Quantifying climatic drivers and identifying spatial risks of vegetation carbon sink instability: a case study of central–western Inner Mongolia

Abstract

Vegetation carbon sinks are critical for the global carbon budget, yet their stability in vulnerable ecosystems remains poorly understood. This gap is pronounced in central–western Inner Mongolia due to the nonlinear responses of vegetation carbon sinks to climatic factors and spatial heterogeneity. This study integrates time-series analysis, XGBoost–SHAP, and an optimized MaxEnt model, utilizing MODIS/Terra MOD17A3HGF net primary productivity (NPP) data, interpolated climate data from the China Meteorological Data Service Centre, and future climate data from the BCC-CSM2-MR model, to identify climatic regulators of vegetation carbon sinks and project their spatial risk patterns. Key findings include: (1) During 2014–2023, regional net ecosystem productivity (NEP) showed a slight upward trend with noticeable interannual fluctuations; comparing 2014–2023, approximately 25,500 km² of vegetated area switched between carbon sink and source states. (2) Temperature-related interannual variability accounted for approximately 62% of NEP fluctuations, with the variability of maximum temperature of the warmest month (bio5_IVA) contributing the most (19.4%). Precipitation-related variability contributed the remaining 38%, and interactions between temperature and precipitation jointly shaped ecosystem instability. (3) High-risk and extreme-risk areas clustered along desert margins, with risk levels declining outward. Risk patterns under SSP126 (+1.76 °C by 2040 relative to 2023) remained stable, whereas SSP245 (+2.28 °C) and SSP370 (+2.53 °C) induced widespread risk-level transitions, highlighting the acute vulnerability of ecological transition zones to climate forcing. Although recent studies highlight the important roles of vapor pressure deficit and soil moisture in regulating dryland carbon sinks, these factors were not directly quantified in this study. Future efforts will integrate them to refine the predictions. This framework will provide a mechanistic basis for early warning and targeted management of vegetation carbon sinks in vulnerable ecosystems.

1 Introduction

Vegetation carbon sinks play a critical role in the global carbon cycle, with their spatiotemporal stability being essential for accurate global carbon budgets and climate adaptation (Guo et al., 2025; Hu et al., 2022; Liu et al., 2022). Carbon sink stability generally refers to the capacity of an ecosystem to maintain a relatively persistent carbon sequestration function under natural or anthropogenic disturbances. It is a multidimensional concept involving temporal variability, resistance, resilience, post-disturbance recovery, and long-term trend persistence (Wu et al., 2026; Shi et al., 2025). In this study, vegetation carbon sink instability is operationally defined from two directly observable aspects: temporal variability and functional state transition. Temporal variability refers to pronounced interannual fluctuations in net ecosystem productivity (NEP), indicating unstable carbon sink strength over time. Functional state transition refers to a shift between carbon sink and carbon source states, particularly a transition from positive NEP to negative NEP, which indicates a change from net carbon uptake to net carbon release. Accordingly, carbon sink risk refers to the modeled likelihood that vegetation functioning as a carbon sink will shift to a carbon source state under given climatic conditions.

Drylands cover approximately 41% of Earth’s land surface and contribute substantially to both the trend and interannual variability of the global terrestrial carbon sink. However, the stability of vegetation carbon sinks in these water-limited ecosystems remains critically understudied, despite their vulnerability to climate change and human interventions. In China, drylands span 6.6 million km² and support nearly 580 million people, yet they face increasing risks of desertification and ecosystem degradation (Li et al., 2021). Central–western Inner Mongolia, located in an arid/semi-arid transition zone, is a typical example, featuring fragmented forests and degraded grasslands. As a major carbon sink area in China, it contributed over 9% of the national total in 2022. However, its vegetation carbon sinks have experienced pronounced fluctuations in recent decades, emerging as a hotspot of instability (You et al., 2023; Shi et al., 2025). These dynamics highlight the urgent need to clarify the climatic mechanisms underlying carbon sink variability and to identify high-risk regions of sink-to-source transitions.

In the context of climate change, understanding regional climatic impacts on vegetation productivity is increasingly important (Xu et al., 2023; Peng et al., 2022). Previous studies indicate that vegetation carbon sinks in Inner Mongolia are primarily driven by precipitation and temperature (Dai et al., 2016; Xie and Ren, 2025; Wu et al., 2025). However, dominant factors vary under different baseline conditions and vegetation types (Meng et al., 2023), and many studies have overlooked the influence of seasonal variability and extreme climatic factors. Moreover, most research has relied on traditional statistical methods, such as correlation analysis (Dong et al., 2024; Lyu et al., 2020; Nie et al., 2025) and multiple linear regression (Qi et al., 2024), which are limited in capturing nonlinear interactions and complex dependencies among variables (Yu et al., 2024; Wang et al., 2023). Machine learning methods demonstrate superior performance in modeling nonlinear systems, selecting high-dimensional features, and interpreting variable importance (Li et al., 2025a). The XGBoost algorithm iteratively constructs weighted decision trees to capture complex data patterns, offering advantages such as high computational efficiency, tolerance for missing data, and robustness against multicollinearity (Pande et al., 2024; Qu et al., 2025). Nevertheless, as model complexity increases, interpretability declines, leading XGBoost to be often regarded as a “black-box” model (Zhou et al., 2022a). To address this limitation, SHapley Additive exPlanations (SHAP) provides a transparent means of quantifying the marginal contribution of each input variable to model outputs (Li et al., 2025b). In recent years, the XGBoost–SHAP framework has shown considerable potential in explaining nonlinear relationships and elucidating the mechanisms driving ecological processes (Zhang et al., 2023a; Yuan et al., 2024).

In addition, effective carbon sink risk management in ecologically vulnerable regions is critical for enhancing system stability. Global climate change has further accelerated carbon sink instability, making ecosystems more vulnerable to disturbances and increasing the likelihood of negative feedback loops (Yuan et al., 2025). Existing approaches have employed various methods to assess risk. For example, Wang et al. (2024) applied landscape metrics combined with multi-criteria decision analysis to identify and classify regional carbon sink risk zones, while Zhang et al., 2023b used the Hurst index to evaluate the persistence and long-term stability of carbon sink time series. However, such approaches are often constrained by static evaluation frameworks, subjective parameter weighting, and limited spatial predictive capacity. To overcome these limitations, a more dynamic and spatially explicit modeling approach is required. The Maximum Entropy (MaxEnt) model has emerged as a promising alternative. It is a powerful machine learning tool capable of predicting the spatial distribution of ecological risks based on environmental covariates and known occurrence data. It has been widely applied in risk prediction for habitat loss, urban flooding, and forest fires (Göltas et al., 2024; Li et al., 2024a; Li et al., 2024b). Its potential for assessing carbon sink risk is increasingly recognized. Furthermore, the integration of scenario-based simulations with the MaxEnt framework allows for forecasting future carbon sink dynamics under climatic and socioeconomic uncertainties, thereby providing a robust scientific basis for adaptive land management (Wu et al., 2024).

Unlike previous studies that mainly described carbon sink patterns or examined linear correlations between vegetation productivity and climatic variables, this study addresses a more specific and unresolved problem: how climatic variability drives the instability of vegetation carbon sinks and how such instability can be translated into spatially explicit risk warnings under future climate scenarios. The contribution of this study lies primarily in the development of an integrated assessment framework that links carbon sink/source transition detection, interpretable machine-learning analysis of climatic drivers, and spatial risk prediction. Within this framework, we further address two empirical questions: (1) whether interannual variability in temperature- and precipitation-related factors produces nonlinear and threshold-dependent responses in NEP fluctuations; and (2) where current and future high-risk zones of sink-to-source transition are likely to emerge in the arid and semi-arid transition zone of central–western Inner Mongolia. Therefore, this study does not merely identify risk areas, but connects process-based carbon sink dynamics, interpretable climatic response mechanisms, and scenario-based spatial risk forecasting. (Figure 1). Specifically, the objectives are to: (1) characterize the spatiotemporal dynamics of vegetation carbon sinks and sink-source transitions from 2014 to 2023; (2) quantify the impacts of climatic factors on carbon sink variability, with a focus on interannual variability amplitude (IVA), using the advanced XGBoost–SHAP interpretable framework; and (3) predict current and future high-risk zones of sink-to-source transition by integrating the key climatic drivers into an optimized MaxEnt model under CMIP6 climate scenarios.

FIGURE 1

2 Materials and methods

2.1 Study area

The study area covers central–western Inner Mongolia, including Hohhot, Baotou, Ulanqab, Ordos, and Xilingol League, with a total area of 386,000 km² (106°12′–116°30′E, 37°43′–45°30′N). Situated within a typical ecological transition zone in northern China, the region is characterised by a temperate arid to semi-arid climate, with pronounced spatial variability in annual precipitation (Figure 2). Ecosystems are highly sensitive to climate disturbances, with drought, desertification, and land degradation posing major challenges. Dominated by desert steppe and typical steppe vegetation, the region is a key component of China’s carbon sink and a globally vulnerable ecosystem. Understanding carbon sink fluctuations and associated risks here provides critical insights for managing vegetation carbon sinks and guiding ecological restoration in fragile northern regions.

FIGURE 2

2.2 Carbon sink assessment

To evaluate the role of ecosystems in carbon cycling, net ecosystem productivity (NEP) was used to characterize the carbon sink/source status of vegetation ecosystems (Li et al., 2020a; Zhang et al., 2022). Before NEP calculation, all input datasets were preprocessed to ensure spatial and temporal consistency. Annual NPP was derived from the MODIS/Terra MOD17A3HGF product and converted from g Cm^-2 yr^-1 to kg Cm^-2 yr^-1. Mean annual temperature and annual precipitation were calculated from interpolated monthly meteorological data, and precipitation was converted from millimeters to meters to match the requirements of the soil respiration model. All raster datasets were projected, resampled, and aligned to the same spatial grid, and non-vegetated areas were masked using land-cover data. The calculation can be expressed as Equation 1:where NPP represents net primary productivity and R_h represents heterotrophic respiration, both expressed in kg Cm^-2 yr⁻¹. A positive NEP indicates net carbon uptake by the ecosystem and is classified as a carbon sink, whereas a negative NEP indicates net carbon release and is classified as a carbon source.

The empirical model was adopted to estimate annual soil respiration (R_s) (Chen et al., 2012). This empirical model was selected because it was developed and validated using observations from diverse Chinese terrestrial ecosystems, including grasslands, and explicitly incorporates the effects of temperature, precipitation, and soil organic carbon on soil respiration. These factors represent the dominant abiotic controls on soil carbon emissions in temperate arid and semi-arid ecosystems, making the model suitable for application in central–western Inner Mongolia. R_s was calculated according to Equation 2:where R_s represents the annual soil respiration (kg Cm^-2 yr^-1), T represents the mean annual temperature (°C), P represents the mean annual precipitation (m), and SOC is topsoil organic carbon density at 0–20 cm. Following Xu et al. (2019), SOC was fixed at 2.74 kg Cm^-2 for the temperate semi-arid region. Model parameters were estimated by nonlinear least squares: R₀ = 1.55, Q = 0.031, K = 0.68, and M = 2.23. Here, Q captures the exponential temperature dependence of R_s, while K and M are half-saturation constants for precipitation and SOC, respectively.

Heterotrophic respiration (R_h) depends on soil temperature, moisture, organic matter, oxygen availability, and pH, complicating its accurate quantification (Xie et al., 2014). To address this, the present study employed a power function regression model (Zhang et al., 2020a), which relates R_h on the basis of R_s. This relationship was fitted using 113 paired Rs–Rh observations from forests, croplands, and grasslands across China and has shown broad applicability for estimating the heterotrophic component of soil respiration. R_h was calculated as Equation 3:

2.3 Sen’s slope and the Mann–Kendall test

Vegetation NEP trends in central–western Inner Mongolia were evaluated using Sen’s slope and the Mann–Kendall (MK) test. Sen’s slope (Equation 4) quantifies the magnitude of a monotonic trend as the median of all pairwise slopes (Kumar et al., 2016).where x_i and x_j are NEP at times i and j, respectively, D_ij is the pairwise slope; and D is Sen’s slope.

Trend direction and significance were assessed using the non-parametric MK test. The test statistic S is defined and calculated as follows (Equation 5):where n is the sample size. The standardized statistic Z is given by Equation 6:

The tie-corrected variance of S is given by Equation 7:where t_p denotes the number of tied groups (i.e., equal data points) in the time series.

2.4 XGBoost–SHAP model

The XGBoost algorithm was applied to quantify the influence of climatic factors on NEP fluctuations. The dataset was randomly divided into 80% for training and 20% for testing. Key hyperparameters, including the number of base learners (n_estimators), learning rate (learning_rate), and maximum tree depth (max_depth), were optimized using a randomized search procedure. Model performance was evaluated using the mean absolute error (MAE), root mean square error (RMSE), and the coefficient of determination (R²). The optimized model demonstrated satisfactory predictive accuracy and reliability (Supplementary Table S1).

SHAP was used to interpret the XGBoost outputs. For each sample, SHAP values were computed to quantify the contribution of individual variables, enabling both global assessment of variable importance and local explanation of predictions (Ariza-Garzón et al., 2020). The SHAP value of variable i is defined as Equation 8:where N represents the full set of input features, S denotes a subset of N that excludes feature i, and f(S) is the model output when only the features in subset S are present. The term f(S∪{i})−f(S) represents the marginal contribution of feature i when it is added to subset S.

2.5 Optimized MaxEnt model

Vegetation areas that experienced carbon sink-to-source transitions during 2014–2023 were used to generate 1,000 presence points via random sampling in ArcGIS 10.8. To reduce the risk of multicollinearity, Pearson correlation analysis was conducted for climatic variables (Supplementary Figure S1), and highly correlated ones (|r|>0.8) were excluded (Yang et al., 2025). To address MaxEnt’s sensitivity to sampling bias and model complexity, which are influenced primarily by feature classes (FCs) and the regularization multiplier (RM), systematic parameter tuning was performed via the ENMeval package in R (v4.2.3). Six FC combinations were evaluated: L (linear), LQ (linear + quadratic), LQH (linear + quadratic + hinge), H (hinge), LQHP (linear + quadratic + hinge + product), and LQHPT (linear + quadratic + hinge + product + threshold). Additionally, eight RM values ranging from 0.5 to 4.0 (in increments of 0.5) were tested. The optimal model configuration was selected based on the lowest corrected Akaike information criterion (delta AICc < 2) (Kass et al., 2021; Li et al., 2020b).

A partition of the dataset allocated 75% to training and the remaining 25% to testing, and the model was evaluated over 10 replicates. All other parameters followed MaxEnt defaults. Model performance was evaluated using the area under the ROC curve (AUC). The logistic output of the optimized MaxEnt model, ranging from 0 to 1, was used as the carbon sink risk index, with higher values indicating a greater likelihood of sink-to-source transition. The risk index was classified into five equal-interval levels: the risk-free zone (0–0.2), the low-risk zone (0.2–0.4), the moderate-risk zone (0.4–0.6), the high-risk zone (0.6–0.8), and the extreme-risk zone (0.8–1.0). In this classification scheme, the moderate-risk zone was regarded as a carbon sink risk transition zone, indicating an intermediate state between relatively stable carbon sink areas and high-risk areas with a greater probability of carbon source conversion. It therefore served as a model-based buffer and early-warning zone for identifying areas sensitive to future risk escalation. The same thresholds were applied to both current and future scenario projections to ensure comparability of risk levels across different climate scenarios.

2.6 Data resources

Based on relevant studies (Pang et al., 2024; Huang et al., 2024) and considering the climate sensitivity of the study area, seven climatic factors were selected. Historical climate data (bioclimatic variables) were derived from monthly observations at 51 meteorological stations (China National Meteorological Data Centre) and spatially interpolated. Net primary productivity (NPP) data for 2014–2023 were obtained from the MODIS/Terra MOD17A3HGF Version 6.1 product (NASA Earthdata). Future climate data were derived from the BCC-CSM2-MR model within the CMIP6 Scenario MIP framework, covering three representative scenarios: SSP126 (low emissions), SSP245 (medium emissions), and SSP370 (high emissions). Monthly near-surface temperature (tas, tasmax, tasmin) and precipitation (pr) data (2021–2040) were processed into bioclimatic variables for simulating future carbon sink/source patterns. Land cover data were obtained from the China Land Cover Dataset (1986–2023; Yang and Huang, 2021). Detailed descriptions of all data sources are summarized in Table 1.

To better capture the effects of climatic variability on NEP, we calculated the multi-year mean of absolute year-to-year differences for each variable from 2014 to 2023 and defined it as the interannual variability amplitude (IVA). This indicator reflects the magnitude of interannual fluctuations, reduces short-term noise, and emphasizes long-term trends and stable associations with NEP responses. The calculation is expressed as Equation 9:where R_IVA is the interannual variability amplitude for a given variable, x_t is the observed value in year t, and T is the total number of years.

3 Results

3.1 Dynamic characteristics of carbon sink/source transitions

During 2014–2023, the annual mean NEP across the study area ranged from 0.117 to 0.148 kg Cm^-2·yr^-1, with a multi-year average of 0.132 kg Cm^-2·yr^-1. To evaluate the reliability of the NEP estimates, comparisons were made with published regional-scale results. The multi-year mean NEP estimated in the present study was generally comparable to the value reported by Cui et al. (2025) for Inner Mongolia during 2001–2021 (0.168 kg C·m^-2·yr^-1), as well as that reported by Li et al. (2022) for Inner Mongolian grasslands during 2010–2020 (0.180 kg C·m^-2·yr^-1). The differences among studies are expected given variations in spatial coverage, vegetation composition, study period, input datasets, and NEP estimation approaches. Vegetation functioning as a carbon sink (NEP > 0) dominated the central–western region of Inner Mongolia (Figure 3). Sink areas consistently exceeded 84% of the landscape, reaching a maximum of 95.19% in 2020. Interannual variability was apparent, with marked declines to 86.08% in 2017 and 84.48% in 2023. In contrast, carbon source areas (NEP < 0) were relatively constrained, ranging between 4.81% and 15.52%. Pronounced expansions in 2017 and 2022 were largely attributable to climatic disturbances such as prolonged spring–summer droughts and sandstorms.

FIGURE 3

FIGURE 4

Sen’s slope (median = 0.001) shows a weak upward trend in NEP during 2014–2023. To characterize spatial patterns, NEP trends were classified into five categories based on Z values: significant increase, slight increase, no significant change, slight decrease, and significant decrease (Figure 5; Table 2). Overall, 78.26% of the area exhibited a slight increase and 6.06% a significant increase, mainly in northern Xilingol, southern Ulanqab, and along the Great Bend of the Yellow River near the Hobq Desert. In contrast, only 0.21% showed a significant decrease, concentrated in western Xilingol, northern Ulanqab, and northeastern Hohhot. An additional 8.43% experienced a slight decrease, generally forming diffuse belts surrounding the significantly declining patches.

FIGURE 5

3.2 Driving mechanisms of vegetation NEP fluctuations

The XGBoost–SHAP analysis demonstrated that temperature factors were the dominant drivers of NEP variation, with bio1_IVA, bio4_IVA, bio5_IVA, and bio6_IVA together accounting for nearly 62% of the total contribution. Precipitation factors (bio12_IVA, bio13_IVA, and bio15_IVA) together accounted for the remaining 38%. The SHAP beeswarm plot shows that key variables exhibit a wide distribution of SHAP values, indicating substantial variability in their contributions to NEP. Both positive and negative SHAP values are observed across variables, suggesting complex and nonlinear effects on NEP (Figure 6). Model evaluation confirmed reliable predictive capacity (R² = 0.701, RMSE = 10.808, MAE = 5.375), supporting the robustness of the identified variable importance patterns.

FIGURE 6

The SHAP dependence plots highlighted distinct response patterns of NEP variability to different climate variability factors (Figure 7). For bio1_IVA (°C/yr), NEP variability was promoted below 0.43. The SHAP values were evenly distributed around zero within the range of 0.43∼0.55. Beyond 0.55, suppression became dominant, whereas beyond 0.57, the promotive effect gradually re-emerged and strengthened. For bio4_IVA, NEP variability was mainly suppressed below 71, and the promotive effect gradually strengthened once exceeding this threshold. For bio5_IVA (°C/yr), NEP variability was mainly suppressed below 0.8, shifted to promotion between 0.8 and 1.08, and became suppressed again beyond 1.08. For bio6_IVA (°C/yr), NEP variability was mainly promoted between 2.4 and 3.4, and suppressed beyond 3.4. For bio12_IVA (mm/yr), NEP variability was mainly promoted below 58, and suppressed beyond this threshold. For bio13_IVA (mm/yr), NEP variability was suppressed below 29, promoted within 29∼42, and suppressed beyond 42. For bio15_IVA, NEP variability was promoted below 16, suppressed between 16 and 24, and promoted again beyond 24.

FIGURE 7

The SHAP interaction matrix revealed four factor pairs with the strongest interactions (mean|SHAP|>1.0): bio1_IVA × bio12_IVA (1.55), bio5_IVA × bio12_IVA (1.50), bio4_IVA × bio12_IVA (1.19), and bio4_IVA × bio1_IVA (1.06). These results highlight the strong thermal-hydrological coupling between temperature variability and annual precipitation variability, which jointly shape the stability of vegetation carbon sinks (Figure 8).

FIGURE 8

The SHAP interaction analysis revealed clear threshold-dependent effects among the key climatic variability factors regulating NEP variability (Figure 9).

FIGURE 9

For bio1_IVA × bio12_IVA higher bio12_IVA strengthened the positive contribution of bio1_IVA when bio1_IVA<0.43, reinforced its negative contribution at 0.43∼0.57, and then alleviated the negative contribution while promoting a return to positive values when bio1_IVA>0.57. When bio12_IVA<58, the moderating role of bio1_IVA was relatively weak; above 58, higher bio1_IVA weakened the negative contribution of bio12_IVA.

For bio5_IVA × bio12_IVA, higher bio12_IVA mitigated the negative contribution of bio5_IVA when bio5_IVA<0.80, attenuated its positive contribution at 0.80∼1.08, and reinforced its negative contribution when bio5_IVA>1.08. When bio12IVA <58, the moderating role of bio5_IVA was relatively weak; above 58, higher bio5_IVA alleviated the negative contribution of bio12_IVA as bio12_IVA increased.

For bio4_IVA × bio12_IVA, higher bio12_IVA strengthened the negative contribution of bio4_IVA when bio4_IVA<71, but shifted to enhancing its positive contribution when bio4_IVA>71. When bio12_IVA <58, the moderating effect of bio4_IVA was relatively weak; above 58, higher bio4_IVA weakened the negative contribution of bio12_IVA.

For bio1_IVA × bio4_IVA, the moderating role of bio4_IVA was relatively weak when bio1_IVA<0.55, strengthened the negative contribution of bio1_IVA at 0.55∼0.57, and became associated with a stronger positive contribution when bio1_IVA>0.57. When bio4_IVA <71, higher bio1_IVA strengthened its negative contribution; above 71, higher bio1_IVA alleviated the negative contribution and promoted a shift toward positive contribution.

3.3 Probability evaluation of vegetation carbon sink risk

3.3.1 Spatial pattern of vegetation carbon sink risk zones

The optimized MaxEnt model (fc = LQHPT, rm = 1) achieved an average training AUC of 0.927 and a testing AUC of 0.921, demonstrating excellent predictive performance and substantially reducing overfitting relative to the default configuration (Supplementary Table S2).

The predicted spatial distribution of carbon sink risk revealed pronounced geographic differentiation (Figure 10). High-risk and extreme-risk zones were mainly concentrated along the margins of the deserts, where vegetation cover is sparse and climatic variability exerts stronger destabilizing effects. These ecotone regions represent ecological “hotspots” of instability, with elevated likelihoods of sink-to-source transitions. In contrast, risk-free areas dominated the landscape, accounting for 81.70% of the study region, primarily in the southern and eastern zones where hydrothermal conditions are relatively stable. Moderate-risk zones were distributed as transitional belts between core grasslands and desert margins, indicating a gradient of increasing vulnerability. At the municipal scale, Xilingol and Ulanqab contained the largest areas of high to extreme risk, reflecting their exposure to both desertification pressures and climatic extremes. Although extreme-risk zones occupied only a small fraction of the total area, their spatial clustering in ecologically fragile belts underscores the disproportionate importance for regional carbon balance and ecological security.

FIGURE 10

3.3.2 Influence of environmental variables on risk probability

Based on the percent contribution and permutation importance results, the analysis identified bio12_IVA and bio1_IVA as the dominant drivers of carbon sink risk zones, with contributions of 25.6% and 17.4% and permutation importance values of 30.8% and 30.9%, respectively (Table 3). Bio6_IVA (22.2%) and bio4_IVA (20.3%) also contributed notably, while other variables showed relatively limited influence. The jackknife test (Figure 11h) further highlighted the strong independent predictive power of bio4_IVA and the critical importance of bio1_IVA to overall model performance.

FIGURE 11

The response curves revealed a nonlinear relationship between carbon sink risk and the interannual variability of climatic factors (Figures 11a–g). Under low variability, the curves remained relatively flat, indicating low risk and stable carbon sink functions. However, once climatic fluctuations exceeded certain thresholds, risk probability increased sharply, highlighting the presence of critical tipping points. Among temperature-related factors, bio1_IVA and bio5_IVA exhibited a rise–fall pattern, with risk levels increasing at moderate variability but declining again under more extreme fluctuations, suggesting that excessive variability may induce ecosystem imbalance or weaken carbon sink capacity. In contrast, bio4_IVA and bio6_IVA showed a sustained increase in risk beyond their thresholds, ultimately stabilizing at high levels. For precipitation-related factors, bio12_IVA emerged as the dominant hydrological driver, while bio13_IVA and bio15_IVA also displayed nonlinear responses. All three precipitation factors were associated with elevated risk under moderate-to-high variability, underscoring the pivotal role of water variability in shaping carbon sink stability in arid and semi-arid regions. While moderate precipitation fluctuations may be tolerated, once ecological thresholds are exceeded, water stress can rapidly amplify instability and intensify sink-to-source transitions.

3.3.3 Projected carbon sink risk zones under future climate scenarios

The annual mean temperature in the study area was 3.78 °C in 2023. Under different emission scenarios, both the projected warming by 2040 and the spatial dynamics of carbon sink risk zones exhibit notable divergence (Figure 12; Supplementary Tables S3–S5).

FIGURE 12

Under SSP126, with annual mean temperature projected to increase by 1.76 °C by 2040, the overall risk declined. The extreme-risk zone contracted by 1,104.25 km² (−12.7%), while the low-risk zone expanded by 2,065 km² (+9.4%). This expansion was primarily attributed to the large-scale downgrading of 2,657 km² from the moderate-risk zone, suggesting that low-emission pathways effectively buffer transitional ecotones and promote ecosystem stabilization.

Under SSP245, with annual mean temperature projected to increase by 2.28 °C by 2040, the overall spatial structure remained relatively stable, but the most pronounced bidirectional transitions occurred within the moderate-risk zone. Approximately 2,957 km² of moderate-risk areas were downgraded to low risk, while about 555 km² were upgraded to high risk. These dynamics highlight the heightened sensitivity and volatility of transitional zones under intermediate climate forcing.

Under SSP370, with annual mean temperature projected to increase by 2.53 °C by 2040, instability was most pronounced. While the extreme-risk zone contracted by 2005 km² (−23%), the moderate-risk zone experienced strong differentiation: about 4,737 km² were downgraded to low risk, whereas 529 km² were upgraded to high risk. Notably, widespread transitions occurred across multiple risk categories, with a total transition area equivalent to 16% of the baseline risk zones, which was the highest among all scenarios. These findings demonstrate that under strong climate forcing, transitional zones are not only the key areas of risk differentiation but also the most active domains of risk-level exchange, underscoring pronounced systemic volatility.

4 Discussion

This study systematically assessed the dynamics of vegetation carbon sinks and the risk of sink-to-source transitions in central–western Inner Mongolia. The analysis revealed pronounced fluctuations in carbon sink strength, clarified the dominant climatic drivers and their nonlinear thresholds, and produced spatially explicit risk maps under current and future scenarios. These findings provide new insights into the mechanisms regulating vegetation carbon sinks and their potential vulnerabilities under climate change.

4.1 Nonlinear mechanisms of climatic drivers

During the study period, vegetation carbon sinks in central–western Inner Mongolia showed a slight overall increase, consistent with Zhang et al., 2023c. This transformation is closely linked to the implementation of large-scale ecological restoration initiatives, such as the Grain-for-Green Program and national shelterbelt projects, which have effectively enhanced vegetation cover and mitigated desertification (Hao et al., 2021).

Previous studies have shown that precipitation is the dominant climatic factor shaping carbon sink patterns in arid and semi-arid regions (Zhou et al., 2022b; Zhang et al., 2020b). In contrast, our results reveal that temperature variability plays a decisive role in driving carbon sink dynamics in central–western Inner Mongolia, thereby complementing existing knowledge of the mechanisms regulating carbon sink stability. Beyond the influence of the mean annual temperature, the roles of extreme and seasonal temperature indices are also non-negligible. Specifically, bio6_IVA shows a particularly strong impact, as greater variability markedly amplified fluctuations, underscoring the pivotal role of winter extremes in regulating subsequent growing-season productivity. The nonlinear responses of bio4_IVA and bio5_IVA further revealed threshold-dependent ecosystem behavior, indicating that vegetation growth responds to climate variability only when it exceeds tolerance limits (Wang and Alimohammadi, 2012). Precipitation variability generally promoted NEP variability, with the positive effects of bio12_IVA, bio13_IVA, and bio15_IVA highlighting the importance of water availability in semi-arid areas (Chai et al., 2025). Although the individual contribution of precipitation was lower than that of temperature, our findings highlight the critical role of their interactions. In particular, the interactions between bio6_IVA and bio12_IVA and between bio4_IVA and bio15_IVA underscore the strong thermal–hydrological coupling that shapes the ecosystem carbon balance (Cui et al., 2025).

Recent studies have increasingly emphasized the importance of vapor pressure deficit (VPD) and soil moisture in regulating dryland carbon and water fluxes. Kannenberg et al. (2024) demonstrated that soil moisture is the dominant driver of gross primary productivity and evapotranspiration in U.S. drylands, with VPD playing a secondary but non-negligible role. Similarly, Zhang et al., 2023d found that in Eurasian drylands, low soil water content exerts a stronger and more sustained stress on vegetation growth than high VPD, and that the dominance of soil moisture stress has been increasing over time. Notably, precipitation and temperature are closely linked to VPD and soil moisture: precipitation directly affects soil moisture, while temperature indirectly regulates VPD and soil evaporation through its influence on atmospheric water demand. Although these two factors were not directly quantified in this study, future work will explicitly incorporate VPD and soil moisture to further refine the mechanistic understanding of dryland carbon sink stability.

4.2 Considering human activities: rationale and limitations

Although this study focused primarily on climatic drivers, vegetation carbon sink dynamics in central–western Inner Mongolia should not be interpreted as being controlled by climate alone. Previous studies have shown that human disturbances, such as grazing and enclosure, can significantly affect regional ecosystem carbon dynamics (Chang et al., 2016; Hou and Yu, 2025). Human activities may influence vegetation carbon sinks through different pathways (Liu et al., 2019). Large-scale ecological restoration programs, such as the Grain-for-Green Program, grazing prohibition, and national shelterbelt projects, can enhance vegetation cover, reduce land degradation, and promote carbon uptake. For example, ecological protection practices in Inner Mongolia have been shown to slow grassland degradation and improve ecosystem conditions (Zheng et al., 2024). In contrast, overgrazing, land reclamation, urban expansion, mining, and infrastructure development may reduce vegetation productivity, fragment habitats, and increase the likelihood of local sink-to-source transitions. Therefore, the observed NEP changes and carbon sink/source transitions are likely the combined outcome of climatic variability, land-use change, ecological restoration, and other human activities.

In this study, climate variables were selected as the primary predictors for three main reasons. First, the objective was to identify how interannual climatic variability regulates carbon sink instability and to project future risk patterns under CMIP6 climate scenarios. Climate variables therefore provide a direct and scenario-consistent basis for both historical analysis and future prediction. Second, temperature and precipitation are dominant environmental constraints in arid and semi-arid ecosystems, where hydrothermal conditions strongly regulate vegetation productivity and soil respiration. Third, consistent long-term and spatially explicit data on human activities, such as grazing intensity, ecological restoration intensity, land-management practices, and urban expansion, remain limited and are difficult to project reliably under future scenarios.

To avoid overlooking anthropogenic influences, a preliminary experiment was conducted before the formal analysis to examine the effects of several socioeconomic and human-activity proxies, including grazing intensity (GI), nighttime light index (NTL), population density (POP), and economic development level (GDP), on vegetation carbon sink variability. The results indicated that these anthropogenic factors exerted much weaker influences than climatic factors in the present framework (Figure 13), which is consistent with the findings of Tang et al., 2025. This pattern may be partly attributed to the relatively scattered distribution of towns and villages within the study area, which limits the overall regional impact of urbanization-related human activities on vegetation carbon sinks. Meanwhile, the positive effects of ecological restoration and protection programs may partly offset the negative effects of grazing pressure, land conversion, and other disturbances at the regional scale.

FIGURE 13

Nevertheless, the weaker explanatory power of these anthropogenic proxies does not imply that human activities are unimportant. Rather, it suggests that their effects may be more localized, management-dependent, and difficult to capture using coarse regional indicators. In particular, the specific contributions of large-scale ecological restoration programs, such as the Grain-for-Green Program, were not fully disentangled from climatic drivers in this study. Therefore, the risk maps produced here should be interpreted as climate-centered estimates of carbon sink instability rather than a full attribution of all possible climatic and anthropogenic drivers. Future studies should integrate high-resolution land-use change data, grazing records, restoration project boundaries, and high-resolution observations such as flux towers and field plots to better evaluate the combined effects of long-term variability and event-driven anomalies on vegetation carbon sink dynamics (Ravi et al., 2025).

4.3 Risk zoning and targeted management strategies under future scenarios

The projected divergence of carbon sink risk zones across scenarios highlights the pivotal role of emission pathways in shaping ecosystem stability. Under SSP126, risk levels exhibited a pronounced shift from medium and high categories toward lower categories, leading to an overall decline in risk. This suggests that low-emission pathways can effectively buffer the vulnerability of desert-margin ecosystems and thereby maintain vegetation carbon sink stability. Such findings are consistent with previous studies demonstrating that low-emission trajectories favor more stable ecosystem functioning and enhanced carbon sink capacity (White et al., 2021; Zhang et al., 2024). By contrast, SSP245 showed the most pronounced bidirectional transitions within medium-risk zones, indicating that transitional ecosystems are highly sensitive under intermediate climate forcing and may oscillate between stability and instability once climatic thresholds are approached. Under SSP370, instability became more pronounced, with widespread risk-level conversions, reflecting the systemic vulnerability of ecosystems under strong climate forcing.

This spatially explicit risk assessment provides a practical basis for targeted carbon sink management in arid and semi-arid regions. For high-risk and extreme-risk zones, which were mainly distributed along desert margins and fragile ecotones, the primary ecological feature is weak vegetation cover, high climatic sensitivity, and a high likelihood of sink-to-source transition. These areas usually have high restoration difficulty because vegetation recovery is constrained by water limitation, wind erosion, soil degradation, and frequent climatic disturbances. Therefore, the management objective should be to prevent persistent carbon source formation and restore basic ecosystem stability. Priority measures include vegetation restoration with drought-tolerant native species, sand stabilization, soil moisture conservation, enclosure or grazing exclusion in severely degraded patches, and strict control of land reclamation, overgrazing, mining, and disorderly infrastructure expansion. Ecological compensation and restoration investment should be preferentially allocated to these zones, especially where high-risk patches overlap with desertification-prone belts.

Moderate-risk zones should be regarded as carbon sink risk transition zones. Ecologically, they represent a buffer and early-warning belt between relatively stable carbon sink areas and high-risk areas prone to carbon source conversion. Compared with high-risk zones, these areas may still retain relatively better vegetation conditions and greater management flexibility, but their carbon sink function is more sensitive to climatic perturbations. The main management objective should therefore be to prevent risk escalation and maintain their carbon sink function before irreversible degradation occurs. Continuous monitoring of NEP anomalies, vegetation cover, drought intensity, and soil moisture conditions should be strengthened. Early-warning thresholds can be established based on climatic and vegetation indicators, so that management actions can be triggered when drought, heat stress, or declining vegetation productivity persists. Adaptive measures such as rotational grazing, seasonal grazing prohibition, restoration of degraded patches, improvement of landscape connectivity, and protection of ecological corridors can help buffer climate perturbations and prevent these zones from shifting into high-risk categories.

Low-risk zones function as relatively stable carbon sink areas and ecological reference zones. These areas generally have better vegetation growth conditions, lower modeled likelihood of sink-to-source transition, and stronger carbon sink persistence. Their management difficulty is relatively low, but they may still be threatened by future land-use intensification or extreme climate events. The management objective should be to maintain existing carbon sink capacity and prevent new degradation. These zones should be conserved as baseline areas for evaluating restoration effectiveness in surrounding risk zones. Management should focus on maintaining current vegetation cover, preventing unnecessary land conversion, avoiding excessive grazing pressure, and preserving landscape connectivity. In areas with stable carbon sink performance, sustainable grazing and ecological use can be allowed under carrying-capacity constraints.

At the regional governance level, carbon sink risk zoning can support differentiated land management and ecological restoration planning. High-risk and extreme-risk zones should be incorporated into priority restoration and ecological protection programs; moderate-risk zones should be included in dynamic monitoring and early-warning systems; and low-risk zones should be maintained as stable carbon sink reserves and ecological benchmarks. This risk-based zoning strategy can help shift carbon sink management from uniform restoration toward spatially targeted intervention, improving the efficiency of ecological investment and enhancing the long-term stability of vegetation carbon sinks in vulnerable dryland ecosystems.

5 Conclusion

References

• 1

  Ariza-GarzónM. J.ArroyoJ.CaparriniA.Segovia-VargasM. J. (2020). Explainability of a machine learning granting scoring model in peer-to-peer lending. Ieee Access8, 64873–64890. 10.1109/access.2020.2984412

  • CrossRef
  • Google Scholar

• 2

  ChaiH.MaJ.ZhangJ.LiJ.MengB.WangC.et al (2025). Nonlinear responses of ecosystem carbon fluxes to precipitation change in a semiarid grassland. Front. Plant Sci.16, 1519879. 10.3389/fpls.2025.1519879

  • CrossRef
  • Google Scholar

• 3

  ChangJ.CiaisP.ViovyN.VuichardN.HerreroM.HavlíkP.et al (2016). Effect of climate change, CO₂ trends, nitrogen addition, and land cover and management intensity changes on the carbon balance of European grasslands. Glob. Change Biol.22 (1), 338–350. 10.1111/gcb.13050

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 4

  ChenS.HuangY.ZouJ.ShiY.LuY.ZhangW.et al (2012). Interannual variability in soil respiration from terrestrial ecosystems in China and its response to climate change. Sci. China Earth Sci.55 (12), 2091–2098. 10.1007/s11430-012-4464-6

  • CrossRef
  • Google Scholar

• 5

  CuiS.ZhaoS.LiC.WuY.KolerskiT.ZhangM. (2025). Modeling and evaluation of net ecosystem productivity and its driving factors in Inner Mongolia. Front. Environ. Sci.13, 1581983. 10.3389/fenvs.2025.1581983

  • CrossRef
  • Google Scholar

• 6

  DaiE.HuangY.WuZ.ZhaoD. (2016). Analysis of spatio-temporal features of a carbon source/sink and its relationship to climatic factors in the Inner Mongolia grassland ecosystem. J. Geogr. Sci.26 (3), 297–312. 10.1007/s11442-016-1269-0

  • CrossRef
  • Google Scholar

• 7

  DongS.DuS.WangX. C.DongX. (2024). Terrestrial vegetation carbon sink analysis and driving mechanism identification in the Qinghai-Tibet Plateau. J. Environ. Manag.360, 121158. 10.1016/j.jenvman.2024.121158

  • CrossRef
  • Google Scholar

• 8

  GöltasM.AyberkH.KücükO. (2024). Forest fire occurrence modeling in southwest Turkey using MaxEnt machine learning technique. iForest-Biogeosciences For.17 (1), 10–18. 10.3832/ifor4321-016

  • CrossRef
  • Google Scholar

• 9

  GuoW.HuangS.LiuL.LengG.HuangQ.ChenD.et al (2025). Global critical drought thresholds of terrestrial carbon sink–source transition. Glob. Change Biol.31 (3), e70129. 10.1111/gcb.70129

  • CrossRef
  • Google Scholar

• 10

  HaoL.WangS.CuiX.ZhaiY. (2021). Spatiotemporal dynamics of vegetation net primary productivity and its response to climate change in Inner Mongolia from 2002 to 2019. Sustainability13 (23), 13310. 10.3390/su132313310

  • CrossRef
  • Google Scholar

• 11

  HouQ.YuX. (2025). Seasonal variation in carbon flux and the driving mechanisms in the grassland ecosystem in a mountain region of Northwest China. Ecol. Indic.179, 114168. 10.1016/j.ecolind.2025.114168

  • CrossRef
  • Google Scholar

• 12

  HuZ.ChenS.HaoY. (2022). Carbon and water fluxes in ecologically vulnerable areas in China. J. Plant Ecol.15 (5), 879–881. 10.1093/jpe/rtac082

  • CrossRef
  • Google Scholar

• 13

  HuangS.TianY.ZhangQ.TaoJ.ZhangY.LinJ. (2024). Spatiotemporal changes and driving mechanism of ecosystem carbon sink in karst peak cluster depression basin in Southwest Guangxi based on the interaction of “water-rock-soil-air-biology”. Ecol. Inf.83, 102800. 10.1016/j.ecoinf.2024.102800

  • CrossRef
  • Google Scholar

• 14

  KannenbergS. A.AndereggW. R. L.BarnesM. L.DannenbergM. P.KnappA. K. (2024). Dominant role of soil moisture in mediating carbon and water fluxes in dryland ecosystems. Nat. Geosci.17, 38–43. 10.1038/s41561-023-01351-8

  • CrossRef
  • Google Scholar

• 15

  KassJ. M.MuscarellaR.GalanteP. J.BohlC. L.Pinilla BuitragoG. E.BoriaR. A.et al (2021). ENMeval 2.0: redesigned for customizable and reproducible modeling of species’ niches and distributions. Methods Ecol. Evol.12 (9), 1602–1608. 10.1111/2041-210x.13628

  • CrossRef
  • Google Scholar

• 16

  KumarM.DenisD. M.SuryavanshiS. (2016). Long-term climatic trend analysis of Giridih district, Jharkhand (India) using statistical approach. Model. Earth Syst. Environ.2 (3), 116. 10.1007/s40808-016-0162-2

  • CrossRef
  • Google Scholar

• 17

  LiY.LiM.LiC.LiuZ. (2020a). Optimized maxent model predictions of climate change impacts on the suitable distribution of Cunninghamia lanceolata in China. Forests11 (3), 302. 10.3390/f11030302

  • CrossRef
  • Google Scholar

• 18

  LiZ.ChenY.ZhangQ.LiY. (2020b). Spatial patterns of vegetation carbon sinks and sources under water constraint in Central Asia. J. Hydrology590, 125355. 10.1016/j.jhydrol.2020.125355

  • CrossRef
  • Google Scholar

• 19

  LiC.FuB.WangS.StringerL. C.WangY.LiZ.et al (2021). Drivers and impacts of changes in China's drylands. Nat. Rev. Earth Environ.2 (12), 858–873. 10.1038/s43017-021-00226-z

  • CrossRef
  • Google Scholar

• 20

  LiX.LinG.JiangD.FuJ.WangY. (2022). Spatiotemporal evolution characteristics and the climatic response of carbon sources and sinks in the Chinese grassland ecosystem from 2010 to 2020. Sustainability14 (14), 8461. 10.3390/su14148461

  • CrossRef
  • Google Scholar

• 21

  LiH.WangQ.LiM.ZangX.WangY. (2024a). Identification of urban waterlogging indicators and risk assessment based on MaxEnt Model: a case study of Tianjin Downtown. Ecol. Indic.158, 111354. 10.1016/j.ecolind.2023.111354

  • CrossRef
  • Google Scholar

• 22

  LiW.JiangR.WuH.XieJ.ZhaoY.LiF.et al (2024b). An integrated urban flooding risk analysis framework leveraging machine learning models: a case study of Xi'an, China. Int. J. Disaster Risk Reduct.112, 104770. 10.1016/j.ijdrr.2024.104770

  • CrossRef
  • Google Scholar

• 23

  LiG.ZouY.YuL.ZhuB.LiX.TaoH.et al (2025a). Dynamic responses of grassland vegetation productivity and carbon use efficiency to climate change on the Mongolian Plateau. Ecol. Front.45 (3), 515–524. 10.1016/j.ecofro.2024.12.004

  • CrossRef
  • Google Scholar

• 24

  LiK.ZhaoJ.LiY.LinY. (2025b). Identifying trade-offs and synergies among land use functions using an XGBoost-SHAP model: a case study of Kunming, China. Ecol. Indic.172, 113330. 10.1016/j.ecolind.2025.113330

  • CrossRef
  • Google Scholar

• 25

  LiuY.ZhangZ.TongL.KhalifaM.WangQ.GangC.et al (2019). Assessing the effects of climate variation and human activities on grassland degradation and restoration across the globe. Ecol. Indic.106, 105504. 10.1016/j.ecolind.2019.105504

  • CrossRef
  • Google Scholar

• 26

  LiuZ.ChenZ.YuG.YangM.ZhangW.ZhangT.et al (2022). Ecosystem carbon use efficiency in ecologically vulnerable areas in China: variation and influencing factors. Front. Plant Science13, 1062055. 10.3389/fpls.2022.1062055

  • CrossRef
  • Google Scholar

• 27

  LyuX.LiX.DouH.DangD.LiS.LiX.et al (2020). Evaluation of grassland carbon pool based on TECO-R model and climate-driving function: a case study in the Xilingol typical steppe region of Inner Mongolia, China. Ecol. Indic.117, 106508. 10.5194/egusphere-egu2020-1574

  • CrossRef
  • Google Scholar

• 28

  MengF.LuoM.WangY.SaC.BaoS.BaoY.et al (2023). Revisiting the main driving factors influencing the dynamics of gross primary productivity in the Mongolian Plateau. Agric. For. Meteorology341, 109689. 10.1016/j.agrformet.2023.109689

  • CrossRef
  • Google Scholar

• 29

  NieZ.WangC.ZhangN. (2025). Faster growth rate induces higher ecosystem productivity in Inner Mongolian grasslands during 2000–2018 years. J. Arid Environ.228, 105347. 10.1016/j.jaridenv.2025.105347

  • CrossRef
  • Google Scholar

• 30

  PandeC. B.EgbueriJ. C.CostacheR.SidekL. M.WangQ.AlshehriF.et al (2024). Predictive modeling of land surface temperature (LST) based on Landsat-8 satellite data and machine learning models for sustainable development. J. Clean. Prod.444, 141035. 10.1016/j.jclepro.2024.141035

  • CrossRef
  • Google Scholar

• 31

  PangB.LiuY.AnR.XieY.TongZ.LiuY. (2024). Spatial and temporal divergence and driving mechanisms of carbon sinks in terrestrial ecosystems in the middle reaches of the Yangtze River urban agglomerations during 2008–2020. Ecol. Indic.165, 112205. 10.1016/j.ecolind.2024.112205

  • CrossRef
  • Google Scholar

• 32

  PengJ.DanL.TangX.YangF. (2022). Trends in carbon sink along the Belt and Road in the future under high emission scenario. Atmos. Ocean. Sci. Lett.15 (3), 100149. 10.1016/j.aosl.2022.100149

  • CrossRef
  • Google Scholar

• 33

  QiS.ZhangH.ZhangM. (2024). Evolutionary characteristics of carbon sources/sinks in Chinese terrestrial ecosystems regarding to temporal effects and geographical partitioning. Ecol. Indic.160, 111923. 10.1016/j.ecolind.2024.111923

  • CrossRef
  • Google Scholar

• 34

  QuJ.XuZ.DongB.WangH.HanY. (2025). Analyzing trade-offs, synergies, and driving factors of ecosystem services in Anhui Province using spatial analysis and XG-boost modeling. Ecol. Indic.171, 113098. 10.1016/j.ecolind.2025.113098

  • CrossRef
  • Google Scholar

• 35

  RaviA.PillaiD.DeshpandeM. V. (2025). Declining net carbon sequestration of west-central Indian ecosystem in response to frequently occurring drought: inference from satellite measurements and modeling. Ecol. Inf.91, 103386. 10.1016/j.ecoinf.2025.103386

  • CrossRef
  • Google Scholar

• 36

  ShiL.HeH.ZhangL.WangJ.RenX.YuG.et al (2025). Stability of China's terrestrial ecosystems carbon sink during 2000-2020. Resources. Conservation Recycl.212, 108007. 10.1016/j.resconrec.2024.108007

  • CrossRef
  • Google Scholar

• 37

  TangX.LiuY.MaM.ZhenH. (2025). Exploring the spatiotemporal variations and influential factors of net ecosystem productivity in Inner Mongolian grassland ecosystem. Remote Sens. Nat. Resour.37, 212–220. 10.6046/zrzyyg.2024019

  • CrossRef
  • Google Scholar

• 38

  WangD.AlimohammadiN. (2012). Responses of annual runoff, evaporation, and storage change to climate variability at the watershed scale. Water Resour. Res.48 (5). W05546. 10.1029/2011wr011444

  • CrossRef
  • Google Scholar

• 39

  WangS.WangW.WuY.LiW.ZhaoS.ChenZ. (2023). Effects of a changing climate and anthropogenic impacts on net primary production in Yinshanbeilu, Inner Mongolia, China. Environ. Impact Assess. Rev.102, 107185. 10.1016/j.eiar.2023.107185

  • CrossRef
  • Google Scholar

• 40

  WangL.ZhaoJ.AiD.ChenG.LinY. (2024). Integrating risk zoning and multifactor analysis: a strategic approach to ecological carbon sink management. Ecol. Inf.82, 102671. 10.1016/j.ecoinf.2024.102671

  • CrossRef
  • Google Scholar

• 41

  WhiteH. J.CaplatP.EmmersonM. C.YearsleyJ. M. (2021). Predicting future stability of ecosystem functioning under climate change. Agriculture. Ecosyst. and Environ.320, 107600. 10.1016/j.agee.2021.107600

  • CrossRef
  • Google Scholar

• 42

  WuY.AlexandrovG.EliseevA. V.ChernokulskyA. (2024). Post-agricultural ecosystems of the Chinese Loess Plateau and the Central Russian Plain as carbon sinks in 2025–2065: evidence from CMIP6 model projections. Izvestiya. Atmos. Ocean. Phys.60 (6), 727–736. 10.1134/s0001433824700737

  • CrossRef
  • Google Scholar

• 43

  WuR.HongZ.DuW.YingH.WuR.ShanY.et al (2025). Analysis of carbon Source/Sink driving factors under climate change in the Inner Mongolia grassland ecosystem through MGWR. Atmosphere16 (5), 607. 10.3390/atmos16050607

  • CrossRef
  • Google Scholar

• 44

  WuC.ZhangM.QiX.XuH.ZengY.KongY.et al (2026). Spatial patterns of vegetation carbon sink stability in China's drylands and its response to the aridity. CATENA270, 110157. 10.1016/j.catena.2026.110157

  • CrossRef
  • Google Scholar

• 45

  XieQ.RenJ. (2025). Climate-Driven dynamics of landscape patterns and carbon sequestration in Inner Mongolia: a spatiotemporal analysis from 2000 to 2020. Atmosphere16 (7), 790. 10.3390/atmos16070790

  • CrossRef
  • Google Scholar

• 46

  XieW.ChenS. T.HuZ. H. (2014). Factors influencing the variability in soil heterotrophic respiration from terrestrial ecosystem in China. Environ. Sci.35 (1), 334–340. Available online at: https://europepmc.org/article/med/24720224.

  • Pubmed Abstract
  • Google Scholar

• 47

  XuL.YuG.HeN. (2019). Increased soil organic carbon storage in Chinese terrestrial ecosystems from the 1980s to the 2010s. J. Geographical Sciences29 (1), 49–66. 10.1007/s11442-019-1583-4

  • CrossRef
  • Google Scholar

• 48

  XuX.JiaoF.LinD.LiuJ.ZhangK.YangR.et al (2023). Carbon sink trends in the karst regions of Southwest China: impacts of ecological restoration and climate change. Land12 (10), 1906. 10.3390/land12101906

  • CrossRef
  • Google Scholar

• 49

  YangJ.HuangX. (2021). The 30 m annual land cover dataset and its dynamics in China from 1990 to 2019. Earth Syst. Sci. Data13 (8), 3907–3925. 10.5194/essd-13-3907-2021

  • CrossRef
  • Google Scholar

• 50

  YangH.ZhangH.WangY.JiaX.HaoL.JinK.et al (2025). Urban bird diversity conservation plan based on the MaxEnt model and InVEST model: a case study of Jinan, China. Ecol. Indic.174, 113463. 10.2139/ssrn.5020174

  • CrossRef
  • Google Scholar

• 51

  YouC.WangY.TanX.CuiE.ZhangB.BianC.et al (2023). Inner Mongolia grasslands act as a weak regional carbon sink: a new estimation based on upscaling eddy covariance observations. Agric. For. Meteorology342, 109719. 10.1016/j.agrformet.2023.109719

  • CrossRef
  • Google Scholar

• 52

  YuZ.MiaoL.LiuQ.KattelG. R.BaoG.ChuaiX.et al (2024). Uncovering the impact of multiple determinants on vegetation NPP in Inner Mongolia. Glob. Ecol. Conserv56, e03341. 10.1016/j.gecco.2024.e03341

  • CrossRef
  • Google Scholar

• 53

  YuanY.GuoW.TangS.ZhangJ. (2024). Effects of patterns of urban green-blue landscape on carbon sequestration using XGBoost-SHAP model. J. Clean. Prod.476, 143640. 10.1016/j.jclepro.2024.143640

  • CrossRef
  • Google Scholar

• 54

  YuanX.ChenX.OchegeF. U.HamdiR.TabariH.LiB.et al (2025). Weakening of global terrestrial carbon sequestration capacity under increasing intensity of warm extremes. Nat. Ecol. Evol.9 (1), 124–133. 10.1038/s41559-024-02576-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 55

  ZhangR.ZhaoX.ZuoX.DegenA. A.LiY.LiuX.et al (2020a). Drought-induced shift from a carbon sink to a carbon source in the grasslands of Inner Mongolia, China. Catena195, 104845. 10.1016/j.catena.2020.104845

  • CrossRef
  • Google Scholar

• 56

  ZhangL.XiaoJ.ZhengY.LiS.ZhouY. (2020b). Increased carbon uptake and water use efficiency in global semi-arid ecosystems. Environ. Res. Lett.15 (3), 034022. 10.1088/1748-9326/ab68ec

  • CrossRef
  • Google Scholar

• 57

  ZhangC.HuangN.WangL.SongW.ZhangY.NiuZ. (2022). Spatial and temporal pattern of net ecosystem productivity in China and its response to climate change in the past 40 years. Int. J. Environ. Res. Public Health20 (1), 92. 10.3390/ijerph20010092

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 58

  ZhangJ.MaX.ZhangJ.SunD.ZhouX.MiC.et al (2023a). Insights into geospatial heterogeneity of landslide susceptibility based on the SHAP-XGBoost model. J. Environ. Manag332, 117357. 10.1016/j.jenvman.2023.117357

  • CrossRef
  • Google Scholar

• 59

  ZhangK.ZhuC.MaX.ZhangX.YangD.ShaoY. (2023b). Spatiotemporal variation characteristics and dynamic persistence analysis of carbon sources/sinks in the Yellow River Basin. Remote Sens.15 (2), 323. 10.3390/rs15020323

  • CrossRef
  • Google Scholar

• 60

  ZhangD.ZhaoY.WuJ. (2023c). Assessment of carbon balance attribution and carbon storage potential in China's terrestrial ecosystem. Resour. Conserv. Recycl.189, 106748. 10.1016/j.resconrec.2022.106748

  • CrossRef
  • Google Scholar

• 61

  ZhangY.ZhangY.LianX.ZhengZ.ZhaoG.ZhangT.et al (2023d). Enhanced dominance of soil moisture stress on vegetation growth in Eurasian drylands. Natl. Sci. Rev.10 (8), nwad108. 10.1093/nsr/nwad108

  • CrossRef
  • Google Scholar

• 62

  ZhangY.ZhangY.YangJ.WuW.TaoR. (2024). Integrating system dynamics, land change models, and machine learning to simulate and predict ecosystem carbon sequestration under RCP-SSP scenarios: fusing land and climate changes. Land13 (11), 1967. 10.3390/land13111967

  • CrossRef
  • Google Scholar

• 63

  ZhengH.HuangY.ZhangW.SongC.ZhangQ.SunW.et al (2024). The implementation of ecological protection in Inner Mongolia has slowed down grassland degradation. Fundam. Res.5, 2719–2730. 10.1016/j.fmre.2024.10.006

  • CrossRef
  • Google Scholar

• 64

  ZhouS.LiuZ.WangM.GanW.ZhaoZ.WuZ. (2022a). Impacts of building configurations on urban stormwater management at a block scale using XGBoost. Sustain. Cities Soc.87, 104235. 10.1016/j.scs.2022.104235

  • CrossRef
  • Google Scholar

• 65

  ZhouY.LiY.LiW.LiF.XinQ. (2022b). Ecological responses to climate change and human activities in the arid and semi-arid regions of Xinjiang in China. Remote Sens.14 (16), 3911. 10.3390/rs14163911

  • CrossRef
  • Google Scholar