Dynamic response of carbon storage to future land use/land cover changes motivated by policy effects and core driving factors

Abstract

The evolution of land use/land cover (LULC) patterns significantly influences the dynamics of carbon storage (CS) in terrestrial ecosystems. In response to future environmental changes, however, most studies fail to synthesize the effects of policy pathways and evolving core driving factors on LULC projections. This article presents a systematic framework to assess the dynamic response of the terrestrial ecosystem CS to future LULC changes. After investigating spatiotemporal characteristics and driving forces, policy effects and future core driving factors are integrated into the improved Markov–future land use simulation model to project LULC across diverse scenarios. Then the Integrated Valuation of Ecosystem Service and Tradeoff model is coupled to explore CS dynamics with LULC changes. This framework was applied to the Weihe River Basin. The finding reveals that the overall proportion of cultivated land, forestland and grassland is above 85% and is significantly influenced by policy effects. Precipitation, temperature, population density and gross domestic product are core driving factors of LULC changes. Equal-interval projection is a viable approach to mitigate policy impacts by avoiding error propagation while coupling future core driving factors to improve LULC projection accuracy. Ecological protection should be emphasized in the future. The rate of increase in CS is 1.25 and 1.63 times higher than the historical trend and economic development scenario, respectively, which alleviates carbon loss from the expansion of built-up land. This research provides a valuable reference for future insight and optimization of ecological conservation strategies.

摘要

政策效应与核心因子驱动下的未来土地利用/土地覆被变化对碳储存的动态影响

土地利用/土地覆被(land use/land cover, LULC)变化显著影响着陆地生态系统碳存储。然而，在响应未来环境变化时，政策路径和不断变化的核心驱动因子的综合作用往往被忽视。本研究提出了一个系统框架以评估陆地生态系统碳储存对未来LULC变化的动态响应。在探究LULC变化时空特征和驱动作用的基础上，将政策效应和未来核心驱动因子整合到改进的Markov–FLUS模型中，预测未来不同情景下的LULC；并耦合InVEST模型探索LULC与碳储量的动态特征。该框架被应用于渭河流域，结果表明，耕地、林地和草地面积总占比超过85%，且受到政策显著影响。降水、温度、人口密度和GDP是LULC变化的核心驱动因素。等间隔预测可以有效减轻政策效应导致的误差迭代，并且通过耦合未来核心驱动因子提高了LULC预测的准确性。未来应注重生态保护，其碳储存增长速度分别是历史趋势和经济发展情景的1.25和1.63倍，减缓了建设用地扩张所导致的碳损失。本研究为未来深入了解和优化生态保护策略提供了宝贵的参考。

INTRODUCTION

The rapid increase in carbon emissions has triggered a series of extreme events, including global warming and extreme climate phenomena (Liu et al. 2022a), posing severe threats to ecological environments, human survival and economic development (ED) (Wang et al. 2022b). To address this challenge, China is actively implementing measures such as carbon capture, storage and other emission reduction technologies to achieve carbon neutrality by 2060 (Du et al. 2023). Carbon storage (CS) in terrestrial ecosystems plays a pivotal role in the global carbon cycle, enhancing future ecosystem services and mitigating climate change effects (Li et al. 2020; Xu et al. 2023). Changes in land use/land cover (LULC) are vital factors that have an impact on terrestrial ecosystems CS (Chuai et al. 2013). These changes encompass both natural processes, such as ecosystem variations caused by climate change (Wang et al. 2022b), and human activities, such as the implementation of ecological restoration policies (Oliveira et al. 2019; Yang et al. 2023) and the rapid urbanization (Tang et al. 2020). These factors combine to exacerbate LULC transformation frequency and pose a challenge to the accurate assessment of future CS. Therefore, investigating the response of CS to future LULC changes and its spatiotemporal evolution considering climate change and policy pathways is crucial for formulating carbon sequestration technology routes and regional prioritization strategies to achieve carbon neutrality.

Multi-scenario modeling is effective for comparing the potential consequences of current decisions and future environments. It is useful to identify potential problems and conflicts in CS under future LULC changes by comparing and evaluating different scenarios (Zhang et al. 2022). Xiang et al. (2022) assessed the response of CS to LULC in Chongqing using Integrated Valuation of Ecosystem Service and Tradeoff (InVEST) by setting different policy scenarios. Yang et al. (2020) assessed and projected changes in LULC and CS in Hubei Province from 2015 to 2030 considering different climate scenarios selected in the Coupled Model Intercomparison Program Phase 6 (CMIP6). These studies have realized the multi-scenario analysis of future CS dynamics by considering policy pathways or climate change, and their results have facilitated different local decisions for development. However, previous studies have lacked an exploration of the potential impact of the combined effects of policy pathways and climate change on future LULC projections under changing environments, which will cause biases in both quantitative requirements and spatial distribution.

The Markov model is extensively employed for projecting LULC requirements, primarily due to its lack of posteriority and cost-effective modeling (Liang et al. 2021; Xiang et al. 2022). Determining the transfer probability matrix of LULC between two different periods is key to future LULC requirement projection and the selected two periods cause the effect of the transfer rate and iteration number will have an impact on the LULC requirement prediction accuracy (Liang et al. 2021). From 1978 to 2000, the Chinese government launched numerous programs, such as ‘Comprehensive Agricultural Development Program’, ‘Soil and Water Conservation Program’, ‘Grain for Green Program’, etc., to address the serious condition of land or ecological systems (Bryan et al. 2018). These programs are characterized by high investment and duration, with remarkable results in the early stages (Yang et al. 2023). All of these have changed the rate of transformation and distribution patterns of the involved landscapes (Liu et al. 2008). The abrupt perturbations can upset the existing system balance, and a new balance will take longer to establish (Wang et al. 2023). However, previous studies have neglected the impact of policy effects. Furthermore, future policy pathway options will also alter the transformation rates of different LULC types, all of which should be taken into account in LULC requirement simulations and projections.

In spatial simulation, future land use simulation (FLUS) can better capture spatial distribution patterns of LULC by coupling the spatial distribution characteristics of driving factors, such as natural and human activities (Liu et al. 2017). Furthermore, it allows for the establishment of distinct transformation rules according to different policy pathways (Liang et al. 2018; Liu et al. 2022b). Driving factors dominate the probability of occurrence of each LULC type, so it is clear that the failure to incorporate the core driving factors varying across scenarios into the future LULC spatial projections weakens the applicability of the projected LULC to the future scenarios (Wu et al. 2022; Zhang et al. 2022). Therefore, identifying the core driving factors of LULC is key for LULC spatial distribution patterns projection. Wu et al. (2023) identified the driving factors of LULC using a geographical detector for each LULC period. Zhou et al. (2020) determined the ranking of the main driving factors of urban development through random forest according to the initial LULC. Notably, previous studies have mostly focused on identifying the driving forces over a single period. However, LULC change is a dynamic process. Focusing only on LULC in one period will cause high uncertainty in driving force identification. The average of the driving forces has the potential to weaken the driving effect of the factors. The comprehensive effect of multiple periods of LULC should be considered in the identification of driving forces to improve future LULC spatial projections.

Coupling Markov and FLUS (Markov–FLUS) can effectively reconstruct the spatial patterns of future LULC under comprehensive consideration of different policy effects and climate change scenarios (Zhang et al. 2022). Nevertheless, the impact of policies on the LULC transformation rate and change in core driving factors on the LULC distribution pattern cannot be ignored under the future changing environment. This is extremely essential to the adaptability of scenarios and the accurate assessment of CS (Liu et al. 2022a; Wu et al. 2022). Therefore, this study establishes a systematic framework based on random forest, improved Markov–FLUS and InVEST model for LULC projection and CS assessment under different future scenarios. The random forest is employed to identify core driving factors, the improved Markov–FLUS model is utilized for LULC projections across various scenarios, and the InVEST model is applied for CS assessments. The framework was applied to the Weihe River Basin (WRB) in Northwest China, where LULC is more prominently affected by climate and policy (Wu et al. 2023). The objectives of this study were to (i) describe the characteristics of LULC change in the WRB under policy effects and identify the core driving factors with the transformation frequency of multi-year LULC; (ii) project LULC under different scenarios with the influence of policy pathways and future core driving factors; and (iii) analyze the spatiotemporal dynamics of CS and the response of CS to changes in LULC under different scenarios. The established framework can provide valuable insights for policymakers to formulate targeted and sustainable optimization strategies for LULC to achieve sustainable development.

STUDY AREA AND DATA PROCESSING

Study area

The WRB is located in northwestern China and is part of the Yellow River Basin, with an area of 134 000 km². The Wei River, which is 818 km long, is the largest tributary of the Yellow River. The WRB is a semi-humid region with an average annual temperature and precipitation of 7.8–13.5 °C and 400–800 mm, respectively (Wu et al. 2022). The Qinling Mountains and Loess Plateau are located south and north of the WRB, respectively. The Loess Hills and Gullies are in the middle and upper reaches, and the Guanzhong Plain is in the lower. Fig. 1 shows the geographical location of the WRB.

Data sources and processing

This study used the global land cover product with a fine classification system at 30 m from 1985 to 2022 (Zhang et al. 2021), obtained from the Big Earth Data Science Engineering Program (https://data.casearth.cn), which was further classified into six categories: cultivated land, forestland, grassland, built-up land, bare land and water body (Wu et al. 2023). Precipitation and temperature are station data from the National Meteorological Science Data Center (https://data.cma.cn), which are interpolated to 1 km resolution raster data after calculating multi-year averages. Elevation data were produced from ASTER GDEM 30 m resolution digital elevation data from Geospatial Data Cloud (http://www.gscloud.cn). The slope was produced from elevation data. The fundamental national geographic information (including national highway, railway, provincial highway, expressway and rivers) is obtained from the National catalogue service for geographic information (https://www.webmap.cn). Distance information was obtained from vector data using Euclidean distance processing. Furthermore, all the map data were processed into spatially consistent 1 km raster data by resampling and spatial interpolation techniques (Wu et al. 2023). Table 1 and Fig. 2 present data sources and driving factors spatial distribution, respectively.

Future gross domestic product (GDP) and population density (POP) data were obtained from different scenarios raster datasets of 1 km × 1 km from Murakami and Yamagata (2019) and Chen et al. (2020), respectively. GDP data are selected for 2020–50 with intervals of 10 a. POP is selected for 2025–40 with intervals of 5 a. Future temperature and precipitation data were obtained from the CMIP6 dataset in the World Climate Research Program (https://esgf-node.llnl.gov/search/cmip6). Data for 2020–40 were selected and multi-year averages were calculated at 5 a intervals. Specifically, precipitation and temperature data from different climate models were downscaled to a 1 km × 1 km resolution based on historical multi-year average data using the Delta method (Xu and Wang 2019). Supplementary Figs S2–S5 show the datasets for GDP, POP, temperature and precipitation for future scenarios.

METHODOLOGY

In the process of estimating CS based on LULC, the reliable spatial distribution pattern of LULC plays a decisive role. To this end, a combined modeling framework was constructed to support the accurate projection of future LULC and its effects on CS. As shown in Fig. 3, the framework includes LULC spatiotemporal characterization, core driving factors identification, LULC projections and future CS dynamic assessment. Specific implementation methods and processes are described in detail below.

Identification of core driving factors with random forest

Random forest is an integrated classifier consisting of several individual decision trees (Wu et al. 2021). It has a high tolerance for random variables and can resolve multicollinearity in high-dimensional data via the random sampling of samples or features (Zhou et al. 2020). Furthermore, the random forest can easily measure the relationship between the input and output variables so that the driving effect of each influencing factor on the change in LULC can be quantified (Cui et al. 2023). It mainly measures the driving force by reducing the Gini index, calculated as follows (Algehyne et al. 2022):

where DF_r is the importance of the rth influencing factor on the change in LULC, representing the value of the driving force; R, I and J are the total number of factors, number of decision trees and number of nodes of a single decision tree, respectively; and D_GI,rij is the reduction value of the Gini index of the rth factor at the ith node of the jth decision tree.

This study utilizes Scikit-learn to build the random forest model (Pedregosa et al. 2011). The main parameters for model construction are the number of decision trees (n_estimators) and the maximum depth of the trees (max_depth). Multiple parameter combinations have been tried, and finally, the parameter combination that guarantees both efficiency and classification accuracy (>0.85) has been determined, with the specific settings as follows: n_estimators = 90, max_depth = 15 and other parameters at default.

Modeling for LULC projection

LULC requirement projection based on improved Markov

Markov model is a mathematical statistical model based on transfer probability i.e. frequently used in LULC requirement projection because the dynamic evolution process of LULC has notable Markov characteristics (Xiang et al. 2022). To reduce the limitation of the unidirectional transfer function of the Markov model, the scenario weight matrix, W_n = Diag(w₁, w₂, …, w_n,), is introduced to explore potential changes in the area of LULC types under different scenarios. The equation of the improved Markov model applied to different scenarios is as follows (Liang et al. 2021):

where S_k+1 and S_k represent the LULC state vectors at k + 1 and k, respectively; ${p^{″}}_{cl}$ is the normalized correction result of ${p^{\prime }}_{cl}$ using Equation (3), as the sum of transfer probabilities for each LULC type is not equal to 1 in ${p^{\prime }}_{cl}$⁠; Diag() is a symbol for diagonal matrix; ${p^{\prime }}_{cl}$ is the transfer probability matrix of p_cl modified by W_n for different scenarios; W_n represents the scenario weight matrix, w_n > 1 indicates an accelerating trend of transitions; 0 < w_n < 1 indicates a decelerating trend; and w_n = 1 indicates no change in the initial transfer probability; p_cl is the transfer probability of LULC type c at initial period transforming into type l at end of study (⁠$0\le p_{cl}\le 1,{\sum }_{l=1}^n{p}_{cl}=1$⁠), and n represents the total LULC type.

LULC spatial distribution projections based on FLUS

The FLUS model is an LULC spatial projection model integrating an artificial neural network (ANN) algorithm and a cellular automata (CA) model (Liang et al. 2018; Liu et al. 2017). ANN is used to obtain the suitability probabilities based on driving factors. The change in the location of each LULC type is guided by the suitability probability values at each pixel. ANN contains an input layer, a hidden layer and an output layer. The calculation formula is as follows:

where P(k,t,l) is the suitability probability of LULC type l at time t on grid cell k, $\sum P(k,t,l)=1$⁠; w_j,l is the weight between the hidden layer and input layer (driving factors); net_j(k, t) represents the signal received by neuron j in the hidden layer at time t, and sigmoid() is the activation function between the hidden and output layers.

The next step was to introduce a roulette selection self-adaptive inertia competition mechanism into the CA model. This enhances the ability to simulate randomness and uncertainty in LULC transformation. It allows the integration of the self-adaptive inertia coefficients, neighborhood weights, transformation costs and suitability probability to achieve rational spatial distribution of each LULC type. The self-adaptive inertia coefficient is the core of this mechanism and the equation is as follows:

where ${\mathrm{I}\mathrm{n}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{t}\mathrm{i}\mathrm{a}}_l^t$ is the self-adaptive inertia coefficient for LULC type l at time t and $D_l^{t-2}$ and $D_l^{t-1}$ are the differences between the current and future requirements for LULC type l at time t − 2 and t − 1, respectively.

Consequently, the combined probability of a certain LULC type for a grid cell can be expressed as:

where ${\mathrm{T}\mathrm{P}\mathrm{r}\mathrm{o}\mathrm{b}}_{k,l}^t$ is the combined probability of grid cell k for conversion from the original LULC to the target l at time t; ${\mathrm{\Omega }}_{k,l}^t$ represents neighborhood weights and $\mathrm{c}\mathrm{o}{\mathrm{n}}_{c\to l}$ represents the transformation cost, which defines the possibility of transition from the original LULC type c to the target l.

Spatiotemporal effects under the influence of policies and core driving factors

Generally, countries and governments will formulate relevant policies in response to sustainability emergencies when land systems are at risk and in emergency situations (Bryan et al. 2018; Liu et al. 2008). Especially in the early stages of policy implementation or development may result in an abrupt transformation of LULC type, where there is a great potential for transformation, together with the influence of human subjective initiative (Du et al. 2023; Wu et al. 2023). Over time the effectiveness of the relevant policies becomes apparent and the emergency situations are mitigated, a new balance is established, and a relatively stable transformation rate is achieved for each LULC type (Liu et al. 2022b; Xiang et al. 2022). Both abrupt changes and unstable growth rates of the LULC types will undoubtedly affect LULC projections by increasing the error propagation.

In this study, after clarifying the abrupt change year of LULC, different time scales are selected for the same target year LULC projection to examine the impact of transformation rate on LULC projection. LULC in 2000 was chosen as the initial year to project the LULC requirements in 2020, setting the projection intervals at 5 years (⁠${p^{\prime }}_{cl}$based on LULC in 2000 and 2005) and 10 years (⁠${p^{\prime }}_{cl}$based on LULC in 2000 and 2010), respectively. When the projection interval is small and unequal, the Markov model performs multiple iterations at the same transition rate to project the target year’s LULC requirements. Therefore, the unequal-interval projection (e.g. 5-year) requires four iterations, whereas the equal-interval projection (e.g. 10-year) requires only two iterations. The appropriate selection prediction interval is effective in responding to policy effects. Comparative analysis by selecting different projection intervals allows assessment of their applicability to LULC requirement projections in the context of human activities and policy impacts.

The core driving factor change dominates the suitability probabilities of each LULC type (Liang et al. 2018). Therefore, the future core driving factors especially for changing significantly should be considered to enhance the accuracy and appropriateness of LULC projection. Comparative analyses were performed with previous studies (Supplementary Fig. S1) without considering the core driving factors, but only qualitative comparative analyses were undertaken due to the different methodologies and study scales (Wu et al. 2022). The comparative analysis can assess the reasonableness of the LULC spatial distribution pattern projections considering the core driving factor changes.

LULC projections in future different scenarios

Three threshold future scenarios including historical trend (HT), ED and ecological priority (EP) were designed to explore future LULC dynamics by combining future climate scenarios and potential policy effects. Detailed descriptions of the different scenarios are provided in Supplementary Material. According to the policy response characteristics, different growth rates for each type of LULC and transfer cost matrices are given for different scenarios to respond to the policy effect.

Future LULC requirements projection in various scenarios can be realized by setting different W_n (Equation (4)). In HT, different LULC types have experienced consistent transformation rates with history, W_n = Diag(1.00, 1.00, 1.00, 1.00, 1.00, 1.00), representing the relative transformation rate of cultivated land, forestland, grassland, built-up land, bare land and water body, respectively; In ED, however, socioeconomic development and population growth have triggered an increase in material demand and accelerated urbanization, resulting in increased transformation rates of cultivated and built-up land, W_n = Diag(1.10, 0.90, 0.90,1.20, 0.90, 1.00). Conversely, recognizing the ecological importance can enable ecoenvironment protection through policy guidance or sustainability measures, resulting in an increased transformation rate of ecological land, W_n = Diag(0.95, 1.20, 1.20, 0.85, 1.00, 1.00) in EP.

For the spatial distribution, the suitability probabilities for different scenarios are obtained by considering core driving factors. The neighborhood weights (Supplementary Table S1) and transformation cost matrix (Supplementary Table S2) were set to reflect the specific assumptions under different policy pathways (Wu et al. 2022). Consequently, this model framework thoroughly evaluates the effects of diverse scenarios on future LULC from quantitative requirements and spatial patterns. It is instructive in revealing the potential future impact of existing decision processes.

CS assessment based on InVEST

The CS module of the InVEST model was used to quantify the CS in this study. The equation for calculating CS is as follows (Liang et al. 2021; Wang et al. 2022b):

where C_total is the total CS (Mg), S_i is the area of LULC type i (ha), n is the number of LULC types, C_i,above is the aboveground biomass carbon density of i (Mg/ha), C_i,below is the belowground biomass carbon density of i (Mg/ha) C_i,soil is soil organic carbon density of i (Mg/ha) and C_i,dead is the dead organic carbon density of i (Mg/ha).

Carbon density data for various LULC types are essential for CS quantification (Chuai et al. 2013). The carbon density primarily refers to previous studies in the basin where the WRB is located, and most of the values were corrected and tested (Hu et al. 2023; Liang et al. 2021; Wang et al. 2022a; Xu et al. 2023). In summarizing the trend of carbon density changes of different LULC types, the carbon density applicable to this study area was acquired through comparative analysis (Table 2).

RESULTS

LULC distribution characteristics in WRB

From 1985 to 2020, cultivated land was always the major LULC type in the WRB, accounting for more than 45%, followed by forestland and grassland, both accounting for more than 20% (Fig. 4). The spatial distribution of LULC in the WRB features strong spatial heterogeneity, with cultivated land and built-up land mainly clustered in the plains around the Weihe River mainstream, forestland in the Qinling Mountains and Luo River Basin (LRB), and grassland in the Ching River Basin (CRB) and upper reaches of the LRB (Fig. 4a). Notably, all LULC types showed a higher dynamic change from 1985 to 2020, showing a distinct abrupt trend ca. 2000 (Fig. 4b). For example, both cultivated land and forestland had opposite trends before and after 2000, with cultivated land rising and then falling, and forestland showing the reverse. Other land types had the same trend, but the discrepancy in the growth rate after the abrupt change was significant. Specifically, built-up land remained in a stable growth trend, but the growth rate after 2000 was significantly faster than that before 2000, with the area proportion increasing from 1.00% to 4.00%.

Driving forces analysis with LULC transformation frequency

The driving force of each factor on the LULC transformation frequency was obtained by random forest, as shown in Fig. 5. Fig. 5a shows the five transformation frequencies calculated from LULC from 2000 to 2020. The area of no transformation accounted for more than half of the total, accounting for 68.59%; one-time transformation accounted for 21.21%; two times transformations accounted for 9.04%; and three and four times transformations were less, accounting for less than 1.00%. High-frequency transformations occur mainly in the upper of the CRB. One and two times transformations were dispersed from the global distribution in the WRB. However, the one-time transformation was considerably more concentrated in the Guanzhong Plain, which is highly related to the irreversibility of built-up land expansion. As a result, GDP and POP associated with human activities show a higher driving force (Fig. 5b). Human activities shaped artificial landscape transformation, such as cultivated land and built-up land, which accounted for a higher and wider transformation in WRB (Wu et al. 2023). Precipitation and temperature also have a strong driving force at various transformation frequencies, which is associated with a high proportion and transformation among cultivated land, forestland and grassland. Moreover, precipitation and temperature are key factors that promote vegetation growth in WRB (Liang et al. 2021). Additionally, factors such as distance from traffic and topography (elevation and slope) also exhibit potential impact on LULC transformations. This is because topography can affect the transformation of cultivated land and the growth of vegetation (Wang et al. 2022b), while infrastructure such as expressways and railways can dramatically change the LULC landscape (Wu et al. 2023).

LULC simulation and projection

Comparative of LULC simulations accuracy with different time scales

The LULC types fluctuated drastically before 2000 due to the prior influence of relevant policies (Fig. 4), which caused visible errors in the LULC simulation. Thus, for a more distinct comparison of different time scale effects, the dataset for LULC after 2000 was selected to minimize the impact of abrupt fluctuations on LULC projections. Table 3 and Fig. 6 present the simulation results of the LULC requirement and spatial distribution in 2020 under different projection intervals with an initial study period of 2000. From Table 3, the errors of the equal-interval simulations were all less than those of the unequal-interval. The average absolute error for the equal-interval was approximately 4.18%, which was even smaller than the unequal-interval of 14.24%. Specifically, the gap in the simulation error was greater for forestland, grassland and bare land, ranging from 2.30 to 7.00 times. This is related to the higher dynamic changes in these three types from 2000 to 2005 (Fig. 4), which increased the error propagation in unequal-interval projections, resulting in poorer simulation results.

Kappa and the figure of merit (FoM) are effective evaluation indicators for the simulation results of the spatial distribution of LULC (Zhang et al. 2022). As shown in Fig. 6, the projected distribution pattern of LULC at equal-interval had good spatial consistency with the actual distribution in 2020. For different projection intervals, the values of Kappa were above 0.8, with 0.81 for the unequal-interval and 0.89 for the equal-interval. However, the FoM values for the equal-interval (FoM = 0.46) were significantly better than those for the unequal-interval (FoM = 0.09). This is because the model simulation performance and accuracy gradually decrease with an increase in the number of iterations in the unequal-interval. Therefore, equal-interval projection can effectively reduce error propagation and improve the accuracy of LULC simulations and projections.

LULC projections under different scenarios

According to available observations from 2000 to 2020, LULC requirements for 2025, 2030, 2035 and 2040 under different future scenarios were obtained by an improved Markov model with equal-interval projections of 5, 10, 15 and 20 years, respectively (Fig. 7). As shown in Fig. 7, the requirements for each LULC type varied distinctly in the three different scenarios. Forestland, built-up land and water body showed an increasing trend. Forestland and water body especially had a significant increase in the EP while built-up land had the highest speed of expansion in the ED, with the proportion increasing from 4.9% in 2020 to 6.3% in 2040. The areas of cultivated land, bare land and grassland showed different levels of decline. In the three scenarios, the decreasing trend in cultivated land ranked EP > HT > ED. Grassland decreases slowly in EP and significantly in the ED. Exceptionally, there is an increasing trend after 2035 under both EP and HT conditions. The disparity in bare land is not visible under different scenarios because of the smaller area.

The LULC distribution patterns under different scenarios have been projected by FLUS based on LULC requirements, setting parameters and future core driving factors. As shown in Fig. 8, the results showed that the expansion of built-up land is notable, mainly concentrated in the plain areas in the middle and lower reaches of the Weihe River mainstream, tending to expand gradually along the direction of the river and road distribution. In particular, the distribution of built-up land was more clustered in the ED than in the two other scenarios, mostly transferred from cultivated land and grassland. Existing forestland is better protected in EP, with higher forestland cover than HT and ED.

CS assessment and dynamics characteristics

CS assessment in different scenarios

The results of CS for each LULC are shown in Table 4, while the intermediate variables (CS at aboveground, belowground, soil organic and dead organic) in the calculations are detailed in Supplementary Table S3. The total CS showed an increasing trend in the WRB under different scenarios; however, the growth rate varied from scenario to scenario. For example, CS increased fastest in the EP, with an average growth rate of 2.00%/5 a, followed by the HT with 1.60%/5 a, whereas the average growth rate slowed significantly in the ED, at 1.23%/5 a. Furthermore, the amount of CS exhibited by each LULC type varied dramatically across scenarios. Specifically, the amount of CS in cultivated land tended to decrease in all three scenarios, with average decline rates of 5.73%/5 a, 4.08/5 a and 2.73%/5 a for EP, HT and ED, respectively. The CS of forestland and grassland showed an increasing trend in EP as a whole, higher than in HT and ED. The CS of built-up land showed an increasing trend; however, the greatest trend was observed in the ED. The CS change in bare land and water body was not significant, even though they produced large biases in the modeling (Table 3). Bare land and water body account for a low percentage overall, causing relatively little quantity bias. Meanwhile, due to the weaker CS capacity of both (Table 1), the impact on the overall CS is minimal.

From the spatial distribution of CS from 2025 to 2040 under different scenarios (Fig. 9), the overall distribution of CS in the WRB showed moderate changes under different scenarios, but the CS in different sub-basins was divisible and had strong spatial heterogeneity. Specifically, higher CS is mainly concentrated in the LRB and Qinling Mountains. Lower CS is mainly concentrated in the Guanzhong Plain, which is located in the middle and lower reaches of the Weihe River. Additionally, areas with lower CS have become increasingly concentrated in the Guanzhong Plain over time with the expansion of built-up land in different scenarios.

Spatial autocorrelation of CS

Figs 10 and 11 show Moran’s I scatterplot (P < 0.05) and spatial correlation analysis, respectively. From 2025 to 2040, global Moran’s I shows an increasing trend in different scenarios and overall maintains around 0.7, indicating that the CS in the WRB has an obvious spatial aggregation and is gradually increasing. The percentage of high–high (H–H) clustering in EP is more than in ED, which indicates that EP better contributes to CS than ED. Furthermore, it is found that most of the points are located in the first quadrant (H–H) and the third quadrant (low–low, L–L), and the number of L–L is redundant to H–H, but the H–H sample points with high significance are further distances from the coordinate origin. This indicates that the H–H clustering effect of CS in the Weihe River is more remarkable but the overall level of CS is relatively low. Similar results can be seen in the spatial correlation analysis (Fig. 11). The hotspot areas were mainly concentrated in the LRB and Qinling Mountains, which are related to the distribution of forestland. Cold spot areas were concentrated in the Guanzhong Plain, which is related to the distribution of built-up land. This further indicated that the CS was highly consistent with the spatial distribution of the LULC.

Spatiotemporal characteristics of CS dynamics

To explore the spatiotemporal characteristics of CS dynamics, the change of CS from 2025 to 2040 compared with the base year (2020) under different future scenarios was investigated based on Emerging Hot Spot Analysis (Fan et al. 2022). Fig. 12 shows the distribution of the spatial variation of CS dynamics and the spatiotemporal hotspot analysis. As shown in Fig. 12A, the transition between CS and carbon loss became more frequent and widespread over time. Specifically, the area of change in the CS gradually expands under different scenarios. From 2025 to 2030, the changes are mainly concentrated in the northern of the WRB and Guanzhong Plain. At the beginning of 2035, the area of change continuously expanded, occurring in the entire basin by 2040.

In Fig. 12B, the hotspot area of CS dynamics in EP is notably greater than that in the other two scenarios, indicating that the CS in EP is superior to that in HT and ED. The most severe carbon loss occurs in ED, there was an evident spatial concentration and progressive intensification of cold spot areas. Compared with EP and HT, new cold spots tended to expand toward the LRB and CRB. It is worth noting that the CS dynamics in the Qinling Mountains are not noticeable, which is mainly related to the lower POP, and the suitable geographic and climatic environments resulting in a high degree of landscape dominance and better ecological protection. Whereas in the LRB and Guanzhong Plain LULC transformation is relatively frequent, causing the new hot and cold spots of CS dynamics to be concentrated.

DISCUSSION

LULC projection with policy effect and driving factor contributions

Policy-driven abrupt and dramatic LULC transformation

LULC types in the WRB had a significant abrupt change in their trend ca. 2000. The reasons for this may be related to a series of policies implemented before 2000, such as the ‘Rural Land Joint Production Contract Responsibility System’, ‘Grain for Green Project’ and a series of measures for ecological conservation (Bryan et al. 2018; Xu et al. 2023). The overall proportion of cultivated land, forestland and grassland is above 85% in the WRB, which further leads to a more prominent influence by policy. Meanwhile, urbanization development programs have also contributed to the expansion of major cities (Wu et al. 2023). Policies related to LULC are proposed to improve the ecoenvironment or to meet development requirements (Bryan et al. 2018). However, the policies implemented have upset the original balance of the LULC system, causing abrupt and dramatic LULC transformation and altering the rate of conversion. These should be considered in LULC projections.

A comparison of the simulation validation revealed that equal-interval projection reduces the potential influence of dramatic increases or decreases for a particular LULC type in a given period. That allows consideration of the comprehensive influence of multi-period transformations in LULC from near to far instead of iterating based on a constant evolutionary trend (Li et al. 2020), and effectively reduces the number of iterations. Therefore, the equal-interval projection is more consistent and applicable to the characteristics of LULC transformation under the influence of policies, which is effective in avoiding error propagation and conducive to improving the credibility of the projection results (Liu et al. 2022b).

Advantages of incorporating future core driving factors in LULC projection

Precipitation, GDP, POP and temperature are the core driving factors for changes in LULC in the WRB, which is highly consistent with the findings of previous studies (Wu et al. 2022; Zhang et al. 2022). However, it is worth noting that this study applies the LULC transformation frequency as the dependent variable, taking into account the multi-year combined effect. It outperforms previous studies that measured the driving effect of factors on LULC only in terms of the driving force of factors on one-period LULC or the average driving force of factors on multi-period LULC (Chang et al. 2022; Wu et al. 2023). It emphasizes the contribution of factors to the number of LULC transformations.

The spatial distribution patterns of LULC under different scenarios are projected by coupling the future core driving factors, which was relatively reasonable compared with the results of Wu et al. (2022). Specifically, built-up land in the WRB was mainly distributed in the Guanzhong Plain under different scenarios, without large-scale expansion into the CRB (Fig. 8; Supplementary Fig. S1). This is the contribution of introducing the future core driving factors of LULC that vary across scenarios, as the spatial distribution of driving factors directly influences the suitability of LULC type (Liang et al. 2018; Liu et al. 2022a). Human activity factor constrains expansion for built-up land, and future population and economic centers are mainly concentrated in the middle and lower Weihe River under different scenarios (Supplementary Figs S4 and S5).

Advantages of combining Markov and FLUS model

Markov–FLUS can accurately project future land use requirements and distribution patterns. Markov model enables alleviating the impact of abrupt fluctuations and policy effects on LULC requirements projections by setting different projection intervals. The improved Markov model allows to setting of different scenario weight matrices (W_n) to fully reflect the potential impact of different policy effects producing a reasonable projection of LULC requirements under different scenarios. The FLUS model is capable of integrating the dual impacts of climate change and human activities in LULC spatial projections by incorporating core driving factors. Therefore, the improved Markov–FLUS model realized LULC projections under different development scenarios, showing potential applicability for accurately assessing the impact of LULC dynamics on CS. The quality of future LULC simulations and projections can only be better realized through a progressive top–down investigation.

Response of CS to changes in LULC

Both CS and its dynamics pronounced clustering patterns across various scenarios. The most prominent hotspots are primarily concentrated within the CRB and at the confluence of the UWRB and MWRB, whereas the cold spot is located in the Guanzhong Plain. These patterns are intricately linked to the distribution of forested and built-up lands, predominantly driven by disparities in the CS capacity of different LULC types (Chuai et al. 2013; Hu et al. 2023). It is evident that transformations in LULC are intimately associated with CS and display spatial heterogeneity.

In the EP, lower economic and population growth rates and relatively suitable climatic conditions promote forestland conservation and expansion (Wang et al. 2022b). Decelerated transformation rates from landscapes with high capacity for CS, resulting in an increasing trend in CS (Liang et al. 2021; Yang et al. 2020). In HT and ED, the expansion of built-up land played a dominant role in the decline of CS. Furthermore, the transformation of the LULC type is frequent in the ED, with the characteristics of higher dynamic, more complex and disorderly. ED lacks the protection of forestland, with some massive deforestation (Liu et al. 2022b), while high-frequency forestland transformation is irreversible in terms of ecological persecution (Xu et al. 2023). Therefore, the CS is lowest and the expansion of the cold spot region is most evident in the ED (Fig. 12). Overall, future different policy effects will cause opposed changes in CS. Scenario simulations could help stakeholders optimally trade-off CS service processes by investigating the interconnections among them. These are valuable for the regional future ecological conservation and sustainable development.

Policy recommendations for CS

• (1) Optimizing LULC structures to enhance ecosystem CS. Intensive LULC development can reduce CS and increase ecological vulnerability (Wang et al. 2022a). Rational optimization of the LULC structure should be mentioned to increase regional CS, considering spatiotemporal characteristics, environmental changes and the impact of human activities on LULC transformation. Areas with lower precipitation and concentrated socioeconomic activities often lack sufficient CS capacity. Specially, both CRB and Guanzhong Plain have lower CS capacity, so efforts should be made to implement ecological restoration and increase urban green space to enhance regional CS capacity and ecological benefits.

• (2) Continuing to strengthen ecological protection policy pathways to reduce carbon emissions from social systems. The results of the study found that EP policy pathways have yielded better CS. It is advisable to appropriately slow down economic growth, reduce fossil fuel consumption and promote clean energy sources to mitigate the negative impact of climate change on the growth of vegetation such as forestland and grassland (Wang et al. 2022b). ED should be pursued while balancing the ecoenvironment and avoiding environmental destruction (Wu et al. 2022). Priority should be given to the protection of original ecosystems and to sustainably strengthening ecological conservation programs to expanse forestland coverage. These will help to rationally and scientifically optimize the LULC distribution pattern and enhance regional CS capacity.

Study limitations and future research

This study highlights the scientific value of the combined modeling framework. However, this study has some limitations. The time scale for equal-interval projections should be chosen rationally according to the extent to which the locality is affected by policies or human activities. Moreover, the typical scenarios and driving factors were only selected, which still leaves a strong uncertainty for the characterization of future LULC distribution patterns. Future studies should consider additional scenarios and more factors and focus on timely factorial information to comprehensively assess the impact of future LULC changes on CS. Additionally, carbon densities were estimated based on literature rather than empirical surveys and failed to inadequately account for the spatiotemporal heterogeneity, which hindered the precise quantification of CS. Overcoming these limitations is crucial to enhance the accuracy of CS assessments in future studies.

CONCLUSIONS

This study established a combined methodological framework for exploring the response of the terrestrial ecosystem CS dynamics to future LULC change under different policy pathways and climate change scenarios. It enables top–down integrated LULC evolution trend analysis, driving force identification, spatial pattern projection and CS dynamics assessment. The framework was applied to WRB and the following conclusions were obtained.

• (1) Cultivated land, grassland and forestland are the main LULC types that are susceptible to ecological conservation policies or projects. Identifying the driving force based on the transformation frequency not only preserves the core driving factors but also better explains the contribution of factors to the number of LULC transformations. Precipitation, POP, GDP and temperature are the core driving factors.

• (2) The equal-interval projection after abrupt LULC changes can mitigate the effects of LULC transformation rate fluctuations. Considering policy effects and future evolving core driving factors improves the reliability and accuracy of LULC projections.

• (3) Rapid ED has resulted in the decrease of CS. Future development should further strengthen the ecological protection policy and optimize the LULC structure according to local conditions to maintain a superior ecological environment. The framework proposed in this study can realize future LULC projections and CS assessments under different scenarios, which can provide valuable references for future LULC policy to achieve carbon neutrality targets.

Supplementary Material

Supplementary material is available at Journal of Plant Ecology online.

Figure S1: Prediction results of LULC distribution patterns in different scenarios without considering future core driving factors (Wu et al. 2022).

Figure S2: Spatial distribution of the gross domestic product (GDP) in the Weihe River Basin under future scenarios (yuan/km²).

Figure S3: Spatial distribution of population density (POP) in the Weihe River Basin under future scenarios (population/km²).

Figure S4: Spatial distribution of average temperature in the Weihe River Basin under future scenarios (°C).

Figure S5: Spatial distribution of average precipitation in the Weihe River Basin under future scenarios (mm).

Table S1: The neighborhood weights for each LULC type.

Table S2: The transformation cost matrix for different scenarios.

Table S3: CS for each level from 2025 to 2040 under different scenarios in the Weihe River Basin (10⁸ Mg).

Funding

This study was supported by the China Postdoctoral Science Foundation (2022M722561), the National Natural Science Foundation of China (51679186), the Water Science and Technology Program of Shaanxi (Program No. 2018slkj-4) and the Natural Science Basic Research Program of Shaanxi (Program No. 2021JLM-45).

Acknowledgements

The authors are grateful to the editors and the anonymous reviewers for their insightful comments and suggestions.

Conflict of interest statement. The authors declare that they have no conflict of interest.

REFERENCES