Carbon Storage Distribution and Influencing Factors in the Northern Agro-Pastoral Ecotone of China

Abstract

Under the global goals of carbon peaking and carbon neutrality, China’s northern agro-pastoral ecotone—an ecologically fragile transition zone with drastic land use/cover change (LUCC)—is characterized by a lack of in-depth understanding of its “land use conflict–carbon sink response” mechanism, which is essential for regional land optimization and carbon neutrality. This study quantified the spatiotemporal dynamics of carbon storage in the zone from 2000 to 2020 using the InVEST model and identified key driving factors by combining the XGBoost model (R² = 0.73–0.88) with the SHAP framework. The results showed that regional total carbon storage increased by 30.11 × 10⁶ tons (a net growth of 0.57%), mainly driven by forest carbon sinks (+65.74 × 10⁶ tons, accounting for 218.3% of the total increase), while cropland and grassland underwent continuous carbon loss (−53.87 × 10⁶ tons and −35.80 × 10⁶ tons, respectively). Spatially, this presents a pattern of “high-value agglomeration in the central–southern region and low-value fragmentation at urban–rural edges”. The Normalized Difference Vegetation Index (NDVI) was the primary driver (average SHAP value: 426.15–718.91), with its interacting temperature factor evolving from air temperature (2000) to nighttime surface temperature (2020). This study reveals the coupling mechanism of “vegetation restoration–microenvironment regulation–carbon sink gain” driven by the Grain for Green Program, providing empirical support for land use optimization and carbon neutrality in agro-pastoral areas.

1. Introduction

In recent years, global climate change has led to frequent extreme weather events, exerting profound impacts on global ecosystems and human society [1]. As a core strategy to address climate change, carbon neutrality has become a global consensus and a key action pathway [2,3]. As an important component of the global carbon cycle, terrestrial ecosystem carbon storage not only occupies a central position in climate regulation [4], but its dynamic changes also directly affect the balance of regional carbon budgets. The global terrestrial ecosystem can absorb approximately 25% of anthropogenic carbon dioxide emissions annually, forming a crucial terrestrial carbon sink [5]. Enhancing the carbon sequestration capacity of ecosystems is therefore regarded as a sustainable strategy to mitigate climate change [6]. Among the driving factors, Land Use and Cover Change (LUCC) is a core factor shaping terrestrial carbon cycle processes [7]. It refers to the long-term dynamic changes in surface land use types (such as cropland, forestland, construction land, water area and unused land) and their cover conditions driven by human activities or natural processes, and is crucial because it directly affects the direction and intensity of carbon cycling by altering vegetation and soil carbon pools. By altering surface vegetation types and soil carbon pool structures, LUCC acts as a key variable triggering changes in regional carbon storage. Activities such as urban expansion and deforestation have significantly weakened the carbon sequestration capacity of ecosystems [8,9]. With the intensification of global warming and the acceleration of urbanization, land use patterns are transforming at an unprecedented rate [10], severely disrupting the carbon cycle and increasing the risk of carbon imbalance. Therefore, in-depth exploration of the dynamic characteristics of LUCC and its mechanism of action on carbon storage is of vital significance for promoting the implementation of regional sustainable development strategies and facilitating the achievement of carbon neutrality goals.

In the LUCC implementation strategy report jointly released by the International Geosphere-Biosphere Programme (IGBP) and the International Human Dimensions Programme on Global Environmental Change (IHDP), it is emphasized that, due to significant regional differences in the impacts and responses of LUCC to global change, regional-scale LUCC research should focus on “key areas”, “hotspot areas”, and “vulnerable areas” [11]. In China, LUCC research in vulnerable ecological regions has mainly concentrated on the Loess Plateau, oases, arid regions, river deltas, and agri-forestry/agro-pastoral ecotones [12]. The northern agro-pastoral ecotone serves as China’s second ecological security barrier following the natural grasslands in pastoral areas; it also functions as an ecological barrier for central and eastern China and an important water conservation area for the Beijing–Tianjin–Hebei region [13]. Characterized by an overall fragile ecological background, the region exhibits rapid responses to external disturbances, high sensitivity, and large variability, making it a typical area where ecological vulnerability, environmental sensitivity, and high agricultural economic risks coexist. This region is undergoing drastic LUCC transitions driven by the Grain for Green Program [14,15]; coupled with semi-arid climatic fluctuations, these transitions result in highly nonlinear dynamics of carbon storage [16,17,18].

A variety of methods are available for assessing terrestrial ecosystem carbon storage, including sample surveys, remote sensing estimation, and model simulation. These traditional methods are often time-consuming and labor-intensive and are not suitable for large-scale regional studies. In recent years, carbon storage estimation methods have become more complex and diverse. The Integrated Valuation of Ecosystem Services and Trade-offs (InVEST) model has emerged as a widely used tool owing to its ease of operation, flexible parameters, and relatively accurate results [19,20]. Studies have applied the PLUS-InVEST model to estimate green and blue carbon stocks in Zhangzhou City, Fujian Province (2000–2020), and predict future trends, highlighting land use as a key driver of carbon stock changes and the study’s significance for China’s “dual-carbon” goals [21]. researchers have utilized the MCCA-InVEST model to simulate carbon storage changes in Chengdu (2025–2035) under three scenarios (natural development, ecological conservation, sustainable development), clarifying the critical role of land use patterns in balancing ecological security and economic growth [22]. Scholars have also employed the PLUS-InVEST model to predict carbon storage trends in the Dongting Lake Basin (2030–2060), with results showing that the ecological conservation scenario effectively curbs regional carbon loss [23]. Existing studies have assessed the spatiotemporal changes in regional ecosystem carbon storage at different scales, such as river basins [24,25], administrative divisions [26,27], and national parks [28,29]. Gaining a deep understanding of the driving mechanisms behind carbon storage changes and analyzing the contribution of various influencing factors are crucial for the precise regulation of ecosystem carbon sequestration functions. Geodetector is widely used in driving factor analysis because it can measure geographical differences and interactions between different driving factors [30,31]. Studies have found that forest expansion driven by ecological restoration projects can significantly increase carbon storage, while the reduction in cropland and rangeland areas caused by urbanization leads to obvious carbon loss [32]; other research has pointed out that accelerated urbanization in the Panxi region of Sichuan Province has resulted in a decrease in cropland and grassland areas, ultimately leading to a decline in regional carbon storage [33]. These studies have confirmed the universal impact of LULC on carbon storage, but in-depth and systematic research on the unique “land use conflict–carbon sink response” mechanism in agro-pastoral ecotones is still lacking. For driving factor analysis, most studies have adopted models such as Geodetector and Optimal Parameters-based Geographical Detector (OPGD) to clarify the impact of driving factors on the spatial distribution of carbon storage [33,34]. Scholars have used the OPGD model to analyze the driving factors of carbon storage in the coastal region of central Vietnam [34], while others have applied the Geodetector model to explore the driving mechanism of carbon storage in the Panxi region [33]. Geographically Weighted Regression (GWR) and Structural Equation Modeling (SEM) have been used to examine the interactions of climate, land use, and socioeconomic variables on carbon storage dynamics [35,36]. Machine learning models and the SHAP (SHapley Additive exPlanations) method have also been extensively applied in such studies. Against this backdrop, machine learning methods, with their advantage in handling complex nonlinear relationships, are increasingly used to identify factors influencing the carbon cycle [37,38]. In particular, the XGBoost (Extreme Gradient Boosting) model, renowned for its efficiency and flexibility, can uncover potential patterns that are difficult to capture using traditional methods [39]. However, due to the complexity of their internal decision-making processes, machine learning models are often regarded as “black boxes”, which makes their interpretations less intuitive. Moreover, machine learning models alone cannot clarify the spatial differentiation of carbon sinks or the specific mechanisms of their driving factors. The SHAP method, derived from game theory (Shapley value), can effectively address this limitation [40]; it provides a transparent and interpretable framework that not only improves the predictive ability of the model but also enhances the understanding of factors influencing carbon sinks.

Current studies on carbon storage mostly focus on relatively stable ecosystems, while exploration of the “land use conflict–carbon sink response” feedback mechanism in ecotonal zones remains insufficient, failing to reflect the nonlinear carbon dynamics of ecologically fragile regions; furthermore, although machine learning models have been applied to analyze the driving factors of carbon storage, few studies integrate interpretable frameworks to quantify the spatial heterogeneity of factor contributions and their interannual evolutionary mechanisms, thereby restricting the precision of carbon sink regulation strategies—this study aims to address three interconnected questions, namely: how the spatiotemporal patterns of carbon storage in China’s northern agro-pastoral ecotone have evolved from 2000 to 2020 and what differences exist in carbon dynamic characteristics across different land use types; which factors dominate the spatial heterogeneity of carbon storage in this region and how their driving mechanisms have changed; and what the coupling mechanism is between LUCC processes and carbon sink function responses in this ecologically fragile zone, as well as how to enhance regional carbon sequestration capacity through land use optimization.

Based on the above context, this study selects China’s northern agro-pastoral ecotone as the research area. As a typical ecologically vulnerable region and an area with active land use transitions, this region’s drastic LUCC processes and complex carbon sink response mechanisms urgently require in-depth analysis. This study integrates the InVEST model, XGBoost machine learning algorithm, and SHAP interpretation framework to construct a comprehensive research scheme for assessment and attribution analysis. First, the InVEST model is used to quantify the spatiotemporal dynamics of carbon storage under long-term land use changes in the region; second, the XGBoost model is employed to deeply explore and model the complex nonlinear relationships between carbon storage changes and multi-dimensional driving factors; finally, the SHAP value interpretation framework is utilized to quantify the contribution of each driving factor to the spatial heterogeneity of carbon sinks and its mechanism of action. The aim is to reveal the coupling mechanism of “land use transition–carbon sink function response” in this critical ecological ecotone, thereby providing scientific support for the optimal management of regional land use and the design of carbon neutrality pathways.

2. Materials and Methods

2.1. Overview of the Study Area

The northern agro-pastoral ecotone is a transitional and ecologically fragile area between the agricultural regions in northern China and the pastoral regions. It is located at the southeastern edge of the Inner Mongolia Plateau and the northern part of the Loess Plateau, covering 154 counties (banners or cities) across 9 provinces (Autonomous Regions) in China, including Inner Mongolia, Heilongjiang, Jilin, Liaoning, Hebei, Shanxi, Shaanxi, Gansu, and Ningxia. It spans more than 10 degrees of latitude from north to south and over 20 degrees of longitude from east to west (Figure 1).

Situated in China’s semi-arid transitional zone, the northern agro-pastoral ecotone has a typical temperate, semi-arid continental monsoon climate, with an annual average temperature of approximately 8 °C and an average annual precipitation of 300–450 mm. Agricultural and pastoral landscapes are interspersed in this area, and the vegetation transitions from forest–steppe zones in the east to typical steppe zones and desert steppe zones in the west. Geographically, the region serves as a transition zone from the Northeast China Plain, North China Plain, and Loess Plateau to the Inner Mongolia Plateau and Qinghai–Tibet Plateau, with elevation increasing from northeast to southwest.

2.2. Data Sources

This study utilizes data spanning 2000 to 2020, with detailed sources and processing procedures elaborated as follows:

Land use data: Derived from the annual 1000 m resolution land use/cover datasets (2000, 2005, 2010, 2015, 2020) provided by the Resource and Environmental Science Data Center of the Chinese Academy of Sciences (CAS) (http://www.resdc.cn, accessed on 21 July 2025). These datasets are classified into 6 primary types (cropland, forestland, grassland, water body, construction land, unused land) and 25 secondary types.

Elevation data: Sourced from the 1000 m resolution Shuttle Radar Topography Mission (SRTM) digital elevation model (DEM) (Version 4.1) via the CAS Resource and Environmental Science Data Center. This dataset covers the entire study area and has undergone preprocessing to ensure topographic continuity, providing foundational data for analyzing elevation-driven carbon spatial patterns.

Land surface temperature (LST) data: Obtained from the MOD11A2 product (8-day composite, 1 km resolution) released by NASA’s Land Processes Distributed Active Archive Center (LP DAAC), accessed through the CAS Resource and Environmental Science Data Center. The dataset includes daytime and nighttime LST in Kelvin, and annual average LST was calculated by averaging valid composites for each year (2000–2020).

Socioeconomic data:

GDP spatial grid data (1 km resolution): Generated by the CAS Resource and Environmental Science Data Center through spatial disaggregation of county-level GDP statistics (from the China Statistical Yearbook) using NPP-VIIRS nighttime light data (to reflect economic activity intensity) and land use data (to weight by built-up area proportion). The dataset was validated against municipal GDP statistics.

Population spatial grid data (1 km resolution): Developed by disaggregating county-level population census data (2000, 2010) and annual estimates (2005, 2015, 2020) from the China Population and Employment Statistical Yearbook, using residential land distribution (from land use data) as a proxy. Sourced from the same CAS database.

Meteorological data: Includes annual average air temperature, precipitation, and potential evapotranspiration (2000–2020) at 1 km resolution, downloaded from the National Tibetan Plateau Scientific Data Center (http://data.tpdc.ac.cn, accessed on 21 July 2025). These data were spatially interpolated by the Peng Shouzhang team using the thin-plate spline method, incorporating 756 national meteorological station records (1980–2020) and topographic covariates (elevation, slope, aspect).

NDVI data: From NASA’s MODIS NDVI product (MOD13A3), a 1 km resolution, 16-day composite dataset (2000–2020) downloaded via https://www.earthdata.nasa.gov, accessed on 21 July 2025. The dataset uses a maximum value composite (MVC) method to reduce atmospheric interference, with values ranging from −2000 to 10,000 (scaled by 0.0001).

Grassland grazing intensity data: Provided by the National Ecological Science Data Center (https://www.nesdc.org.cn, accessed on 21 July 2025) as 1 km resolution annual grid data (2000–2020), expressed in sheep units per hectare (SU/ha). The dataset integrates county-level livestock statistics (sheep, cattle, goats) from the China Rural Statistical Yearbook with grassland area data (from land use datasets).

Carbon density data:

For grasslands, forests, and croplands: We prioritized the 2010s carbon density dataset of China’s terrestrial ecosystems by Xu et al. [5], which includes aboveground biomass carbon (AGC), belowground biomass carbon (BGC), and soil organic carbon (SOC, 0–100 cm depth) densities. For the study area, the aboveground, belowground, and soil carbon densities of grasslands and forests, as well as the soil carbon density of croplands, were extracted from this dataset. For other land use types, relevant literature [40,41,42,43] was referenced, and adjustments were made based on the research results to obtain carbon density data suitable for the study area (

Table 1. Carbon density index of each land use type in study area (unit: t·hm²).

	Above	Below	Soil	Dead Organic Matter
Cropland	4	25.6	58.18	0.72
Forestland	33.06	7.67	83.99	0.18
Grassland	1	4.75	75.06	0.45
Water area	3	0	0	0
Construction	2.5	0.08	78	0
Unused land	1.3	8.6	31.4	0

).

To ensure spatial consistency, all datasets underwent the following processing using ArcGIS 10.8: Resampling to a 1 km resolution using nearest-neighbor interpolation. They were converted to the WGS1984 geographic coordinate system with consistent row/column counts (1981 × 1885) by cropping to the study area boundary. Temporal consistency: For multi-year datasets, they were aligned to 5-year intervals to match the research period, with missing years interpolated using linear regression.

2.3. Methods

This study employs an integrated technical approach of the InVEST model, XGBoost algorithm, and SHAP framework to investigate carbon storage dynamics and driving mechanisms in the northern agro-pastoral ecotone. To accurately characterize long-term (2000–2020) and large-scale carbon storage spatiotemporal dynamics, the InVEST model integrates four carbon pools and links land use data for quantitative simulation, providing high-precision baseline data. Given the strong nonlinearity in carbon storage dynamics (driven by overlapping LUCC transitions and semi-arid climate fluctuations), the XGBoost algorithm uncovers complex multi-factor coupling relationships inaccessible to traditional linear models. To address the XGBoost “black box” limitation, the SHAP framework quantifies driver contributions, visualizes impacts and interactions, ensuring rigorous mechanism analysis.

Together, they form a complete technical chain of “storage quantification–modeling–mechanism interpretation,” aligning with core research content and objectives.

2.3.1. InVEST Model

The Carbon Storage and Sequestration module in the InVEST model is a key tool for calculating the total carbon storage of terrestrial ecosystems. Based on the compositional characteristics of ecosystem carbon storage, this module divides the carbon storage of each land use type into four basic carbon pools: the aboveground biomass carbon pool (carbon in living plants), belowground biomass carbon pool (carbon in plant roots), soil carbon pool (organic carbon in soil), and dead organic matter carbon pool (carbon in litter and dead plants). In this study, this model was used to calculate the carbon storage of the study area, with the formula as follows:

$$C_{\mathit{total}}={\sum }_{i=1}^n{S}_i\times \left(C_{i-\mathit{above}}+C_{i-\mathit{below}}+C_{i-\mathit{soil}}+C_{i-\mathit{dead}}\right)$$

(1)

In the formula, $C_{\mathit{total}}$ represents the total carbon storage; $S_i$ represents the area of land use type $i$; $C_{i-\mathit{above}}$, $C_{i-\mathit{below}}$, $C_{i-\mathit{dead}}$, and $C_{i-\mathit{soil}}$ represent the carbon densities in aboveground biomass, belowground biomass, soil, and dead organic matter for each land use type, respectively.

2.3.2. XGBoost Model

XGBoost (Extreme Gradient Boosting) is an ensemble learning decision tree model based on the boosting framework, proposed by Chen and Guestrin in 2016 [39]. As an efficient implementation of Gradient Boosting Decision Trees (GBDTs), XGBoost exhibits significant differences from traditional GBDT approaches. Compared with conventional gradient boosting tree methods, it demonstrates a notable enhancement in both performance and accuracy. By incorporating a regularization term into the loss function, XGBoost effectively controls model complexity, while its use of the second-order Taylor expansion of the loss function for function approximation significantly improves the model’s fitting accuracy and generalization ability. The XGBoost model is expressed as

$$\widehat{y_i}=\sum _{k=1}^K{f}_k\left(x_i\right), f_k\in \mathcal{F}\mathcal{ }\left(i=\mathrm{1,2},\dots ,n\right)$$

(2)

The formula $[F=\left\{f\left(x\right)=w_{q\left(x\right)}∣q:R^m\to \{\mathrm{1,2},\dots \},\hspace{0.17em}T,\hspace{0.17em}w\in R^T\right\}]$ refers to the Classification and Regression Tree (CART), where $q\left(x\right)$ represents the mapping of samples to the leaf nodes of the tree structure, $T$ denotes the number of leaf nodes, and $w$ stands for the weight corresponding to the leaf node index.

The objective function of the XGBoost model consists of a loss function and a complexity function, and its mathematical expression is

$$Obj=L+\Omega$$

(3)

where $L$ is the loss function and $\Omega$ is the complexity function.

$$L={\sum }_{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2$$

(4)

$$\Omega =\gamma T+{\displaystyle \frac{1}{2}}\gamma {\sum }_{j=1}^T{w}_j^2$$

(5)

Here, $\gamma T$ is the L1 regularization term, which serves to control the sparsity of the model; ${\displaystyle \frac{1}{2}}\gamma \sum _{j=1}^T{w}_j^2$ is the L2 regularization term, which functions to smooth the model parameters. On this basis, the prediction function of the $t$-th tree is added iteratively, and the second-order Taylor expansion is employed to approximate the objective function, as follows:

$$Ob{j}^t\approx {\sum }_{i=1}^n\left[{\left(y_i-\widehat{y_i}\right)}^2+2\left(y_i-\widehat{y_i^{t-1}}\right)f_t\left(x_i\right)-h_i{f}_t^2\left(x_i\right)\right]+\Omega$$

(6)

$$Ob{j}^{\left(t\right)}\approx \sum _{i=1}^n\left[{\mathit{g}}_{\mathit{i}}{w}_{q\left(x_i\right)}+{\displaystyle \frac{1}{2}}{\mathit{h}}_{\mathit{i}}{w}_{q\left(x_i\right)}^2\right]+\gamma T+{\displaystyle \frac{1}{2}}\sum _{j=1}^T{w}_j^2$$

$$={\sum }_{j=1}^T\left[\left({\sum }_{i\in I_j}{\mathit{g}}_{\mathit{i}}\right)w_j+{\displaystyle \frac{1}{2}}\left({\sum }_{i\in I_j}{\mathit{h}}_{\mathit{i}}+\lambda \right)w_j^2\right]+\gamma T$$

(7)

2.3.3. SHAP

SHAP provides a technical method for explaining model outputs. SHAP values are derived from the Shapley value in game theory and are constructed as an interpretive model inspired by game theory. This model helps researchers understand the impact and directionality of each feature on specific outcomes and the model as a whole. Therefore, this study employs the TreeSHAP method to explain the predictions made by the XGBoost model on its training dataset. Its calculation formula is as follows:

$${\phi }_i={\sum }_{S\subseteq \mathcal{N}\setminus \left\{i\right\}}{\displaystyle \frac{\left|S\right|!\hspace{0.17em}\left(n-\left|S\right|-1\right)!}{n!}}\left(f_S\left(S\cup \left\{i\right\}\right)-f\left(S\right)\right)$$

(8)

where ${\phi }_i$ is the Shapley value of the $i$-th feature; $\mathcal{N}$ represents the set of all features; $S$ represents a subset that does not contain the $i$-th feature; $f_S\left(S\cup \left\{i\right\}\right)$ represents the model output when the $i$-th feature is included; $f_S$ is the model output when the $i$-th feature is not included.

2.3.4. Model Performance Evaluation Metrics

To quantitatively assess the accuracy of the model predictions, three statistical metrics were used: coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). R² reflects the goodness of fit between the predicted values and the observed values, ranging from 0 to 1. MAE represents the average of the absolute differences between predicted and observed values, reflecting the average prediction deviation. A smaller MAE indicates better model performance. RMSE measures the average magnitude of the prediction errors, with units consistent with the target variable. A smaller RMSE indicates higher prediction accuracy. Their definitions and calculation formulas are as follows:

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

(9)

$$MAE={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n\left|y_i-{\widehat{y}}_i\right|$$

(10)

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

(11)

3. Results

3.1. Characteristics of Land Use Change

3.1.1. Dynamic Evolution of the Areas of Major Land Use Types

From 2000 to 2020, the land use system in the agro-pastoral ecotone in Northern China was dominated by cultivated land and grassland. Although the combined proportion of these two land types decreased from 89.1% in 2000 to 86.0% in 2020, they remained the main components of the regional land use structure. Additionally, major land use types exhibited differentiated area evolution trends during the study period (Figure 2 and Figure 3;

Table 2. Land transfer matrix of the study area from 2000 to 2020.

		Cropland	Forestland	Grassland	Water Area	Construction	Unused Land	Total
Cropland	area (km2)	128,041	14,457	52,753	2866	9873	5254	213,244
	Proportion (%)	60.04%	6.78%	24.74%	1.34%	4.63%	2.46%	100%
Forestland	area (km2)	11,717	46,153	19,964	567	1163	1033	80,597
	Proportion (%)	14.54%	57.26%	24.77%	0.70%	1.44%	1.28%	100%
Grassland	area (km2)	51,625	22,678	158,809	2037	4972	13,896	254,017
	Proportion (%)	20.32%	8.93%	62.52%	0.80%	1.96%	5.47%	100%
Water area	area (km2)	3078	576	2017	4527	416	2113	12,727
	Proportion (%)	24.18%	4.53%	15.85%	35.57%	3.27%	16.60%	100%
Construction	area (km2)	6536	678	2574	276	2902	508	13,474
	Proportion (%)	48.51%	5.03%	19.10%	2.05%	21.54%	3.77%	100%
Unused land	area (km2)	6160	1327	13,503	1084	1089	27,086	50,249
	Proportion (%)	12.26%	2.64%	26.87%	2.16%	2.17%	53.90%	100%
Total		207,157	85,869	249,620	11,357	20,415	49,890	624,308

).

Cultivated land and grassland showed a significant slow reduction trend. Affected by the combined effects of agricultural structure adjustment and ecological policies, the area of cultivated land fluctuated slightly interannually and decreased from 213,244 km² in 2000 to 207,157 km² in 2020, with a net reduction of 6087 km² over 20 years, representing a reduction rate of 2.85%. Influenced by semi-arid climate fluctuations and grazing activities, the area of grassland showed slight interannual rebounds but an overall continuous downward trend, decreasing from 254,025 km² in 2000 to 249,620 km² in 2020, with a cumulative reduction of 4405 km² and a reduction rate of 1.73%. This reflects the certain degradation pressure faced by the regional grassland ecosystem.

Forestland and construction land exhibited typical expansion characteristics. Driven by ecological restoration projects such as the “Grain for Green Project”, the area of forestland increased steadily and continuously, rising from 80,605 km² in 2000 to 85,869 km² in 2020, with a net increase of 5264 km² and a growth rate of 6.53%. It became the fastest-growing ecological land type, indicating a positive restoration trend in the regional ecosystem. Promoted by urbanization and industrialization, construction land showed the most prominent expansion trend, with its area rapidly increasing from 13,474 km² in 2000 to 20,415 km² in 2020, a net increase of 6941 km² and a high growth rate of 51.51%. Its expansion intensity was significantly higher than that of other land types, reflecting the strong restructuring effect of human socioeconomic activities on land use space.

The areas of water bodies and unused land showed slight reduction or stabilization characteristics. Affected by uneven precipitation distribution caused by climate change and human water use activities, the area of water bodies decreased from 12,727 km² in 2000 to 11,357 km² in 2020, with a net reduction of 1370 km² and a reduction rate of 10.76%, potentially weakening the region’s hydrological regulation capacity. Under the dual effects of “development and utilization” and “ecological management”, the area of unused land tended to stabilize, slightly decreasing from 50,249 km² in 2000 to 49,890 km² in 2020, with a net reduction of 359 km² and a reduction rate of less than 0.72%. This indicates that the development intensity and ecological protection in the region are balanced.

3.1.2. Core Pathways of Land Use Conversion

Quantitative analysis of the land use transfer matrix (Table 2) for the period 2000–2020 revealed clear core pathways of land use conversion in the study area. The area and proportion characteristics of each conversion process are detailed as follows (Figure 2 and Table 2):

Two-way Conversion between Cultivated Land and Grassland: The core interactive process of land use conversion. Mutual conversion between cultivated land and grassland was the most frequent type of land use transformation during the study period. From 2000 to 2020, the area of cultivated land converted to grassland reached 52,753 km², accounting for 24.74% of the total transferred-out area of cultivated land. In turn, the area of grassland converted to cultivated land was 51,625 km², representing 20.32% of the total transferred-out area of grassland. The scales of these two conversions were comparable, forming the main interactive pattern of land use conversion in the region.

Conversion of Cultivated Land and Grassland to Forest Land: The primary source of forestland increase. Conversion of cultivated land and grassland to forestland served as the main pathway contributing to forest land expansion. Specifically, the area of grassland converted to forestland was 22,678 km², accounting for 8.93% of the total transferred-out area of grassland; the area of cultivated land converted to forestland was 14,457 km², making up 6.78% of the total transferred-out area of cultivated land. Together, these conversions accounted for 70.6% of the total increase in forestland area, acting as the primary source of forestland expansion.

Conversion of Cultivated Land and Grassland to Construction Land: The dominant spatial source of construction land expansion. The expansion of construction land primarily encroached on cultivated land and grassland. From 2000 to 2020, the area of cultivated land converted to construction land was 9873 km², accounting for 4.63% of the total transferred-out area of cultivated land; the area of grassland converted to construction land was 4972 km², representing 1.96% of the total transferred-out area of grassland. Together, these conversions accounted for 79.5% of the total increase in construction land area. Additionally, the area of forestland converted to construction land was 1163 km², making up 1.44% of the total transferred-out area of forestland.

Conversion of Unused Land to Grassland and Cultivated Land: The dominant direction of unused land conversion. Unused land was mainly converted to grassland and cultivated land. During the study period, the area of unused land converted to grassland was 13,896 km², accounting for 27.67% of the total transferred-out area of unused land; the area of unused land converted to cultivated land was 5254 km², representing 10.45% of the total transferred-out area of unused land. Together, these conversions accounted for 38.12% of the total transferred-out area of unused land, serving as the dominant direction of unused land conversion.

3.2. Carbon Storage Change

3.2.1. Spatial Distribution of Carbon Storage

The carbon storage in the study area exhibits a significant “core–periphery” spatial differentiation pattern. High-carbon-storage areas have long been stably concentrated in the central and southern regions, covering the forest–grass composite zone in the northern Loess Plateau and the high-coverage grassland zone at the southeastern edge of the Inner Mongolia Plateau. These areas have a high proportion of forestland and grassland and deep soil carbon pools, forming a contiguous core carbon sink area. Low-carbon-storage areas have a fragmented and mosaic-like distribution, mainly concentrated in the agricultural areas of northeastern river valleys (such as the upper reaches of the Liaohe River), urban–rural transition zones, and the edge zones of the northwestern desert steppe. In these areas, the proportion of construction land and the intensity of cultivated land use are high, while overall carbon density is relatively low.

From the temporal perspective, carbon storage changes from 2000 to 2020 were characterized by localized adjustments rather than abrupt global shifts in the carbon sink capacity (Figure 4). Areas with increased carbon storage were concentrated in the core zones of high-value areas, mostly due to vegetation carbon accumulation from forestland expansion. In contrast, areas with decreased carbon storage expanded along the edges of low-value areas, mainly caused by carbon loss as construction land encroached on cultivated land and grassland. Meanwhile, over 70% of the areas remained basically unchanged. Although high-carbon-storage areas showed a slow expansion trend (shifting from locally concentrated red patches in 2000 to more continuous high-value areas between 2005 and 2020, especially in the central, southern, and southeastern regions), the spatial extent of low-carbon areas in the northwest was relatively fixed. Therefore, the overall “core–periphery” differentiation pattern remained stable for a long time.

3.2.2. Moderate Growth and Offset Effect Among Land Use Types

Forestland was the major carbon sink, and construction land was a potential carbon source. From 2000 to 2020, the total carbon storage of the study area increased slowly from 5278.10 × 10⁶ t to 5308.21 × 10⁶ t, with a net increase of 30.11 × 10⁶ t over 20 years and a growth rate of only 0.57%. Overall, the carbon sink function showed the characteristic of “low-intensity growth–overall stability” (

Table 3. Changes in carbon storage of different land use types from 2000 to 2020 (unit:10⁶ t).

	2000	2005	2010	2015	2020	2000–2020
Cropland	1887.21	1877.00	1874.99	1877.61	1833.34	−53.87
Forestland	1006.76	1030.34	1034.00	1033.96	1072.50	65.74
Grassland	2064.21	2054.69	2055.33	2039.49	2028.41	−35.80
Water area	3.82	3.66	3.68	3.70	3.41	−0.41
Construction	108.57	112.82	115.00	134.13	164.50	55.93
Unused land	207.53	209.32	207.39	204.08	206.05	−1.48
total	5278.10	5287.83	5290.39	5292.97	5308.21	30.11

). Offset effects among different land use types influenced the growth of regional carbon storage. Forestland was the only carbon sink type with excess contribution, with carbon storage increasing from 1006.76 × 10⁶ t to 1072.50 × 10⁶ t, a net increase of 65.74 × 10⁶ t, accounting for 218.3% of the total increase, which was far beyond its scale. In contrast, cropland, grassland, unused land, and water bodies all showed carbon losses, with net reductions of 53.87 × 10⁶ t, 35.80 × 10⁶ t, 1.48 × 10⁶ t, and 0.41 × 10⁶ t, respectively. The total carbon loss of these four land use types reached 91.56 × 10⁶ t, offsetting 1.4 times the increase from forestland. In addition, although the carbon storage of construction land increased from 108.57 × 10⁶ t to 164.50 × 10⁶ t (with a net increase of 55.93 × 10⁶ t), considering its carbon density characteristics and ecological attributes, this increase was essentially a carbon source loss. The original carbon storage of cropland (9873 km²) and grassland (4972 km²) occupied by the expansion of construction land was calculated as (9873×(4.00 + 25.60 + 58.18 + 0.72) + 4972×(1.00 + 4.75 + 75.06 + 0.45))×10⁻⁴ = 82.3 × 10⁶ t, far exceeding its own carbon increase. This indicates that the expansion of construction land poses a potential threat to the regional carbon balance.

3.2.3. Coupling of Carbon Dynamics and Land Use Type Conversion

The coupling between carbon dynamics and land use conversion has clear path dependence (

Table 4. Core pathways of land use conversion: area, carbon density change, and carbon storage change.

Land Use Type Conversion (2000–2020)	Area Change (km2)	Carbon Density Change (t·hm2)	Carbon Storage Change (×106 t)
Cropland→Cropland	128,041	0	0
Cropland→Forestland	14,457	36.4	52.62
Cropland→Grassland	52,753	−7.24	−38.01
Cropland→Water area	2866	−85.5	−24.43
Cropland→Construction	9873	−7.92	−8.74
Cropland→Unused land	5254	−47.2	−24.98
Forestland→Cropland	11,717	−36.4	−42.27
Forestland→Forestland	46,153	0	0
Forestland→Grassland	19,964	−43.64	−91.05
Forestland→Water area	567	−121.9	−7.03
Forestland→Construction	1163	−44.32	−5.12
Forestland→Unused land	1033	−83.6	−8.67
Grassland→Cropland	51,625	7.24	37.22
Grassland→Forestland	22,678	43.64	98.97
Grassland→Grassland	158,809	0	0
Grassland→Water area	2037	−78.26	−16.02
Grassland→Construction	4972	−0.68	−0.33
Grassland→Unused land	13,896	−39.96	−55.4
Water area→Cropland	3078	85.5	26.23
Water area→Forestland	576	121.9	7.11
Water area→Grassland	2017	78.26	15.99
Water area→Water area	4527	0	0
Water area→Construction	416	77.58	3.23
Water area→Unused land	2113	38.3	8.02
Construction→Cropland	6536	7.92	5.21
Construction→Forestland	678	44.32	3.05
Construction→Grassland	2574	0.68	0.17
Construction→Water area	276	−77.58	−2.11
Construction→Construction	2902	0	0
Construction→Unused land	508	−39.28	−2.01
Unused land→Cropland	6160	47.2	28.9
Unused land→Forestland	1327	83.6	11.17
Unused land→Grassland	13,503	39.96	54.07
Unused land→Water area	1084	−38.3	−4.13
Unused land→Construction	1089	39.28	4.3
Unused land→Unused land	27,086	0	0

). The carbon sink or source effect is directly determined by conversion direction and carbon density difference. Ecological restoration conversions are core carbon sinks. For example, Grassland → Forestland brings the largest carbon gain (98.97 × 10⁶ t) with a carbon density increase of 43.64 t·hm². Cropland → Forestland adds 52.62 × 10⁶ t carbon. In contrast, degradation and urbanization conversions are major carbon sources. Forestland → Grassland causes the biggest carbon loss (91.05 × 10⁶ t) due to a 43.64 t·hm² carbon density decrease. Cropland → Construction leads to an 8.74 × 10⁶ t loss. This “sink–source offset” makes regional carbon storage show “low-intensity growth and overall stability”.

The coupling mechanism follows the rule of “land use carbon density gradient + conversion direction”. Conversions from low-carbon to high-carbon types stably increase carbon storage, such as Unused land → Grassland and Water area → Forestland. Conversions from high-carbon to low-carbon types reduce it, like Forestland → Unused land and Water area → Cropland. Notably, construction land expansion is a hidden carbon source. Though it shows a slight carbon increase in some pathways (Water area → Construction), it essentially occupies high-carbon cropland and grassland. This mechanism provides a precise basis for land use optimization. We should prioritize converting low-carbon land to forestland or grassland. We also need to restrict high-carbon land occupation by construction land to enhance regional carbon sequestration.

3.2.4. Soil Carbon Pool and Components of Carbon Pool

The aboveground carbon pool increased gradually from 390.90 × 10⁶ t in 2000 to 406.70 × 10⁶ t in 2020 (

Table 5. Carbon storage of each carbon pool. (unit: 10⁶ t).

	Above	Below	Soil	Dead Organic Matter	Total
2000	390.90	771.71	4087.25	28.24	5278.09
2005	396.59	770.03	4093.07	28.13	5287.83
2010	397.50	769.31	4095.44	28.13	5290.38
2015	397.93	768.47	4098.51	28.06	5292.97
2020	406.70	757.82	4115.99	27.69	5308.21

), a net increase of 15.80 × 10⁶ t and a growth rate of approximately 4.04%. On the whole, it exhibited “slow and small-scale growth in the early stage and accelerated growth in the later stage”; between 2000 and 2015, the annual increase was only 0.43–5.69 × 10⁶ t, while between 2015 and 2020, the increase reached 8.77 × 10⁶ t, reflecting a significant enhancement in aboveground vegetation carbon sequestration capacity (such as biomass accumulation) during the later stage.

The underground carbon pool decreased slowly from 771.71 × 10⁶ t in 2000 to 757.82 × 10⁶ t in 2020, a net decrease of 13.89 × 10⁶ t and a decline rate of approximately 1.80%. The decline continued slowly but accelerated in the later stage: between 2000 and 2015, the annual decrease was 0.72–1.68 × 10⁶ t, while between 2015 and 2020, it expanded to 10.65×10⁶ t, indicating a more significant loss trend in underground carbon storage in the later stage.

The soil carbon pool increased steadily from 4087.25 × 10⁶ t in 2000 to 4115.99 × 10⁶ t in 2020, a net increase of 28.74 × 10⁶ t and a growth rate of approximately 0.70%. The soil carbon pool was always the component with the highest proportion of carbon storage (accounting for 77.4% of total carbon in 2000 and 77.5% in 2020) and maintained sustained, stable growth. In particular, the increase reached 17.48 × 10⁶ t between 2015 and 2020, further consolidating its core position as a “stabilizer” of regional carbon storage.

The dead organic matter carbon pool decreased slightly from 28.24 × 10⁶ t in 2000 to 27.69 × 10⁶ t in 2020, a net decrease of 0.55 × 10⁶ t and a decline rate of approximately 1.95%. Overall, its change was gentle with a slight decrease: between 2000 and 2015, annual fluctuations were only 0.07–0.11 × 10⁶ t, and between 2015 and 2020 the decrease was only 0.37 × 10⁶ t, suggesting that the dynamics of this carbon pool were relatively stable.

3.3. Influencing Factors of Carbon Storage

3.3.1. Ranking of Single-Factor Importance and Independent Effects

To clarify the driving logic behind the spatial heterogeneity of carbon storage in the study area, this study selected 10 indicators as independent variables, including air temperature (X1), potential evapotranspiration (X2), population density (X3), nighttime surface temperature (X4), Normalized Difference Vegetation Index (NDVI, X5), grazing intensity (X6), GDP density (X7), elevation (X8), daytime surface temperature (X9), and precipitation (X10). The analysis was conducted by combining the XGBoost model and the SHAP interpretation framework. The coefficient of determination (R²) for the test sets of the XGBoost model from 2000 to 2020 ranged from 0.73 to 0.88, showing a good fitting effect and providing reliable support for quantifying the intensity of factor effects.

From the analysis results (Figure 5), natural factors play a dominant role in the distribution of carbon storage, with NDVI (X5) emerging as the most critical driving factor. Between 2000 and 2020, its average SHAP values were 503.54, 566.99, 426.15, 596.17, and 718.91, in sequence, which were not only much higher than those of other factors but also showed an overall increasing trend year by year. A high NDVI value corresponds to a strong positive SHAP value, which means that vegetation biomass accumulation (such as the increase in photosynthetic products from forestland expansion) exerted significant positive feedback on carbon sinks. In contrast, low-NDVI areas, such as the edge of the northwest desert steppe, generally have low carbon storage due to insufficient vegetation coverage. This characteristic is highly consistent with the dynamic law of carbon storage in different land types in the study area, i.e., “forestland increases carbon sinks while grassland reduces carbon”. As the core representative of topographic factors, elevation (X8) maintained a stable impact intensity on carbon storage for a long time, with its SHAP value consistently ranking among the top three—163.63 in 2000 and rising to 461.00 in 2020. Medium–low elevation areas, such as the forest–grass composite zone in the northern Loess Plateau, possess more suitable hydrothermal conditions for vegetation growth and soil organic carbon accumulation, thus forming core carbon sink areas. Conversely, high-elevation areas, such as the northwest desert steppe, exhibit generally low carbon density due to harsh climatic conditions. Among temperature factors, the daytime surface temperature (X9) showed fluctuating changes in importance: its SHAP value was 436.31 (ranking second) in 2000, dropped to 207.16 (ranking third) in 2010, and rose back to 743.80 (ranking second) in 2020. This fluctuation reflects the dynamic regulatory role of surface temperature on vegetation photosynthetic efficiency and soil microbial activity. Precipitation (X10) showed the most obvious decline in importance, with its SHAP value decreasing from 144.95 (ranking fourth) in 2000 to 96.47 (ranking ninth) in 2020. Its action direction shifted from a significant positive effect in the early stage (sufficient precipitation can promote vegetation growth) to being close to the zero axis in the recent stage. This change is presumably related to the enhanced vegetation drought resistance and altered soil water retention capacity under the area’s semi-arid climatic conditions. In contrast, human factors such as population density (X3) and GDP density (X7) retained a low importance, with SHAP values consistently below 85 (at 58.67 and 70.26, respectively, in 2020). This indicates carbon storage changes in the study area remain dominated by natural drivers, while anthropogenic interference has not yet become a leading factor.

3.3.2. Multi-Factor Interaction

2000: Temperature–Precipitation–NDVI Synergy-Dominated Stage (Figure 6a).

During this stage, land use in the study area was dominated by cropland (213,244 km²) and medium-to-low-coverage grassland (254,025 km²), and the core mechanism of carbon sink regulation was the basic synergy of “vegetation–climate”. A high NDVI value was optimized when accompanied by a suitable air temperature of 20–25 °C and sufficient precipitation of 350–450 mm, which maximized the accumulation of vegetation photosynthetic products. Conversely, low temperatures below 15 °C, low precipitation of less than 300 mm, or high-temperature droughts above 30 °C significantly inhibited vegetation growth and reduced the carbon sink capacity. Spatially, areas with elevations below 1000 m became the core zones of “NDVI–temperature–precipitation” synergistic carbon sink enhancement due to concentrated precipitation and suitable temperatures. Conversely, in high-elevation areas, the interaction effect of various factors was weak due to insufficient hydrothermal conditions.

2005–2015: Daytime Surface Temperature–NDVI–Elevation Synergy Deepening Stage (Figure 6b–d).

With the continuous advancement of the Grain for Green Project, forestland in the study area increased from 80,605 square kilometers to 85,869 square kilometers, accompanied by simultaneous increases in grassland coverage. The combined effects of the vegetation canopy and litter layer changed the surface energy balance, causing the daytime surface temperature (X9) to replace air temperature (X1) as the core factor interacting with NDVI. During this stage, high NDVI values showed a more stable carbon sink effect under medium–low daytime surface temperatures of 15–20 °C. The forestland canopy reduced the surface temperature by 3–5 °C, effectively mitigating the inhibitory effect of high temperatures on photosynthetic enzyme activity. At the same time, the thermal regulation effect of topography showed more obvious differentiation. Medium-elevation areas (1000–1500 m) were suitable for moderate-to-high daytime surface temperatures, forming “elevation–temperature–vegetation” adaptive carbon sink zones. Notably, potential evapotranspiration (X2) and NDVI showed a “stress-buffering” inverse relationship: in high-NDVI areas, evapotranspiration exceeding 800 mm inhibited carbon sinks due to intensified water stress, whereas low-NDVI areas showed a weak positive deviation due to reduced vegetation water consumption.

2020: Nighttime Surface Temperature–Precipitation–Daytime Surface Temperature Precise Coupling Stage (Figure 6e).

Affected by the regional climate warming trend, the increase in nighttime surface temperature in the study area (1.2 °C) exceeded that of the daytime surface temperature (0.8 °C), making the nighttime surface temperature (X4) a key regulatory factor. In high-NDVI areas, when the nighttime surface temperature exceeded 18 °C, plant respiration rates increased, leading to more carbon loss; concurrently, the enhanced activity of soil microorganisms accelerated organic carbon decomposition. For example, in the forestlands of the central and southern regions, the carbon increase in 2020 was 32% lower than that in 2010, thus reflecting this mechanism. In addition, the threshold of hydrothermal coupling became more precise: sufficient precipitation (>400 mm) needed to coincide with a suitable daytime surface temperature of 15–22 °C to achieve stable carbon sink enhancement. When the daytime surface temperature exceeded 25 °C, even with sufficient precipitation, water stress developed due to excessive transpiration. Meanwhile, the synergistic effect of potential evapotranspiration and high temperatures (daytime > 28 °C and nighttime > 18 °C) further intensified water stress, inhibiting vegetation carbon sequestration. The overall driving mechanism showed a significant characteristic of “micro-environmental precise regulation”.

4. Discussion

4.1. XGBoost Model Accuracy Evaluation

The XGBoost model exhibits an excellent fitting performance and computational efficiency in handling complex, multi-factor nonlinear relationships. However, its inherent black-box nature limits the interpretability of the internal logic of model outputs, which greatly limits the in-depth exploration of the driving mechanisms of carbon sinks. To address this issue, this study adopted the XGBoost–SHAP model to analyze the driving factors influencing carbon storage in the northern agro-pastoral ecotone.

The results show that the model performed well on both the training and test sets (the split ratio of the training set to the test set is 8:2). SHAP values effectively quantified the relative importance of various influencing factors. Results from 10-fold cross-validation confirmed the high reliability of the model (

Table 6. Model Performance Evaluation.

	Training Set			Testing Set
Year	R2	RMSE	MAE	R2	RMSE	MAE
2000	1	0.02	0.01	0.79	584.25	382.95
2005	1	96.98	72.6	0.8	746.37	457.96
2010	1	4.88	3.62	0.83	671.36	456.22
2015	1	0	0	0.73	844.88	505.32
2020	0.99	132.23	99.26	0.88	530.66	400.35

). In all experimental years (2000, 2005, 2010, 2015, and 2020), the coefficient of determination (R²) for the training set was close to 1.00, indicating excellent goodness of fit for the training data. Meanwhile, the training residuals were randomly distributed within a narrow range close to zero, with no systematic biases observed (Figure 7).

For the test set, the predicted values and the actual values showed a significant linear correlation, with R² values ranging from 0.73 to 0.88, which proves that the model has good generalization to unseen data. Although the residuals in the test set showed a wider range than those in the training set, they were still randomly distributed around zero, with no systematic bias relative to changes in predicted values being observed. This result not only aligns with a common phenomenon in machine learning, where the error of the test set is usually slightly higher than that of the training set, but also indicates that the model did not suffer from severe overfitting.

4.2. Changes in Carbon Storage

From 2000 to 2020, carbon storage in the agro-pastoral ecotones showed an overall increasing trend, consistent with the conclusions of some scholars. Tong Rongxin et al. [44] calculated the soil carbon storage of various provinces in China from 2000 to 2020 and showed that provinces with large areas in northern China exhibited a significant increasing trend in soil carbon storage. Similarly, Huang Yan et al. [45] reported that ecosystem carbon storage in North China also increased during the same period. The present study shows that the carbon storage increased from 5278.10 × 10⁶ tons in 2000 to 5308.21 × 10⁶ tons in 2020, closely aligning with the research results of Liu Mengzhu et al. [43]. However, it should be noted that this study reveals a “carbon storage structural contradiction”. Specifically, forest carbon sinks increase by 65.74 × 10⁶ tons, while croplands and grasslands suffer a total carbon loss of 89.67 × 10⁶ tons. The latter offsets 1.4 times the carbon gain from forests. The reason for this contradiction is as follows: Studies by Liu et al. [43] only focused on the positive impact of the “Grain for Green Program” on carbon storage. They failed to quantify the negative effects of “two-way conversion between croplands and grasslands” and “construction land occupying high-carbon-density land” in the agro-pastoral ecotone. The unique pattern of “low-intensity total growth combined with high-intensity structural contradiction” in this region differs from the reported conclusion that “soil carbon storage in northern Chinese provinces shows unidirectional growth” [44]. This difference highlights the unique ecological background of the agro-pastoral ecotone, namely its “intense land use conflicts”. It should be noted that, in the model configuration adopted in this study, the soil carbon density has no interannual changes; therefore, the dynamic changes in carbon storage were mainly driven by the mutual conversion of different land use types in the region. As a highly sensitive human–land transition zone, agro-pastoral ecotones are jointly affected by human activities and ecological policies. On the one hand, population growth and economic development have promoted the reclamation of some wastelands; on the other hand, the long-term implementation of ecological policies, such as the Grain for Green Program, has significantly improved regional vegetation coverage. Taken together, these have reshaped the land use structure and spatial distribution pattern of carbon storage.

From the perspective of land-type contributions, croplands, forests, and grasslands have consistently acted as the core carriers of regional carbon storage throughout the study period, accounting for over 93% of the total. Among them, the expansion of forestland and construction land is the main driver of the increase in carbon storage. In particular, the continuous advancement of the Grain for Green Program has converted croplands with a low-carbon-sequestration capacity into forests with a high-carbon-sequestration capacity, making this land use transition a key pathway for enhancing regional carbon sink function. Similar studies have confirmed that the implementation of large-scale ecological projects, such as the Three-North Shelterbelt Program and the Grain for Green Program, have exerted positive and sustained effects on the improvement of regional carbon storage by changing land cover types [46,47,48,49]. The results of this study are consistent with these conclusions, further highlighting the important regulatory role of ecological policies in the process of regional carbon sequestration enhancement.

From the perspective of carbon storage components, the soil carbon pool has always been the core carrier of carbon storage in the agro-pastoral ecotones, far exceeding aboveground biomass, belowground biomass, and dead organic matter. In 2000, soil carbon storage in the study area was 4087.25 × 10⁶ tons, accounting for 77.4% of the total carbon storage (5278.09 × 10⁶ tons) in that year; by 2020, it had increased to 4115.99 × 10⁶ tons, accounting for 77.5% of the total carbon storage (5308.21 × 10⁶ tons). This proportion remained basically stable over a 20-year period, staying above 77%, highlighting the dominant position of the soil carbon pool in regional carbon storage. This conclusion is consistent with the finding of Deng et al. [47] on the Loess Plateau, which holds that “soil carbon pools dominate carbon storage”. However, this study further reveals that the growth of soil carbon pools in the agro-pastoral ecotone (28.74 × 10⁶ tons) mainly comes from litter input driven by forest expansion. As shown in Table 4, the conversion of “grassland to forest” contributes 98.97 × 10⁶ tons to carbon storage. In contrast, the growth of soil carbon on the Loess Plateau depends more on vegetation root accumulation after croplands are returned to forests/grasslands (under the Grain for Green Program). This difference provides empirical evidence for the differentiation of soil carbon sink enhancement paths in different ecologically fragile areas. This is closely related to the ecological background and land use types of the agro-pastoral ecotones, primarily dominated by cropland, grassland, and forestland. Croplands, long affected by agricultural activities, receive annual inputs of crop residues into the soil; though tillage accelerates organic carbon decomposition, they still maintain a high soil carbon baseline. Grasslands, characterized by dense root systems and continuous litter input, continuously accumulate soil organic carbon, especially the deep roots of perennial herbs, which can stably store carbon in the deep soil and reduce the decomposition risk. Forests gradually supplement the soil carbon pool through litter layer decomposition and root exudate input, while the relatively stable microenvironment under forest canopies (e.g., lower temperature and higher humidity) slows organic matter decomposition, thus improving the soil carbon storage efficiency. In contrast, other carbon pools, such as the aboveground biomass (406.70 × 10⁶ tons in 2020), belowground biomass (757.82 × 10⁶ tons in 2020), and dead organic matter (27.69 × 10⁶ tons in 2020), showed slight growth under the influence of factors such as forest expansion, but their total amounts were always far lower than that of the soil carbon pool.

4.3. Carbon Storage Influencing Factors

In terms of natural factors, the Normalized Difference Vegetation Index (NDVI) is the most critical determinant influencing the distribution of carbon storage, which is consistent with previous research findings [50,51,52]. This is consistent with the conclusion of Wu et al. [50] that “NDVI dominates carbon storage in Jilin’s ecosystems”. But this study further reveals the “dynamic interaction mechanism” between NDVI and temperature. In 2000, it relied on air temperature; from 2005 to 2015, it shifted to daytime surface temperature; in 2020, it focused on nighttime surface temperature (Figure 6). This is significantly different from the conclusion of Lai et al. [37] that “NDVI and precipitation dominate carbon storage” in Hainan Island. The core reason is that under the semi-arid climate of the agro-pastoral ecotone, vegetation growth relies more on the “microenvironment regulated by temperature”. It does not rely on precipitation supply in humid areas. This finding improves the theory of differentiation of carbon storage driving factors in different climate zones. Through photosynthesis, plants absorb carbon dioxide from the atmosphere and convert it into organic matter. Part of this organic matter is stored in plant biomass, while another part enters the soil through root exudates. A higher NDVI usually indicates an increase in vegetation biomass, thereby enhancing the carbon sequestration capacity of ecosystems [53].

We observed that the interaction effect between NDVI and temperature factors exhibited significant interannual variability, reflecting the complexity of the carbon cycle response to hydrothermal conditions in the agro-pastoral ecotones. In 2000, the interaction between NDVI and air temperature (X1) was the strongest. Then, the land use pattern in the study area was dominated by croplands and medium-to-low-coverage grasslands, where the air temperature directly regulated the crop growth cycle and photosynthetic efficiency of forage grass. Early spring warming advanced the crop sowing period and extended the growing season, while suitable summer temperatures (20–25 °C) promoted the accumulation of photosynthates. Conversely, high temperatures and drought inhibited vegetation growth. Therefore, air temperature served as a key synergistic factor through which NDVI influenced carbon storage.

From 2005 to 2015, the core interaction shifted to daytime surface temperature (X9), a transition closely linked to the advancement of the Grain for Green Program. During this period, the forested area in the study region increased from 80,605 km² to 85,869 km², accompanied by a significant increase in grassland coverage. The shading effect of vegetation canopies altered the surface energy balance: forest canopies reduced the intensity of solar radiation directly reaching the ground, preventing sharp increases in surface temperature, while the litter layer of grasslands also buffered fluctuations in surface temperature. At this stage, daytime surface temperature better reflected the actual microenvironment for vegetation growth than air temperature. In summer, the daytime surface temperature of forested areas was 3–5 °C lower than that of bare land, which alleviated the heat-induced inhibition of photosynthetic enzyme activity. Meanwhile, a suitable daytime surface temperature in grasslands promoted the root uptake of water and nutrients, thereby enhancing the positive driving effect of NDVI on carbon storage. Thus, daytime surface temperature replaced air temperature as the dominant interacting factor.

By 2020, the core of the interaction further shifted to nighttime surface temperature (X4), consistent with the observed trend of increased nighttime surface than daytime surface temperatures against the backdrop of regional climate warming (average nighttime surface temperature increased by 1.2 °C from 2000 to 2020, compared with a 0.8 °C increase in daytime surface temperatures). Nighttime surface temperatures primarily regulate the carbon balance by influencing plant respiration and soil microbial decomposition. When the nighttime surface temperature is below 15 °C, the consumption of photosynthates by plant respiration is low, which is conducive to biomass accumulation; however, when the temperature exceeds 18 °C, enhanced respiration leads to increased carbon loss, while elevated microbial activity accelerates soil organic carbon decomposition. In 2020, nighttime surface temperatures generally rose in high-NDVI areas of the study region (e.g., forests in the central and southern parts), highlighting the interaction between NDVI and nighttime surface temperature. In areas with high NDVI but excessive nighttime surface temperatures, the growth rate of carbon storage slowed down significantly, explaining why the forest carbon increase in 2020 (18.54 × 10⁶ t) was lower than that in 2010 (27.24 × 10⁶ t). Most existing studies focus on the impact of daytime temperature on carbon storage [51]. Nighttime surface temperature > 18 °C inhibits carbon storage in high-NDVI areas with the conclusion of Ren et al. [38] that “carbon storage in Central Asia has low temperature sensitivity”. The difference stems from the fact that the vegetation in the agro-pastoral ecotone is mainly herbs and deciduous forests. High nighttime temperatures are more likely to increase respiratory consumption. While Central Asia is dominated by desert vegetation, its carbon metabolism has a weaker response to temperature. This provides new evidence for the sensitivity pattern of carbon cycle to nighttime warming in ecologically fragile areas.

4.4. Uncertainty Analysis

Firstly, there are still limitations in the regional adaptability of current carbon density parameters. Constrained by the distribution of sampling points in the existing literature, field data for the northwestern segment of the study area (the Hexi Corridor in Gansu Province and the desert steppe area in eastern Ningxia) are insufficient. This leads to potential deviations in the carbon density parameters of grasslands and unused land in this sub-region, which are higher than those in the central and southern regions. Meanwhile, the model fails to distinguish between different tillage methods for croplands and age structures for forests, and such differences in subdivided attributes may further introduce deviations into carbon storage estimation. For future research, field sampling can be conducted in typical sample plots of the northwestern segment, and high-resolution remote sensing data can be integrated to differentiate subdivided land use types. This will help further calibrate carbon density parameters and reduce uncertainty.

Secondly, in the calculation of carbon storage and its spatiotemporal variations, the model does not incorporate microparameters associated with biogeochemical processes—such as photosynthetic rate and soil microbial activity—that play a crucial role in the carbon sequestration process. As a result, it is difficult to fully reflect the intrinsic mechanism of the ecosystem carbon cycle.

Furthermore, the carbon density parameters in the model do not account for interannual dynamic changes. This simplifies the variations in carbon storage to a single outcome of land use type conversion, which may underestimate the impact of interannual fluctuations of environmental factors on carbon pools.

Land use change data in this study were mainly acquired through remote sensing technology. Although remote sensing technology has become increasingly advanced, subjective judgments made by operators during image interpretation may cause deviations in land use classification accuracy, which, in turn, affects the accuracy of carbon storage estimation.

5. Conclusions

This study integrated the InVEST model, XGBoost machine learning algorithm, and SHAP interpretation framework to quantify the spatiotemporal dynamics of carbon storage in China’s northern agro-pastoral ecotone from 2000 to 2020 and to identify its key driving mechanisms. The main conclusions are as follows:

1. From 2000 to 2020, regional carbon storage showed “low-intensity growth with overall stability,” while structural contradictions persisted:

Positive progress: Total carbon storage increased moderately, driven by ecological restoration projects (e.g., Grain for Green Program), with forestland emerging as the core carbon sink. The soil carbon pool remained dominant (accounting for >77% of total storage) and maintained steady growth, acting as a “carbon stabilizer.”

Core contradictions: Cropland and grassland (dominant land types) experienced continuous carbon loss, offsetting over 1.4 times the forestland carbon gain. Construction land expansion, though increasing its own carbon storage, encroached on high-carbon-density cropland and grassland, posing a potential threat to regional carbon balance.

2. Regional Differences: Spatial Heterogeneity and Targeted Management

Carbon storage exhibited a stable “core–periphery” spatial pattern, requiring differentiated management strategies:

Central–southern high-carbon zones: Characterized by contiguous forest-grass landscapes and deep soil carbon pools, these areas should prioritize consolidating carbon sink advantages (e.g., forest tending and grassland degradation prevention).

Northwestern low-carbon zones: With fragmented vegetation and harsh conditions, these areas need enhanced field sampling to refine carbon density parameters and promote desert steppe enclosure projects to improve vegetation coverage.

Urban–rural transitional zones: Dominated by construction land and intensive cropland, these areas require strict control of construction land expansion and promotion of green infrastructure (e.g., urban forests) to mitigate carbon loss.

3. Driving Mechanisms: Natural Dominance and Evolutionary Interactions

Natural factors dominated carbon storage heterogeneity, with driving mechanisms evolving interannually:

Primary driver: NDVI was the most critical factor—high NDVI (e.g., in expanded forestland) promoted carbon sink enhancement, while low NDVI (e.g., desert steppe edges) limited carbon storage.

Temperature–NDVI interaction evolution: The core temperature factor interacting with NDVI shifted over three stages: from air temperature (2000) to daytime surface temperature (2005–2015), and finally to nighttime surface temperature (2020, with >18 °C inhibiting carbon storage in high-NDVI areas).

Secondary factors: Elevation remained important (favorable for carbon accumulation at 1000–1500 m), while precipitation’s influence declined (due to improved vegetation drought resistance). Human factors (e.g., population/GDP density) had limited impact, indicating natural processes still dominated.

4. Practical Implications: Policy Recommendations for Carbon Neutrality

In alignment with China’s “double carbon” goals and regional ecological protection strategies, the following actionable suggestions are proposed: Strengthen ecological policy synergy. Continue advancing the Grain for Green Program and the Three-North Shelterbelt Program, prioritizing the conversion of low-carbon-density cropland (sloping or low-productivity plots) to high-carbon-density forestland/grassland. Establish a “carbon sink benefit evaluation mechanism” for ecological projects, linking policy subsidies or ecological compensation funds to verified carbon sequestration increments to enhance implementation effectiveness. Strictly restrict the conversion of high-carbon-density cropland and grassland to construction land. In croplands, promote straw returning to supplement soil organic carbon; in grasslands, implement rotational grazing to reduce degradation and maintain the stability of the underground carbon pool (the soil carbon pool accounts for over 77% of total carbon storage). Mitigate risks for high-NDVI areas (central–southern forests), increase forest canopy density via mixed afforestation of evergreen and deciduous species to reduce nighttime surface temperature, and promote heat-tolerant native vegetation varieties to alleviate respiration-induced carbon loss under high temperatures.