather than as direct causal proof. In addition, microclimatic influences were not measured directly at each sampling site, but were represented indirectly through spatial proxies of near-surface hydrothermal conditions, including land surface temperature, freeze–thaw cycle days, soil moisture, and snow-related variables [2]. These proxies capture important aspects of local energy and water regulation, but they cannot fully resolve fine-scale microclimatic heterogeneity beneath plant canopies or across microsites.

The broader relevance of these findings lies in their potential applicability to other alpine regions of the Tibetan Plateau and to similar cold, moisture-sensitive ecosystems. In particular, many alpine grasslands across the Tibetan Plateau are also shaped by permafrost-related processes, freeze–thaw activity, moisture limitation, and strong spatial heterogeneity in plant productivity. From this perspective, the general ecological mechanisms identified here may be transferable to regions with comparable climatic and landscape settings. Nevertheless, direct spatial extrapolation should be approached with caution. The strength and form of SOC–environment relationships may differ across regions because of variation in soil type, permafrost extent, vegetation composition, topographic complexity, and land-use or grazing intensity [65]. Therefore, the main scalable contribution of this study is not a direct transfer of model parameters, but an interpretable analytical framework that combines field observations, multi-source environmental predictors, and RF SHAP analysis to identify dominant SOC controls in environmentally heterogeneous alpine systems.

5. Conclusions

This study quantified the spatial heterogeneity of surface soil organic carbon (SOC) in the Yellow River Source Region, identifying the dominant hydrothermal and biotic controls using an integrated RF–SHAP framework. The results revealed significant spatial variability in SOC, with moisture-related processes acting as a central mediator between thermal conditions, vegetation productivity, and soil properties. Thermal regimes and freeze–thaw dynamics directly influenced SOC, while vegetation productivity and soil properties primarily affected SOC through moisture-mediated pathways. These findings underscore the hierarchical and nonlinear nature of SOC regulation in alpine headwater ecosystems. From a sustainability perspective, maintaining favorable hydrothermal conditions is crucial for stabilizing soil carbon stocks. Prioritizing wetland conservation, adaptive grazing management, and reducing soil disturbance in high-SOC areas will help mitigate carbon loss, especially under climate warming. The integrated modeling approach developed in this study provides a transferable method for identifying vulnerable carbon pools and supporting sustainable land management in other alpine and climate-sensitive regions.