Integrated Carbon Stock Simulation in Jiangsu Province Using InVEST and Random Forest Under Multi-Scenario Climate and Productivity Pathways

Abstract

Carbon stock plays a crucial role in regulating atmospheric carbon dioxide concentrations and represents a vital ecological function for mitigating climate change and supporting long-term environmental sustainability. Jiangsu Province, a typical region experiencing rapid urbanization and land-use transformation in eastern China, serves as a representative case for regional-scale carbon assessment. This study employs the InVEST model, integrated with multi-source remote sensing data, a random forest algorithm, and a control variable approach, to simulate the spatiotemporal dynamics of carbon stock in Jiangsu Province under a set of climate, productivity, and population scenarios. Three scenario groups were designed to isolate the individual effects of climate change, gross primary productivity, and population density from 2020 to 2060, enabling a clearer understanding of the dominant drivers. The results indicate that the coupled model estimates Jiangsu’s 2020 carbon stock at 1.52 × 10⁹ t C, slightly below the 1.82 × 10⁹ t C estimated by the standalone InVEST model, with the coupled results closer to previous estimates. Compared with InVEST alone, the integrated model significantly improves numerical accuracy and spatial resolution, allowing for finer-scale pattern recognition. By 2060, carbon stock is projected to decline by approximately 24.4% across all scenarios. Among the features, climate change exerts the most significant influence, with an elasticity coefficient range of −37.76–1.01, followed by productivity, while population density has minimal impact. These findings underscore the dominant role of climate drivers and highlight that model integration improves both predictive accuracy and spatial detail, offering a more robust basis for scenario-based assessment. The proposed approach provides valuable insights for supporting sustainable carbon management, real-time monitoring, and provincial-scale decarbonization planning.

1. Introduction

As one of the most urgent challenges worldwide, climate change deeply influences ecosystems, economies, and human communities [1]. The primary cause of the problem is the ongoing rise in greenhouse gas emissions, which endangers ecosystem stability in addition to causing global warming. In recent years, research has increasingly focused on strategies to enhance carbon stocks (CSs) and reduce carbon emissions as means to mitigate these impacts [2]. Growing carbon stocks in ecosystems has been demonstrated to be a successful strategy for lowering the concentration of CO₂ in the atmosphere, which is crucial for reducing the greenhouse effect and adapting to climate change [3]. As a critical function of terrestrial ecosystems, carbon stocks capture and sequester atmospheric CO₂, significantly alleviating the concentration of greenhouse gases. This function is essential for regulating global climate and maintaining the carbon balance [4]. Thus, comprehensive studies of the dynamic shifts in carbon stock and the factors that influence them serve as a crucial basis for developing carbon neutrality policies and attaining sustainable development in addition to offering a scientific basis for combating climate change. Therefore, accurate simulation of carbon stock dynamics under different influencing factors has become a key issue in global climate change research.

Field surveys are generally accepted as the most trustworthy source of empirical data for carbon stock estimations [5]. However, quick regional assessments are severely hampered by the high expense and logistical limitations of long-term, extensive monitoring [6,7]. Remote sensing data can serve as a valuable complement to field surveys; however, their accuracy is limited by regional and vegetation-type constraints [8]. To address these limitations, the Integrated Valuation of Ecosystem Services and Tradeoffs (InVEST) model provides a practical framework that is often integrated with land-use simulation models to evaluate the spatial and temporal impacts of land-use changes on carbon stocks. For instance, the PLUS model has been coupled with InVEST to assess land-use change and carbon stock dynamics under three future scenarios, effectively linking spatially explicit land-use data with carbon stock predictions [2]. Similarly, the FLUS model and InVEST have been used together to simulate carbon stock changes under multiple land-use scenarios, offering insights into how different development pathways influence carbon storage [9]. However, the predictive capability of the InVEST model primarily relies on a simple linear connection between land-use types and carbon density and often overlooks the critical influence of dynamic land-use and land-cover changes (LUCCs), despite increasing evidence that urbanization and land transformation significantly impact carbon sequestration capacity [9,10]. Furthermore, it finds it difficult to fully represent the impact of intricate nonlinear driving forces, such as socioeconomic issues and climate change. The lack of flexibility in handling multi-dimensional nonlinear relationships limits its ability to capture complex ecological mechanisms under future scenario uncertainty.

With the advancement of artificial intelligence and computer technology in recent years, machine learning approaches have been increasingly popular in carbon stock research due to their strong data processing and pattern recognition skills [11]. For instance, a study used the random forest model (RF) to examine the spatiotemporal pattern of carbon stock [12]. Machine learning algorithms can handle multi-dimensional data, identify intricate nonlinear relationships, and predict carbon stocks with greater flexibility and accuracy than classic ecological process models. Coupling machine learning models with ecological process models such as InVEST enables a more robust and comprehensive simulation framework. While the InVEST model provides interpretable biophysical mechanisms and spatial simulation capacity, the random forest algorithm complements it by learning nonlinear patterns from high-dimensional drivers such as AGPP, climate, and socioeconomic data [8,13]. This synergy not only compensates for InVEST’s linear limitations but also enhances model adaptability to complex and uncertain future scenarios. Therefore, the coupled I-RF model integrates mechanism-driven and data-driven strengths, enabling improved accuracy, spatial detail, and interpretability in carbon stock simulation.

Climate change and alterations in land-use structure are two critical factors affecting the variations in carbon stores within China’s terrestrial ecosystems [14,15]. Traditional ecological process models, including the InVEST model, effectively represent the spatial and temporal dynamics of historical carbon stocks; however, they inadequately address the multifaceted influencing factors. They generally rely solely on land-use type and carbon density to predict carbon stocks, overlooking the significant roles that climate change, biophysical attributes, and socioeconomic dynamics may play in shaping carbon storage capacities [16]. In carbon cycle research, annual gross primary productivity (AGPP) is a crucial ecological indicator that characterizes the photosynthetic capacity of vegetation. As a key flux driving the terrestrial carbon cycle [17,18], it captures both vegetation growth and the carbon exchange processes between terrestrial ecosystems and the atmosphere. However, the quantitative relationship between AGPP and carbon stock and its underlying mechanism are still unclear. The majority of previous research has concentrated on qualitatively examining the beneficial and detrimental effects of different factors on the dynamics of carbon stocks. For instance, logistic regression and LMDI decomposition have been used to examine the effects of precipitation, elevation, and GDP on carbon stock in Qingcheng County [16]. However, within a unified research framework, studies that systematically quantify the independent contributions of multiple driving factors remain limited. Although some studies have attempted to simultaneously consider climate, biological productivity, and socioeconomic factors, most have conducted simulations under a single, unified development scenario. This approach fails to adequately capture the complex interactive effects arising from multiple future scenario combinations, including varying climate pathways, population growth trends, and ecosystem productivity changes. Such limitations hinder an accurate representation of the diversity and uncertainty in future carbon stock dynamics, thereby constraining the scientific prediction of carbon cycle processes and the formulation of informed management decisions.

A crucial precondition for accomplishing the regional carbon neutrality target is the scientific quantitative assessment of carbon dynamics in model provinces, given the notable variations in the resource characteristics, economic development level, and environmental circumstances among Chinese provinces [19]. Jiangsu Province, located in eastern China, represents an ideal case for provincial-scale carbon modeling due to its dual identity as both a major economic hub and a significant carbon emitter. Its high reliance on fossil fuels, intensive land use, and rapid urbanization have resulted in growing ecological pressures and heightened carbon management challenges [20,21]. In response, the province has prioritized low-carbon city development in recent years. However, rapid urban expansion is exceeding the spatial carrying capacity, intensifying ecological and environmental pressures [22].

To assess the impact of urbanization on regional carbon dynamics, it is essential to simulate the spatial–temporal evolution of carbon stock and identify its key drivers. Yet, the current research remains limited in capturing the nonlinear effects of climate variability, ecosystem productivity, and anthropogenic activities. To address this gap, this study integrates the process-based InVEST model with the random forest (RF) algorithm, forming a hybrid I-RF model. This fusion improves spatial accuracy while preserving ecological interpretability, enabling more robust assessments under varying urbanization scenarios [19]. Although Jiangsu Province is selected as the research area due to its representativeness in urbanization intensity and carbon emission levels, the I-RF modeling framework established in this study is designed to be transferable to other provincial-level regions with appropriate data support. This study focuses on Jiangsu Province as the research area with the following objectives: to compare the carbon stock simulation results from the InVEST model with those from a coupled machine learning-enhanced InVEST model (random forest, RF), termed the I-RF model, in order to establish a high-precision provincial-scale carbon stock simulation method, as well as to identify the key drivers of carbon stock changes, for systematically quantifying their independent effects through a factor substitution design provides a scientific basis for targeted carbon neutrality strategies and evidence-based policy formulation.

2. Materials and Methods

2.1. Study Area

Jiangsu Province lies in the heart of China’s eastern coastal region, encompassing the lower reaches of the Yangtze and Huai Rivers. It is a vital part of the Yangtze River Delta, located between 30°45′–35°08′ N latitude and 116°21′–121°56′ E longitude, covering an area of approximately 107,200 km². The province’s terrain is mainly flat plains with some hills and gentle slopes, supporting extensive river networks and lakes.

Jiangsu Province, which lies at the meeting point of the subtropical and temperate monsoon climates, has a warm and humid climate. The province receives substantial yearly precipitation, with concurrent rainfall and heat, fostering optimal conditions for plant development and carbon sequestration. Jiangsu’s humid climate and advanced ecosystems confer significant carbon stock capacity. Figure 1 depicts the precise geographic location and prevailing land-use trends of Jiangsu Province.

2.2. Data Collection and Processing

This study used six variables, classified into four primary categories, climate, geography, vegetation, and societal factors, to model carbon stock in Jiangsu Province. Climate variables include annual cumulative precipitation and annual mean temperature; the geographical variable is indicated by the Digital Elevation Model (DEM); the vegetation variable is defined by the Normalized Difference Vegetation Index (NDVI); and the social variable is represented by population density (PD). This study employed future AGPP, climate, and land-use data under three scenarios to forecast future carbon stock. Among these datasets, NDVI, land-use cover, and DEM were derived from remote sensing products. Comprehensive information regarding the data and their origins is provided in

Table 1. Brief information on the data used in this study.

Dataset	Data Source	Spatial Resolution	Time Coverage
DEM	Geospatial Data Cloud (https://www.gscloud.cn/)	30 M	-
PRE	National Tibetan Plateau Data Center (https://data.tpdc.ac.cn/)	1 KM	2020
TMP	National Tibetan Plateau Data Center (https://data.tpdc.ac.cn/)	1 KM	2020
NDVI	National Ecosystem Science Data Center (https://www.nesdc.org.cn/)	30 M	2020
PD	LandScan Platform (https://landscan.ornl.gov)	1 KM	2020
Land-Use Cover	National Glacier, Frozen Soil, and Desert Science Data Center (ncdc.ac.cn)	30 M	2020
Future Climate	https://ds.nccs.nasa.gov/	0.25°	2020–2060 (at 5-year intervals)
Future Population	https://figshare.com/	1 KM	2020–2060 (at 5-year intervals)
Future GPP	National Tibetan Plateau Data Center (https://data.tpdc.ac.cn/)	0.25°	2020–2060 (at 5-year intervals)
Future Land Use	http://www.geosimulation.cn/	1 KM	2020–2060 (at 5-year intervals)

. To ensure input data reliability, we performed pixel-based distribution checks in ArcGIS Pro 3.1 for all variables, and outliers were treated using conditional thresholding where necessary. The overall workflow is shown in Figure 2.

The future climate variables PRE (annual cumulative precipitation) and TEM (mean temperature) and annual gross primary productivity (AGPP) used in this study are sourced from the CMIP6 climate model BCC-CSM2, covering the period 2020 to 2060 with an original spatial resolution of approximately 0.25°. To maintain spatial consistency among datasets, all data with a resolution exceeding 1 km were resampled to a standardized 1 km resolution. Due to the initial resolution of future climate data and AGPP data being below 1 km, Inverse Distance Weighting (IDW) interpolation was utilized to downscale these datasets to a 1 km resolution, hence improving their spatial continuity.

To ensure data quality before introducing them to the RF model, we conducted data distribution analysis, removed outliers using statistical methods (e.g., IQR), and normalized all continuous variables to ensure consistency across different units and scales. These steps minimized biases in data distribution and improved model performance. In addition, a Yeo–Johnson transformation was applied to the target variable (carbon stock) to reduce skewness and enhance prediction accuracy.

2.3. Modeling Methods

2.3.1. InVEST Model

As a prominent tool in carbon stock studies, the InVEST model stands out for its distinct advantages over other models like CASA, FORCCHN, and LPJ-GUESS, namely its lower data requirements, faster processing speed, and higher efficiency in practical applications [23,24]. Its accuracy has also been validated through practical applications [8]. The InVEST model [3,25] was employed to quantitatively evaluate the carbon sequestration capacity of ecosystems in Jiangsu Province.

The formula for calculating carbon density across different land-use types is outlined below:

$$C_m=C_{m-above}+C_{m-below}+C_{m-soil}+C_{m-dead}$$

(1)

$C_m$ denotes the total carbon density of the land-use type (t/hm²). $C_{m-above}, C_{m-below}, C_{m-soil},$ and $C_{m-dead}$ indicate aboveground biomass carbon density (t/hm²), belowground biomass carbon density (t/hm²), soil organic carbon density (t/hm²), and dead organic carbon density (t/hm²), respectively.

The equation for determining the overall carbon stock across various land-use categories is as follows:

$$C_m=\left(C_{m-above}+C_{m-below}+C_{m-soil}+C_{m-dead}\right)\times S_m$$

(2)

Subsequently, carbon density is consolidated according to land-use area data, with $C_m$ denoting the total carbon stock for a specific land-use type (t) and $S_m$ indicating the area of that land-use type (hm²). The total carbon stock of the study area is calculated through the aggregation of carbon stocks from different land-use types.

$$C_{total}={\sum }_{m=1}^n{C}_m\times S_m$$

(3)

$C_{total}$ denotes the total carbon stock within the study area (t C), $C_m$ indicates the total carbon density for land-use type m (t/hm²), and $S_m$ signifies the area corresponding to that land-use type (hm²).

Table 2. Carbon density of different land-use types in Jiangsu Province (t/hm²) [26].

Land-Use Type	Aboveground Carbon Density	Belowground Carbon Density	Soil Carbon Density	Dead Organic Matter Carbon Density
Cropland	46.50	80.70	108.40	1.00
Forest	20.36	67.50	22.53	18.39
Grassland	4.30	86.50	170.00	7.80
Water Area Body	0.00	0.00	0.00	0.00
Built-Up Land	0.00	0.00	71.00	0.00
Unused Land	0.74	0.13	69.92	0.00

presents the carbon densities associated with various land-use types utilized for calculating carbon stock in Jiangsu Province, as applied in this study using the InVEST model. These values are sourced from [26], a study also based in Jiangsu Province, and are derived from regional land-use-based carbon assessments. Further explanation is provided below the table.

2.3.2. RF Model

The RF model serves as an efficient machine learning approach for analyzing large-scale datasets, grounded in Breiman’s classification tree algorithm [27]. This approach effectively mitigates the overfitting issues prevalent in conventional classification methods, remains resilient to absent variables, and enhances the predictive accuracy of influential factors through the lens of variable contribution [28].

This research utilized six feature variables (NDVI, PRE, TMP, AGPP, PD, DEM) selected based on carbon cycle mechanisms and pre-screening from Jiangsu Province in 2020, alongside carbon stock data as the target variable, to initialize and train a random forest classifier comprising 80 decision trees. Ten-fold cross-validation was conducted entirely within the 80% training set to select the optimal parameters and assess performance during training. To ensure the model’s generalization ability, the sample points were randomly divided into training and testing sets, utilizing 80% of the samples for model construction, and 20% held out as an independent test set, which remained untouched throughout training and parameter tuning and was only used for final model evaluation.

The coefficient of determination $R^2$ was utilized to assess the accuracy of the prediction model [13]. The $R^2$ value varies between 0 and 1, where values approaching 1 signify enhanced model performance and an ideal fit. In the formula, $y_i$ denotes the observed value, $\widehat{y_i}$ indicates the model predicted value, $\overline{y}$ signifies the average of the observed values, and $n$ represents the sample size. The numerator indicates the sum of squared prediction errors (residual sum of squares, RSS), while the denominator reflects the variance of the observed values (total sum of squares, TSS). When $R^2$ approaches 1, it signifies strong explanatory power of the model for the data; conversely, a low $R^2$ indicates poor model fit, potentially due to significant errors or underfitting. The calculation of $R^2$ is represented by the following formula:

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

(4)

2.4. Scenario Settings

This research utilized climate variables from CMIP6 across three Shared Socioeconomic Pathway (SSP) scenarios [29]. The scenarios are grounded in socioeconomic development narratives concerning a sustainable world, a middle path, and fossil fuel development [30].

Table 3. Scenario Settings.

Scenario Type	Specific Scenario	Description
Population Scenario; AGPP Scenario; Climate Scenario	SSP1-2.6	Low-emission pathway, an optimistic scenario of global mitigation efforts
	SSP2-4.5	“Middle road” development pathway, sustainable but not extreme
	SSP5-8.5	High-emission pathway, fossil fuel-driven economy-prioritized scenario

presents the detailed descriptions of the scenarios.

This study developed a systematic scenario simulation framework to examine the varying effects of climate change, population growth, and AGPP on carbon stock in terrestrial ecosystems (Figure 3). This framework utilizes the control variable method to integrate climate scenarios, population changes, and AGPP from three socioeconomic pathways, resulting in three complementary experimental designs: the Population Pressure Scenario (PPS) group, Productivity Variation Scenario (PVS) group, and Climate Driving Scenario (CDS) group.

The PPS investigates the distinct impact of population change scenarios on ecosystem carbon stock. The specific setting employs three population change scenarios: the sustainable pathway, middle pathway, and fossil fuel pathway. Climate variables and AGPP variables exhibit synchronous and consistent changes in accordance with the scenarios of each individual population. The integration of the aforementioned population changes (three types) with the climate and AGPP scenarios (three types) results in the formation of nine specific scenarios, enabling a systematic analysis of the marginal effects of changes in population factors on the carbon stock.

The PVS measures the distinct impact of AGPP variations on carbon stock while controlling for climate and population variables. This scenario group establishes various gradients of annual gross primary productivity change scenarios, wherein climate and population variables are altered synchronously and consistently with each specific AGPP scenario. The analysis focuses on the impact of AGPP changes on ecosystem carbon stock accumulation or loss. The integration of the aforementioned AGPP changes (three types) with the climate and population scenarios (three types) results in nine distinct scenarios, enabling a precise quantification of the independent effect of productivity change on carbon stock variations.

This study maintains constant population density and annual AGPP in the CDS, concentrating on assessing the effects of varying climate change scenarios on the ecosystem carbon stock. The specific setting employs three climate variable scenarios, SSP1-2.6, SSP2-4.5, and SSP5-8.5, wherein population density and AGPP exhibit synchronous and consistent changes corresponding to each individual climate scenario. The aforementioned climate changes, categorized into three types, in conjunction with the population and AGPP scenarios, also classified into three types, yield a total of nine distinct scenarios. These scenarios are designed to differentiate the effects of varying intensities of climate changes on alterations in the carbon stock.

(a)

PD: All three scenarios show a pattern of rising initially, then declining afterward. Under SSP1-2.6, the population in the study area is projected to decline from 2035 to 51.76 million by 2060, indicative of decreasing fertility rates and an aging demographic. The SSP2-4.5 population increases until 2035, after which it gradually decreases, reaching 54.43 million by 2060. Under SSP5-8.5, there is an initial increase followed by a decline post-2035, resulting in a reduction to 52.22 million by 2060.

(b)

TEM: All scenarios indicate a long-term trend of increasing temperatures. Under SSP1-2.6, the temperature exhibits variable increases, rising from 16.51 °C in 2020 to 17.50 °C in 2060. Under the SSP2-4.5 and SSP5-8.5 scenarios, temperatures are projected to increase, reaching 18.49 °C in SSP2-4.5 and 18.19 °C in SSP5-8.5, up from 16.46 °C and 17.07 °C, respectively. The SSP5-8.5 scenario exhibits significant fluctuations, indicative of the pronounced variability associated with climate change under high-emission conditions. In the analysis of warming magnitude, the SSP2-4.5 scenario exhibits the highest increase at 2.03 °C, followed by SSP5-8.5 at 1.13 °C and SSP1-2.6 at 0.98 °C.

(c)

PRE: Precipitation patterns indicate that SSP1-2.6 experiences fluctuations, ranging from 886.24 mm in 2020 to 1256.95 mm in 2060. SSP2-4.5 displays a decreasing trend followed by an increase, starting at 959.66 mm in 2020 and reaching 824.64 mm in 2060. Conversely, SSP5-8.5 shows a fluctuating decline, decreasing from 1071.92 mm in 2020 to 1029.45 mm in 2060.

(d)

AGPP: SSP1-2.6 exhibits fluctuating decreases, declining from 693.48 in 2020 to 684.16 g C·m⁻²·yr⁻¹ in 2060. In contrast, SSP2-4.5 demonstrates a consistent increase, rising from 668.46 in 2020 to 732.28 g C·m⁻²·yr⁻¹ in 2060. Similarly, SSP5-8.5 also shows an increase, surpassing the magnitude of SSP2-4.5, with values increasing from 676.28 in 2020 to 793.86 g C·m⁻²·yr⁻¹ in 2060.

3. Results

3.1. Comparison of Provincial Carbon Stock Simulation Results Between the InVEST Model and the I-RF Model

This study, utilizing the InVEST model, determined a total carbon stock of 1.816078 × 10⁹ t C (Figure 4a). The distribution pattern indicates a concentration that is “high in the northeast and low in the southwest” [21]. Higher values are largely distributed across cropland regions, while lower values are chiefly located in urban areas to the south of the Yangtze River and in northern cities like Xuzhou, characterized by elevated levels of urbanization.

The I-RF model output for carbon stock in Jiangsu Province is 1.516438 × 10⁹ t C (Figure 4b), which is similar to the results from the InVEST model alone (R² = 0.3350). While this R² value appears modest, it reflects the complexity of carbon dynamics in heterogeneous landscapes rather than poor model performance. Although this moderate correlation reflects divergence in spatial patterns, it demonstrates I-RF’s enhanced capability to capture heterogeneity, particularly in highly urbanized regions (e.g., southern Jiangsu). The model achieved RMSE = 7308.10 t C and MAE = 5860.62 t C on the independent test set. Furthermore, feature importance analysis revealed the NDVI and population density (PD) as the dominant drivers, signifying the joint control of vegetation coverage and anthropogenic activities on carbon sequestration, while climatic variables provided supplementary explanatory power.

The I-RF model prediction results demonstrate enhanced spatial precision, effectively delineating the spatial heterogeneity characteristics of carbon stock in the study area, especially the pronounced northeast–southwest gradient pattern.

3.2. Spatiotemporal Dynamic Simulation and Trend Analysis of Provincial Carbon Stock Across Various Scenarios

This study simulated the potential carbon stock in Jiangsu Province and evaluated the contributions of three key driving factors: climate, AGPP, and population. In each driving scenario, other non-primary driving factors were held constant at SSP2-4.5, and the variations in that driving factor across the SSP1-2.6, SSP2-4.5, and SSP5-8.5 scenarios were analyzed. SSP2-4.5 was selected as the baseline scenario due to its representation of a moderate emission pathway. This scenario is differentiated from the stringent climate policy constraint of SSP1-2.6 and the elevated emission scenario of SSP5-8.5. This choice mitigates the extreme influence of non-primary driving factors on the analysis outcomes, thereby clarifying the effects of various driving factors.

3.2.1. Carbon Stock Dynamics Under the PPS

Given the negligible spatial distribution variations across the three population change scenarios, the spatial analysis presents results solely from the SSP2-4.5 scenario (Figure 5). From 2020 to 2030, the spatial distribution of carbon stock in the study area showed relative stability; however, between 2035 and 2040, low-value carbon stock areas in the southwestern region displayed a trend of spatial expansion. By 2045–2050, carbon stock in the northeastern section of the study area increased, whereas the southern region exhibited consistently low carbon stock characteristics, suggesting an ongoing imbalance in the regional distribution of carbon stock. By 2055 and 2060, the spatial distribution of carbon stock in Jiangsu Province approached equilibrium, typically remaining within the range of 9000–12,000 t C/ha, without the emergence of significant high-value or low-value clusters. The carbon stock disparities between urban and rural areas exhibited a convergence trend, indicating a gradual decrease in the spatial heterogeneity of regional carbon stock.

Figure 6 shows that, among the nine SSP scenarios, carbon stock variations were most significant under the SSP5-8.5 scenario: carbon stock rose sharply to 1.66 × 10⁹ t C in 2030 and then swiftly decreased to 1.01 × 10⁹ t C by 2040, indicating a reduction of approximately 38.8%, followed by a modest recovery to 1.01 × 10⁹ t C by 2060. The decline in carbon stock under the SSP1-2.6 climate scenario was more gradual, exhibiting a reduction of approximately 24.4% by 2060 relative to 2020 levels. Under identical climate and AGPP scenarios, variations in population scenarios exhibited minimal effects on the carbon stock changes.

3.2.2. Carbon Stock Dynamics Under the PVS

Due to the overall similarity in the carbon stock spatial distribution patterns among the three AGPP scenarios, only the results from the SSP2-4.5 scenario are presented for spatial analysis. Figure 7 illustrates that from 2020 to 2040, low-value carbon stock areas progressively expand from the southwest to the northeast while still exhibiting notable gradient distribution characteristics. This suggests a degree of spatial heterogeneity in the carbon sequestration capacity among various geographical units. Between 2045 and 2050, low-value areas diminished, while high-value areas developed predominantly in the eastern coastal regions, likely associated with the marine environment’s potential impact on carbon stock in these regions. During the period 2055–2060, a near-equilibrium state was observed in the spatial distribution of carbon stock across Jiangsu Province. Low-value regions were mainly concentrated in the southern Jiangsu urban agglomerations, while carbon stock disparities in other areas narrowed, generally stabilizing between 9000 and 15,000 t C/ha. This trend indicates that the interplay of long-term climate change and human activities is leading to a stabilization of carbon stock patterns in Jiangsu Province, suggesting a transition towards a relatively balanced developmental stage.

From a temporal perspective (Figure 8), scenarios in which AGPP varies while climate and population remain constant exhibit overall carbon stock trends that are relatively analogous to those observed in population change scenarios. Prior to 2040, variations in carbon stock between the AGPP and population change scenarios were negligible. Under the SSP1-2.6 scenario, carbon stock in 2030 was recorded at 1.71 × 10⁹ t C, 1.68 × 10⁹ t C, and 1.70 × 10⁹ t C, showing variations of merely 2.05%. After 2040, under the same climate and population scenarios, carbon stock across various AGPP scenarios began to display notable differences, with the SSP5-8.5 scenario showing the most significant variations. In the SSP5-8.5 climate and population scenario for 2060, carbon stock for AGPP scenarios SSP1-2.6, SSP2-4.5, and SSP5-8.5 was measured at 0.97 × 10⁹ t C, 0.99×10⁹ t C, and 1.01 × 10⁹ t C, respectively, indicating differences of approximately 4.7%. Carbon stock fluctuations were notably pronounced under the SSP5-8.5 scenario: with AGPP aligned to SSP1-2.6, the carbon stock peaked at 1.67 × 10⁹ t C in 2030 and subsequently declined sharply to 0.99 × 10⁹ t C by 2040, representing a decrease of approximately 40.7%. Following this, the stock continued to fall, reaching 0.97 × 10⁹ t C by 2060.

3.2.3. Carbon Stock Dynamics Under the CDS

Under the AGPP and population scenarios aligned with SSP2-4.5, notable variations were detected in the spatial distribution of carbon stock across three climate scenarios (SSP1-2.6, SSP2-4.5, and SSP5-8.5). Figure 9 illustrates that, within the SSP1-2.6 scenario, regions with high-value carbon stock (18,000–21,000 t C/ha) were predominantly located in northern cropland areas, especially in the years 2030 and 2040. By 2055, the spatial distribution of carbon stock began to stabilize, and by 2060, there was a notable shift in its distribution gradient, indicating a gradual decline from north to south. The regions located south of the Yangtze River demonstrated consistent low carbon stock traits with slight interannual variations, typically remaining under 1800 t C/ha. In contrast, the northeastern part of Jiangsu Province reliably served as a primary high carbon stock zone, sustaining carbon levels between 15,000 and 21,000 t C/ha. Meanwhile, the central area exhibited notable spatiotemporal changes, indicating a responsive mechanism of land use to climate fluctuations.

As shown in Figure 10, within the SSP2-4.5 scenario, the spatial distribution of carbon stock exhibited a relatively stable pattern from 2020 to 2050, with a general decline observed from the northeast to the southwest. Throughout this timeframe, the spatial range of low-value carbon stock regions demonstrated a notable trend of expansion, as the low-value areas in the southwest progressively extended towards the northeast. From 2045 to 2050, there was a notable change in the spatial pattern of carbon stock, as the expansion direction of low-value areas shifted, indicating a movement from the northeast to the southwest. The eastern coastal areas have shown a significant rise in carbon stock, likely linked to the restoration of vegetation and the optimization of land use in these coastal regions. By 2055–2060, the spatial distribution of carbon stock throughout the study area showed a trend towards equalization, remaining at a relatively stable level overall, although urban areas south of the Yangtze River continued to exhibit comparatively lower carbon stock levels.

Significant spatial differentiation in carbon stock was observed across Jiangsu Province from 2020 to 2045 under the SSP5-8.5 scenario (Figure 11). Other regions, with the exception of the southern urban areas of Jiangsu, exhibited significant fluctuations, indicating significant spatiotemporal differences. In 2025, the northeastern region experienced a substantial decrease in carbon stock. However, by 2030, carbon stock in northern Jiangsu had increased significantly, with the majority of areas reaching 18,000–21,000 t C/ha. Additionally, areas south of the Yangtze River experienced growth in carbon stock. The northern region experienced a decline in carbon stock from 2035 onward, and by 2040, the province’s carbon stock had further decreased, with a spatial distribution that was approaching equilibrium. This trend was indicative of the redistribution characteristics of carbon sequestration capacity across different regions. The northeastern region experienced a degree of recovery in carbon stock by 2045. Consequently, the spatial distribution of carbon stock across the province stabilized from 2050 to 2060, with regional disparities gradually diminishing and overall levels remaining within the range of 15,000–21,000 t C/ha. This suggests that carbon stock gradually stabilized after experiencing significant fluctuations in the previous period. This could be attributed to the dynamic equilibrium of carbon cycling under long-term climate change, as well as the combined effects of ecosystem self-regulation and land-use optimization.

The majority of scenarios experienced fluctuating declines after reaching their peak values around 2030 (approximately 1.45 × 10⁹–1.71 × 10⁹ t C) (Figure 12). Short-term climate change exerted a stronger influence on carbon stock compared to long-term socioeconomic pathways, as demonstrated by the significantly greater differences in carbon stock between various climate scenarios (up to approximately 0.54 × 10⁹ t C) than the differences observed between the various AGPP or population scenarios under the same climate setting. Furthermore, the greatest temporal fluctuations were observed under the SSP1-2.6 climate scenario, while the SSP5-8.5 climate scenario also exhibited significant volatility, though its carbon stock levels were generally lower throughout the study period. From a long-term perspective, the average carbon stock across all scenarios declined from approximately 1.45 × 10⁹ t C in 2020 to approximately 1.02 × 10⁹ t C by 2060, equivalent to a decrease of nearly 30%. By 2060, the carbon stock capacities of various scenarios largely converged within the range of 0.94 × 10⁹–1.17 × 10⁹ t C. The three combinations under the SSP1-2.6 climate scenario exhibited an upward trend.

3.3. Quantitative Assessment of the Sensitivity of Provincial Carbon Stock to Key Driving Factors

Given the significant differences in carbon stock between SSP1-2.6 and SSP5-8.5, this study selected these two scenarios as benchmarks and calculated the elasticity coefficients of carbon stock in response to each driving factor. These coefficients were derived as the ratio of the relative percentage change in carbon stock to the relative percentage change in a single driver, corresponding to three scenario groups: PPS, PVS, CDS. Compared to conventional sensitivity indices, this approach offers a more generalized metric for quantifying the relative influence of drivers and capturing their temporal dynamics.

An analysis of the elasticity coefficients revealed the distinct impacts of population, AGPP, and climate change on carbon stock. As shown in Figure 13, the PPS scenario exhibited minimal variation, with elasticity values ranging from −0.27 (in 2030) to −0.03 (in 2060), suggesting that even under extreme population growth assumptions, carbon stock remains largely unaffected. This further confirms the limited direct influence of population variables on the carbon cycle. In contrast, the AGPP (PVS) elasticity coefficients showed a fluctuating but generally positive impact over time. The elasticity was initially negative in 2020 (−0.05) before becoming positive and rising to a peak of 0.32 in 2055 and then declining slightly to 0.28 in 2060. This trend indicates that while initial changes in AGPP might have a minor negative effect, rising ecosystem productivity in later periods can significantly enhance carbon sequestration capacity.

In comparison, the climate (CDS) scenario demonstrated the most dramatic fluctuations. The elasticity value was −5.09 in 2020, dropping to a minimum of −37.76 in 2045. It then briefly turned positive (1.01) in 2055 before falling again to −3.53 by 2060. These variations underscore the nonlinear, phase-specific, and spatially heterogeneous response of carbon stock to climate change, particularly when only temperature is altered. The sharp decline between 2035 and 2045 highlights the potential for extreme warming under SSP5-8.5 to severely impair carbon sequestration, likely due to mechanisms such as reduced vegetation productivity and heightened water stress.

Spatially, the elasticity of carbon stock also exhibited notable regional variation across Jiangsu Province (Figure 14). Between 2030 and 2040, northern Jiangsu shifted from positive to negative climate elasticity values, indicating a reduced capacity to adapt to warming and an increasing risk of carbon sink degradation. In contrast, between 2040 and 2060, southern Jiangsu showed a growing distribution of positive elasticity values between 0 and 10, suggesting a comparatively stronger capacity for climate adaptation and ecosystem recovery in the south. This spatial heterogeneity highlights the importance of incorporating regional differentiation into future carbon management strategies, with particular attention to the protection and adaptive governance of high-sensitivity areas.

4. Discussion

The carbon stock estimation for 2020 produced by the I-RF model is generally consistent with provincial-level results reported for Jiangsu (1.641008 × 10⁹ t C) [26], suggesting the model’s suitability for regional-scale simulations. Further analysis indicates that the modeled spatial patterns and temporal trends of carbon stock in Jiangsu Province demonstrate significant correlation with existing research findings: carbon loss areas are primarily concentrated in cities south of the Yangtze River, and addressing these losses may require implementing non-natural climate solutions, such as Carbon Capture and Utilization [31]. High-value areas are mainly distributed in eastern coastal cities, primarily benefiting from the superior natural conditions and higher vegetation coverage in these regions, thus forming stronger carbon sequestration capacity [21]. Overall, the I-RF model results accurately reflect the actual situation in Jiangsu Province, demonstrating that this assessment method has high reliability and effectively captures the spatiotemporal dynamic characteristics of carbon stock.

Population density showed a relatively minor impact on carbon stock compared to other factors, which may be because the study only considered population numbers without fully accounting for the dynamic influences of specific human activities. Many studies have confirmed that human activities are the primary drivers of land-cover change, while land-use changes themselves result from the combined effects of natural constraints and socioeconomic factors [19]. Human activities profoundly affect regional carbon stock by altering ecosystem resilience and accelerating urbanization processes [13]. Previous research has found that in areas with less human disturbance, carbon stock loss rates significantly decrease. This study found that population density exerts a minor influence on carbon stock in Jiangsu Province, which contrasts with urban-scale research. For example, studies in Nanjing reported that population size determines resource demand, driving changes in the land-use structure [12], while research in Wuhan identified population density as the main driver of construction land expansion and carbon stock decline [2]. This contrast suggests that carbon stock driving mechanisms at the provincial scale may exhibit significant scale effects. Future research could consider the spatiotemporal variability of carbon stock effects caused by the coupling relationship between dynamic changes in population density and human activity intensity.

Prior studies have identified a notable positive association between GPP and carbon stock [17], indicating that vegetation photosynthetic capacity is an important factor driving carbon stock changes. However, this study found that under the PVS, AGPP had a limited impact on carbon stock, and the sensitivity of carbon stock to AGPP showed an increasing trend over time. Furthermore, the intensity of AGPP’s influence was significantly weaker than climate change factors (which could reach ±11%). This difference may stem from the model’s more comprehensive consideration of carbon cycle complexity. Unlike some studies suggesting that increased AGPP leads to proportional increases in carbon stock, this study indicates that this relationship is significantly modulated by the temporal scale and climate context. For instance, under the SSP5-8.5 scenario, despite continuous increases in AGPP, carbon stock still experienced a sharp decline during 2035–2040. This suggests that under extreme climate conditions, carbon pool decomposition induced by climate change may offset or even exceed the carbon sink effects brought by increased AGPP. This finding emphasizes the limitations of relying solely on enhanced vegetation photosynthesis to strengthen carbon sequestration potential within the framework of climate change.

By examining changes in carbon stocks under the CDS [32], this study reaffirms that climate change is a dominant driver of ecosystem carbon dynamics. Notably, carbon stock exhibited a non-linear response to temperature, first increasing and then decreasing, especially under the SSP5-8.5 scenario. A sharp rise to 1.65 × 10⁶ t C in 2030 was followed by a dramatic drop to 1.01 × 10⁶ t C in 2040. In 2045, a sudden spike in climate elasticity was observed, likely due to the convergence of temperatures under SSP1-2.6 and SSP5-8.5, which reduced the elasticity denominator and inflated the value—a mathematical artifact rather than an ecological anomaly. Additionally, we identified a spatially anomalous pattern: carbon stock in northern Jiangsu declined sharply in 2040 but rebounded by 2045. This may reflect temperature increases and reduced precipitation in 2040, followed by cooler, wetter conditions in 2045. Such fluctuations highlight the complex pathways through which climate variables affect ecosystem carbon cycling. Temperature and precipitation regulate key ecological attributes—such as species composition and vegetation biomass—that jointly determine carbon density [33].

However, in this region, precipitation plays a secondary role. With over 50% of vegetation types classified as agricultural ecosystems, water availability is largely controlled by human irrigation rather than rainfall [19]. This aligns with the relatively stable carbon responses across scenarios and supports the notion of a lag effect in soil carbon pools responding to short-term climate variability. As carbon sequestration is inherently a long-term ecological process, this temporal inertia may buffer abrupt climatic impacts. To guide region-specific carbon sink enhancement, future studies should explore how climate-induced shifts in agricultural ecosystems affect long-term carbon dynamics.

This study combined the InVEST model with a random forest model to create an effective framework for predicting carbon stock. Although this approach did a good job of capturing spatiotemporal patterns of changes in carbon stock, there are still ways to make it even better. To further improve the model’s capacity to replicate intricate ecosystem carbon cycling mechanisms, future research might take into account adding more precise soil biogeochemical parameters, coupled carbon–nitrogen–phosphorus cycling processes, and the severity of human management interventions. The accuracy and dependability of provincial carbon stock estimation models would be enhanced by taking into account additional soil factors in addition to vegetation parameters, giving decision-makers in carbon sink management stronger scientific backing.

The study’s findings offer important scientific and methodological contributions by clarifying the relative importance of different drivers and revealing complex interactions influencing carbon stocks across scales. These results can directly support real-time carbon monitoring, inform adaptive local policy design, and guide urban and regional planning efforts aimed at carbon management and climate resilience.

Based on the above findings, the following policy implications are proposed for Jiangsu Province:

(1)

Protect and monitor high-carbon-value areas such as the northeastern cropland and eastern coastal natural ecosystems, where carbon stock remains relatively high and stable under various scenarios, particularly considering their critical role under the CDS. Establish long-term carbon monitoring networks in these regions to track changes and guide adaptive management. Additionally, promote renewable energy and reduce industrial emissions in high-emission areas by integrating carbon targets into local plans and offering subsidies or tax incentives.

(2)

Restrict urban expansion into high-quality farmland, especially in the northeast, to prevent further carbon stock decline as indicated by the stable yet vulnerable carbon distribution patterns under the population-dominated (PPS) scenario. Concurrently, expand urban green space systems to enhance urban carbon sequestration capacity, addressing the relatively low carbon stock observed in southern urban agglomerations.

(3)

Implement adaptive land management and ecological restoration strategies in the central and southern regions where carbon stock shows greater variability and lower values under high-emission climate scenarios (SSP5-8.5). This includes constructing green infrastructure corridors to improve ecosystem connectivity and resilience, which is crucial given the observed spatial heterogeneity in climate sensitivity and the potential degradation of carbon sinks. Additionally, adopt sustainable farming practices to enhance ecosystem productivity and soil carbon storage. In eastern coastal regions, prioritize high-carbon vegetation planting and wetland restoration to maximize the marine environment’s positive impact on carbon stocks. Establish long-term GPP monitoring networks in areas with significant spatial heterogeneity (e.g., southwestern and northeastern Jiangsu) to track ecosystem productivity changes and guide adaptive management.

(4)

Implement ecological restoration in low-carbon southern urban areas identified as climate-sensitive, focusing on degraded wetlands, forests, and agricultural lands. Establish ecological corridors linking these areas to high-carbon coastal regions to improve connectivity and enhance carbon sequestration, especially between 2040 and 2060 when recovery potential is highest. Priority should be given to regions with high climate elasticity, particularly northern Jiangsu areas where sensitivity shifted from positive to negative between 2030 and 2040.

5. Conclusions

This study developed a novel and replicable approach by integrating the InVEST ecological model with a random forest algorithm to simulate carbon stock dynamics in Jiangsu Province from 2020 to 2060. Compared to the standalone InVEST model, the coupled model improved accuracy while preserving spatial pattern consistency. A scenario-splitting framework was also constructed to independently quantify the effects of climate change, ecosystem productivity, and population density via a control variable design. By combining ecological process mechanisms with machine learning’s nonlinear data-mining capabilities, this approach advances regional carbon stock modeling and addresses limitations of traditional single-scenario studies. The results show that Jiangsu’s carbon stock in 2020 was 1.516438 × 10⁹ tons and is projected to decline by 24.4% by 2060 across all SSP scenarios, following a spatial pattern of “higher in the northeast, lower in the southwest.” Among the drivers, climate change had the greatest impact (elasticity coefficient −37.76–5.09 to 1.01), followed by annual gross primary productivity (−0.05 to 0.32), while population change had the minimal direct effects (−0.27 to −0.03). Notably, under the SSP5-8.5 scenario in the PPS, carbon stock declined by up to 38.8% between 2030 and 2040, highlighting the vulnerability of regional carbon sinks to high-emission pathways. These findings contribute to environmental sustainability research by providing technical support for decarbonization planning and sustainable carbon management in fast-developing urban regions.

Despite these strengths, several limitations should be acknowledged. First, the projection of future variables (e.g., climate, population, AGPP) relies on CMIP6 models, which inherently involve uncertainties related to emission scenarios and model structures. These may propagate through the simulation chain and affect the reliability of carbon stock forecasts. Second, certain datasets were resampled from coarser resolutions (e.g., 0.25°) to 1 km using IDW interpolation. Although this ensured spatial alignment, it may introduce moderate spatial uncertainty in heterogeneous areas.

Future research could further improve the model by (1) developing integrated human–environment system frameworks that better represent interactions between land use, human activities, and carbon processes; (2) incorporating more refined soil biogeochemical and vegetation physiological parameters to enhance model realism; and (3) embedding policy-sensitive variables to enable differentiated carbon forecasts under various governance pathways.