China, located at the intersection of East and Central Asia, spans approximately 18 to 53° north latitude and 73 to 135° east longitude. The country features a diverse topography, ranging from expansive plains and rolling hills in the east to towering plateaus and mountain ranges in the west, with a general west-to-east elevation gradient. China's climate is complex, exhibiting considerable spatial variability in both annual precipitation and temperature. The eastern monsoon region, characterized by abundant rainfall and a pronounced north–south temperature gradient, supports a wide variety of plant species. In contrast, the arid regions of western China experience minimal precipitation and sparse vegetation cover. This climatic diversity has given rise to a wide range of vegetation types, from moist evergreen broad-leaved forests to arid desert steppes. Additionally, there are significant regional differences in the seasonal and interannual fluctuations of these vegetation patterns.

Since the late 20th century, China has undertaken several large-scale ecological projects. These include the Three-North Shelterbelt Project (initiated in 1978), which aims to improve the ecological conditions of northern China's arid and semi-arid regions; the Grain for Green Program (launched in 1999), which focuses on converting farmland back to forest and grassland in ecologically fragile areas of western and northern China; and the Natural Forest Protection Program (started in 1998), which primarily targets the Yangtze River Basin and the upper reaches of the Yellow River. These initiatives have significantly influenced vegetation cover and ecosystem functions across the country.

For the purposes of this study, China's vegetation is classified into four major categories: forest, grassland, cropland, and non-vegetated areas (Fig. S1 in the Supplement).

DGVMs offer a wide range of capabilities, serving as essential tools for systematizing our understanding of the functionality and universality of vegetation and ecosystems. The development of DGVMs has enhanced our comprehension of the empirical relationships among functional plant traits and their interactions with the environment. These models also allow for predictions of the impacts of climate change on ecosystems and the resulting biogeochemical feedbacks. Specifically, DGVMs can simulate surface fluxes of CO2 and water, as well as the dynamics of water and carbon reservoirs and their responses to factors such as climate change, atmospheric composition, and land use changes.

In this study, we employed 14 DGVM models (CABLE-POP, CLASSIC, CLM5.0, DLEM, IBIS, ISAM, ISBA-CTRIP, JULES, LPJ-GUESS, LPX, OCN, ORCHIDEEv3, SDGVM, and VISIT), selected for their superior performance in simulating gridded monthly LAI and GPP, to explore vegetation dynamics (Zou et al., 2023). These models were coordinated by the Comparative Productivity in Net Biota project (Trends in the Land Carbon Cycle, TRENDY), which is part of a broader initiative. Within the TRENDY project, all DGVMs employ uniform spin-up durations. Initialization of simulations utilizes the 1700 baseline conditions with model-specific plant functional types (PFTs). Pre-industrial climate forcing data (1901–1920) are recycled through model-dependent cycling or stochastic sampling until a carbon equilibrium state is achieved (Sitch et al., 2024). Despite uniform TRENDY spin-up protocols, inherent structural differences among DGVMs (e.g., carbon cycling, vegetation dynamics) generate divergent initial states (LAI, GPP) that propagate into early transient simulations. Although such discrepancies may impact model–data comparisons in the first few decades, their effects are minimal during our study period (2003–2019). Concurrently, all participating DGVMs adopt LUH2, leveraging its CMIP-optimized architecture to standardize forcing and enhance DGVM compatibility as the unified land use change dataset. All models were subjected to scenarios S1, S2, and S3. Additionally, they were driven by historical climate fields and land use data from the Climate Research Unit's National Center for Environmental Prediction (CRUNCEP) dataset. In S1 simulations, the models were influenced solely by increasing atmospheric CO2 concentrations, with a constant climate based on cyclic climate averages and variability from the early 20th century. The S2 scenario varied both atmospheric CO2 concentrations and climate, while the S3 scenario included variations in CO2 concentrations, climate, and land use. For this study, the S2 scenario, which involved varying atmospheric CO2 concentrations and climate, was used to obtain simulated LAI and GPP. The difference between the S3 and S2 simulations (S3–S2) was utilized to evaluate the models' representation of the impacts of land use changes on LAI and GPP. In the TRENDY project, all DGVMs provide monthly outputs at a standardized spatial resolution of 0.5°×0.5°.

This study utilizes the reprocessed MODIS LAI in China from 2003 to 2019 (Yuan et al., 2011) to validate the simulated LAI of the 14 DGVMs. The reprocessing of the observed LAI data involves filling gaps in the original MODIS LAI (MOD15A2H and MYD15A2H) datasets, which are identified as low quality, using a modified spatiotemporal filtering (mTSF) technique. This is followed by a second round of processing using the TIMESAT SG filtering technique to generate the final products. For this study, the observed LAI data were resampled into monthly gridded data with a resolution of 0.5°×0.5°.

To validate the simulated GPP of DGVM, we used the CSIF dataset, which is processed by a neural network as a proxy for remotely sensed observations (Zhang et al., 2018). The CSIF dataset is derived from MODIS surface reflectance and OCO-2 SIF inversion data, trained using machine learning algorithms on daily SIF observations to produce a cloud-independent CSIF dataset. This dataset accurately captures the seasonal dynamics of satellite-observed SIF and shows a high degree of consistency with ecosystem GPP, making it a suitable proxy for GPP. In this study, the CSIF data were aggregated from a 4 d period to a monthly scale, and the spatial resolution was sampled to a 0.5°×0.5° grid using mean aggregation. Meanwhile, to enable a direct comparison of SIF and GPP values, both SIF and GPP data were normalized at both the annual and monthly scales.

FLUXNET, established with the aim of offering a high-quality, shared dataset for land surface model validation and development, is a global network of micrometeorological flux measurement sites. We used the FLUXNET dataset to analyze the seasonal response of LAI and GPP. In this study, we used monthly-scale GPP data from FLUXNET to validate the seasonal correlation of GPP at the site scale. Due to the limited duration of consistently observed GPP at the sites in China, they are not suitable for long-term trend validation of GPP but can be used for seasonal validation.

We utilize the monthly CO2 concentration data from the Mauna Loa Observatory in Hawaii (https://gml.noaa.gov/ccgg/trends/data.html, last access: 26 September 2025) as a representative regional indicator for CO2 concentrations within China.

To reduce uncertainty in analyzing LAI responses to environmental factors, we employed the same dataset that drives the DGVM. Specifically, historical temperature and precipitation data were sourced from the reanalyzed CRUNCEP dataset. We utilized monthly data from the atmospheric stress component of CRUNCEP (https://gdex.ucar.edu/datasets/d314003/, last access: 26 September 2025) and standardized the spatial resolution to 0.5°×0.5°.

Given that the CRUNCEP dataset does not include a variable for light radiation, this study utilizes monthly-scale net solar radiation data sourced from ERA5 (https://cds.climate.copernicus.eu/datasets/reanalysis-era5-land-monthly-means?tab=download, last access: 26 September 2025). The spatial resolution of this data is sampled at 0.5°×0.5°.

The IGBP classification scheme in MCD12Q1 of MODIS was used to categorize China into four types: forest (tree cover), grassland, cropland (crop), and non-vegetated areas (non-vegetable) (Fig. S1). The image resolution was initially resampled to 0.5°×0.5° using the majority rule. Evergreen coniferous forest, evergreen broadleaf forest, deciduous coniferous forest, deciduous broadleaf forest, mixed forest, woody savanna, and savanna were then combined into a single category termed forested land. Closed shrubland, open shrubland, grassland, and permanent wetland were collectively classified as grassland. Farmland and agricultural land (natural vegetation) were grouped together as cropland. Urban and built-up land, permanent snow and ice, and unutilized land were also categorized as non-vegetation.

We used the percent tree cover data from the Terra MODIS vegetation continuum field (VCF) product dataset, which provides subpixel-level estimates of global surface vegetation, as an indicator of actual tree cover to validate the accuracy of models for land use change in forested land. The spatial resolution of the tree cover data was aggregated to a 0.5°×0.5° annual scale through mean aggregation.

This study also utilized the statistical yearbooks of each Chinese province to assess the influence of human activities on vegetation in selected regions. For provinces predominantly characterized by grassland land use types (Tibet, Sichuan, Qinghai, and Inner Mongolia), the intensity of grazing was determined by counting the number of livestock enclosures reported in the respective province's statistical yearbooks. In contrast, for provinces primarily featuring cropland land use types (Henan, Heilongjiang, and Jiangsu), the extent of crop cultivation was assessed by calculating the area dedicated to farming practices as reported in the statistical yearbooks.

The Theil–Sen median method was employed to determine interannual trends in extensive time series, with significance testing conducted via the Mann–Kendall (MK) test. The primary metric for evaluating model performance focused on the interannual trends of LAI and GPP, as simulated by the DGVM model at both regional and pixel scales. To spatially evaluate the DGVM model's performance, a t test was executed using a running sample size of 9, based on a 3×3 sliding window of remotely sensed observation images.

To elucidate the mechanisms underlying the trends in annual LAI and GPP changes and to address the limitations of DGVMs, we analyzed the impacts on vegetation dynamics from two perspectives: human activities and environmental drivers. This study assessed the influence of human activities on vegetation dynamics by examining variations in grazing intensity and crop cultivation area, both sourced directly from statistical yearbooks. Grazing intensity was calculated by tallying the number of cattle and sheep confined and then multiplying this by an empirical estimate of pasture consumption per animal. To evaluate the response of annually observed and simulated LAI to environmental factors (atmospheric CO2 concentration, precipitation, temperature, and radiation) from 2003 to 2019, the multiple linear regression equations were primarily employed to determine the sensitivity of these observed and simulated LAI values to environmental variables: 1LAI=a×CO2+b×pre+c×tem+d×rad+ε, where LAI and CO2 are annual average LAI and carbon dioxide concentration, respectively; pre, tem, and rad are the annual average precipitation, temperature, and radiation, respectively; a, b, c, and d are regression coefficients; and ε is the residual error term, which is the amount of influence of anthropogenic activities on vegetation dynamics. Both the dependent and independent variables were normalized.

Pearson correlation coefficients were used to assess the seasonal performance of LAI and GPP for each model at both regional and pixel scales. Multi-year averages of observed and simulated LAI and GPP, calculated from January to December, were determined. Subsequently, their respective correlation coefficients (r) and significance (p) were independently calculated. A positive correlation coefficient with a p value less than 0.5 suggests that the model demonstrates strong seasonal cycle performance. Additionally, the site GPP data from FLUXNET were utilized to evaluate the seasonal performance of DGVMs at various sites within China.

Meanwhile, the cross-correlation function (CCF) was used to analyze the correlation between monthly-scale LAI and GPP data across all years, alongside climate factors – specifically precipitation, temperature, and radiation – that may influence the seasonal dynamics of vegetation. CCF is a statistical tool employed to quantify the correlation between two time series, which reflects the correlation of one series with another at varying time lags. In the context of climate data analysis, CCF can detect seasonal variations by identifying cyclical patterns and analyzing seasonal lags. Our objective is to determine the seasonal correlation between LAI and GPP in relation to these environmental factors and to identify any temporal delay in the vegetation's response to seasonal climate changes. Additionally, we aimed to quantify the extent to which these environmental factors impact vegetation growth and the subsequent lag effects they produce.

In general, the LAI and GPP of all models exhibit an upward trend, aligning with the direction of remotely sensed observation (Fig. 1). Regarding the spatial distribution of the observed LAI (Fig. S2 in the Supplement), most of the area experienced an increase in LAI from 2003 to 2019, with 60.31 % of the area showing a significant upward trend and a total of 88.11 % of the area exhibiting some level of increase. The regions with significant increases were primarily located in areas dominated by forested and cultivated land use types. The increasing trend of observed LAI is 0.012 m2m-2yr-1, which aligns with the trend simulated by the multi-model ensemble (MME) (0.0094 ± 0.0058 m2m-2yr-1), suggesting that the simulated LAI effectively captures the overall interannual trend of LAI. However, the normalized interannual trend of the observed CSIF is 0.16 yr-1, which exceeds the normalized GPP trend simulated by the MME (0.086 ± 0.021 yr-1), indicating that the models may underestimate the actual GPP interannual trend. The spatial distribution pattern of the observed CSIF data closely mirrors that of the observed LAI (Fig. S3 in the Supplement), with most areas displaying an increasing trend – 66.4 % of the area showing a significant increase and a total of 93.22 % of the area showing some increase.

Spatially, the comparison between simulated and observed LAI reveals that all models exhibit both overestimations and underestimations across most regions of China (Fig. 2). In the MME (Fig. 2a), only about one-third of the area demonstrates strong performance in simulating LAI trends. The normalized trend comparisons between simulated GPP and observed CSIF mirror the performance observed with LAI (Fig. 3). However, most models tend to underestimate GPP over a larger area than they do LAI. Although the DLEM model outperforms other models in simulating long-term GPP trends, its accuracy remains constrained below 50 % relative to observational benchmarks (Fig. 3e). Notably, the MME underestimates GPP in more than 50 % of the area (Fig. 3a).

Combining the results from Figs. 2 and 3, we observed that the regions where the models overestimate or underestimate LAI and GPP display similar overall distribution patterns. These patterns closely correspond with the spatial distribution of vegetation and land use types in China. To enhance the accuracy and relevance of our analysis, we classified the study area based on these land use types. As shown in Fig. S4 in the Supplement, the overall simulated LAI and GPP trends in forested land were generally underestimated. The LAI trend for MME in forested land was 0.015 ± 0.011 m2m-2yr-1, and the normalized trend of GPP was 0.13 ± 0.025 yr-1. These values are significantly lower than the trend value of observed LAI in forested land (0.027 m2m-2yr-1) and the normalized trend of observed CSIF (0.16). In contrast, the trends of observed LAI in grassland and cropland were 0.0047 and 0.086 m2m-2yr-1, respectively, which fall within the margin of error of the multi-model trend. The LAI trend values for the models in grassland and cropland were 0.0088 ± 0.0066 m2m-2yr-1 and 0.011 ± 0.0072 m2m-2yr-1, respectively. The normalized trends of observed CSIF in grassland and cropland were 0.086 and 0.13 yr-1, respectively. In grassland, the normalized MME's GPP was 0.091 ± 0.026 yr-1, which includes the trend of observed CSIF within its error margin. However, the normalized post-trend value of GPP for MME in cropland (0.093 ± 0.030 yr-1) was relatively low.

Upon calculating the percentage of area misestimated for forest land, grassland, and cropland, respectively (Fig. S5 in the Supplement), distinct patterns of misestimation became evident across different land use types. The models underestimated the LAI trend in forest land by up to 71.26 % of the area, with overestimation occurring in less than 10 % of the area. Conversely, for grassland, the models overestimated the LAI trend in 60.80 % of the area, with underestimation occurring in less than 10 %. In cropland, the models overestimated the LAI trend for 20.90 % of the area and underestimated it for 41.80 %. Overall, the models predominantly underestimated the LAI trend in forest land, displayed varying degrees of misestimation in cropland.

The variation in the impact of environmental factors on vegetation LAI can be understood as a combination of the trends in these factors and their sensitivity to LAI (Fig. 4). Among these factors, the sensitivity of LAI to CO2 concentrations – both observed and simulated data – was significantly higher compared to other environmental variables. The annual normalized trend increase in CO2 concentration was 0.20, markedly surpassing that of precipitation (0.035), temperature (0.048), and light radiation (-0.036) (Figs. 4 and S6–S8 in the Supplement). Thus, when considering both trends and sensitivities, CO2 concentration stands out as the most dominant environmental factor driving changes in both observed and most simulated LAI.

The study identified CO2 concentration as the most significant environmental factor influencing changes in both observed and simulated LAI. A spatial comparison was conducted to assess the impact of CO2 concentration on simulated LAI changes versus those observed in observed LAI (Fig. 5). The results indicated that in most models, the influence of CO2 concentration on LAI changes deviated significantly from actual observations. Even the best-performing model, CLM5, accurately accounted for less than 40 % of the area with accurately observed LAI changes, with large areas showing either overestimation or underestimation. These misestimated areas displayed a consistent pattern in their spatial distribution, mirroring the spatial trends observed in the model's overall LAI trend performance (Fig. 2), where significant overestimation occurred in grasslands. Regarding spatial differences in the sensitivity of observed and simulated LAI to precipitation, temperature, and radiation effects (Figs. S9–S11 in the Supplement), the models still exhibited substantial areas of underestimated and overestimated misestimation for these three environmental factors. However, the proportion of well-performing models increased compared to those influenced by CO2 concentration. Notably, the spatial distribution of simulated LAI changes in response to CO2 concentration closely resembled the spatial distribution of actual LAI changes (Figs. 2 and 5).

Site-scale seasonal performance was validated using data from five FLUXNET sites in China, comparing simulated GPP with observed CSIF (Fig. S12 in the Supplement). The results in Fig. S12 demonstrate that a significant portion of the simulated GPP and CSIF data exhibit strong seasonality across the different sites, closely aligning with the seasonal patterns observed in the GPP data from the FLUXNET sites. The seasonal correlations between the FLUXNET sites and the simulated GPP and CSIF exceeded the significance threshold.

The overall seasonal correlations between simulated LAI and GPP indicate that most models performed well in terms of seasonal accuracy in vegetated areas (Fig. 6). With a few exceptions, such as the DLEM model, most models demonstrated that seasonal correlations with observed LAI exceeded the threshold, indicating significant seasonal performance. Additionally, all models showed exceptionally high seasonal correlations between GPP and CSIF, surpassing the threshold. In summary, the models effectively capture the seasonal performance of both LAI and GPP, whether considered globally or across different land use types.

Spatially, most areas exhibited good to very good seasonal performance for LAI and GPP across most models (Figs. 7 and 8). Specifically, the percentage of areas with good seasonal performance for LAI in MME was 95.60 %, while for GPP it was 88.58 %. When analyzing land use type classification in China, regions exhibiting weak seasonal performance (i.e., p>0.05) in LAI and GPP were predominantly found in non-vegetated areas. Therefore, among vegetated regions, the seasonal accuracy of simulated LAI and GPP across China is remarkably high.

The CCF analysis revealed statistically significant correlations between observed and simulated LAI and key climatic variables – precipitation, temperature, and solar radiation (Figs. 9–11). Similarly, the CCF correlations with all simulated GPP are also very high, except for the DLEM. The correlation coefficients of these models with climate factors all surpass the threshold of significance. Specifically, the CCF seasonal correlations of observed and simulated LAI and GPP with precipitation and temperature slightly surpassed those with radiation. Additionally, the lagged effects of precipitation and temperature were not as pronounced as radiation on LAI and GPP growth promotion. Specifically, the CCF correlations of observed LAI with precipitation and temperature were 0.96 and 0.95, as the correlations of observed CSIF were 0.95 and 0.94, respectively. These values were marginally higher than the CCF correlations of observed LAI and CSIF with radiation, at 0.90 and 0.89. In contrast, the CCF correlations of simulated LAI for MME with precipitation, temperature, and radiation were 0.94, 0.97, and 0.94, while those of simulated GPP with the three climate factors were 0.95, 0.99, and 0.96, respectively. This implies that the simulated LAI and GPP may overestimate the seasonal response to climatic factors. However, in non-vegetated areas, the CCF correlations of observed CSIF with all three environmental factors were lower, with coefficients of 0.76, 0.70, and 0.65 for precipitation, radiation, and temperature, respectively. Nonetheless, the CCF correlations of the simulated GPP with the three climate factors remained very high for the MME, showing coefficients of 0.88, 0.94, and 0.95 for precipitation, radiation, and temperature, respectively. Although the high CCF intercorrelations between seasonal simulated LAI and climate factors are evident, the models generally captured the lagged effects of these factors on LAI for periods 1–2 months longer than observed in remotely sensed data, except for radiation. This time lag is more pronounced in the simulated LAI response compared to actual observations, particularly when compared to the simulated GPP response.

Increasing concentrations of CO2, climate change, nitrogen deposition, and land use collectively shape long-term vegetation greening trends (Piao et al., 2013; Li et al., 2024). Our study indicates a significant misestimation of trends in simulated LAI and GPP at both pixel and regional scales in China (Figs. 1–3). Overall, the observed LAI trends demonstrate reasonable consistency with model simulations within uncertainty bounds (Fig. 1a), indicating that current DGVM frameworks can effectively capture the overall trend of vegetation dynamics. However, the overall agreement contrasts with substantial spatial discrepancies in trend misestimation, as evidenced by pronounced spatial misestimations in China (Fig. 2). The primary reason for these inaccuracies in simulated LAI and GPP trends at pixel scale appears to be the limited ability of DGVMs to capture vegetation responses to elevated CO2 concentrations (Figs. 4 and 5). This aligns with previous findings that underscore the significant role of CO2 fertilization in global vegetation changes (Zhu et al., 2016; Piao et al., 2020). Differences between observed and simulated LAI trends reveal a spatial pattern similar to those seen in the disparities between observed and simulated LAI responses to CO2 concentration, indicating that misestimation of simulated LAI trends likely stems from the model's inadequate representation of LAI response to CO2 variations (Figs. 2 and 5). Besides, the model's representation of LAI response to climate change appears to have a minor effect on trend accuracy, which is less critical than the response to CO2 (Figs. 4 and 5, Figs. S9–S11). On the other hand, the spatial patterns of dominant environmental and anthropogenic drivers affecting observed LAI responses differ markedly from those affecting simulated responses (Fig. S13 in the Supplement). In particular, the areas where dominant drivers in the model diverge from actual drivers are prominent in southeastern and south-central forested regions, northern and western grasslands, and northern crop areas in China (Fig. S13a and b). For example, in south-central and southeastern forests, the model suggests that temperature is the primary driver of LAI trends, whereas CO2 and anthropogenic factors are the actual dominant influences. Conversely, in northern and western grasslands, the model emphasizes CO2 dominance, while precipitation is, in reality, the primary driver. These discrepancies partly explain the model's underestimation of LAI response to CO2 in southern forested areas and its overestimation in northern grasslands (Fig. 5). Parameterization and model structure are fundamental for ecosystem models to generate realistic projections, playing a critical role in their accuracy (Luo et al., 2016). On the one hand, due to the complexity of physiological processes and environmental interactions in ecosystems, model accuracy is highly sensitive to parameterization. However, DGVMs often simplify these interactions, using empirical functions to model leaf-level photosynthetic processes – such as CO2 partial pressure, carboxylation, and electron transport (Reich et al., 2014; Wang et al., 2017). This approach can result in modeled stomatal conductance responses to environmental factors that deviate from observed reality, particularly during the light reaction phase (Buckley, 2017; Mcdermid et al., 2021). Simulating stomatal conductance under varying conditions (temperature, moisture, vapor pressure, and light radiation) on a global scale is challenging, as models typically rely on a single temperature response curve, which fails to capture ecosystem variability (Powell et al., 2013; Trugman et al., 2018; Lin et al., 2015). Furthermore, the species-specific sensitivity of stomatal conductance to CO2 is often overlooked, limiting the model's ability to reflect diverse physiological responses (Buckley, 2017). DGVMs generally employ PFTs to simulate vegetation responses to environmental changes (Rogers et al., 2017), but PFTs simplify plant functional and adaptive traits, reducing the accuracy of physiological responses under climate change (Rogers et al., 2014). Given substantial interspecies differences in photosynthetic and water-use efficiencies, a single PFT approach is insufficient to represent the varied responses of diverse plant types to environmental shifts (Rezende et al., 2016; Rogers et al., 2014).

On the other hand, model behavior is also shaped by equations representing real-world processes, which affect the model's capacity to simulate system functions accurately (Luo et al., 2016). Studies show that current models fail to capture nitrogen and phosphorus limitations on CO2 fertilization effects (Terrer et al., 2019). Nutrient limitations, particularly of nitrogen and phosphorus, can lead to reduced Rubisco – an essential, nitrogen-rich enzyme in photosynthesis – which, in turn, downregulates photosynthetic capacity and decreases CO2 assimilation by vegetation (Ainsworth and Rogers, 2007; Long et al., 2004; Terrer et al., 2016, 2018). DGVM assumptions of C–N and C–P coupling are thus inadequate. While plants employ various symbiotic strategies to mitigate nitrogen limitation, DGVMs often overlook these mechanisms (Sulman et al., 2019). Additionally, evidence from phosphorus-limited ecosystems suggests that elevated CO2 may aid plants in acquiring or mineralizing phosphorus (Nazeri et al., 2013). However, models generally assume a relatively closed phosphorus cycle, and the fundamental dynamics of vegetation under phosphorus limitation remain poorly understood, introducing uncertainty in C–P coupling (Medlyn et al., 2016).

While linear models are useful for analyzing the contribution of environmental factors to long-term vegetation trends, determining the impact of human activities on vegetation dynamics remains a subject of debate in previous studies (Ge et al., 2021; Liu et al., 2022). This uncertainty arises from the complexity of human activities and the diverse conditions across different study areas (Liang et al., 2023). In China, specific human activities, such as enhancing the productivity of cropland and implementing reforestation policies, have been shown to significantly enhance vegetation growth (Zhang et al., 2000). Conversely, urbanization and other similar human activities have been found to have a considerable negative effect on vegetation growth (Wang et al., 2022b; Lin et al., 2019; Jiang et al., 2022). To quantify the effects of land use changes driven by human activities, two sets of scenarios were applied in DGVMs, where the S3 scenario applies historical changes in land use, and the other S2 scenario applies a time-invariant pre-industrial land cover distribution. By difference of the two scenarios, the dynamic evolution of vegetation in response to land use change can be quantified (Friedlingstein et al., 2022b).

In this study, we analyzed the trends of grazing intensity and crop acreage in provinces predominantly characterized by grassland and cropland. This analysis was conducted by examining the number of livestock and crop acreage from the respective provinces' statistical yearbooks. Additionally, tree cover data are utilized to represent actual regional trends in forested land, allowing us to assess the model's ability to capture the effects of human activities on LAI. The S3–S2 scenario was employed to evaluate the influence of land use change on vegetation LAI and GPP as perceived by the DGVM model (Fig. S14 in the Supplement). Overall, the model did not show a significant effect of land use change on LAI in forested land and grassland areas within the MME. Spatially, most simulated LAI and GPP trends exhibited similar patterns under the S3–S2 scenario (Figs. S15 and S16 in the Supplement), with most of the area showing insignificant changes (72.47 % and 87.31 % of the area showed insignificant changes in LAI and GPP, respectively). In terms of tree cover change data (Fig. S17 in the Supplement), there was a significant increase in tree cover in the southern and northeastern regions of China, with 80.60 % of the area experiencing growth and 31.54 % showing significant growth. Tree cover increased across all land use types in China, with the highest annual growth rate (0.417) occurring in forested land (Fig. S17). Data on grazing intensity and cropped area were obtained from the statistical yearbook for grassland and cropland, respectively, and trends in grazing intensity or cropped area were compared with trends in normalized LAI from the model's S3–S2 scenario (Fig. S18 in the Supplement). The results indicated that the model's representation of LAI changes due to land use change was inadequate in several provinces dominated by cropland and grassland. For instance, the model's normalized LAI trend does not accurately reflect the actual anthropogenic changes on vegetation LAI in the Inner Mongolia (IM) province, where the changes were insignificant. In Qinghai (QH), Tibet (TI), Sichuan (SC), Heilongjiang (HLJ), and Hunan (HN), the direction of the LAI trend in the model's S3–S2 scenario contradicts the direction of anthropogenic effects on vegetation LAI. Consequently, the model fails to adequately capture the intensity or direction of anthropogenic impacts on vegetation LAI.

Although DGVMs can dynamically adjust to imposed land use change, there are substantial limitations for models to respond to land use changes, which struggle to ensure the quality of input data and incorporate land management practices and accurately simulate ecosystem responses to change. In the TRENDY project, most of DGVMs use the HYDE and LUH2 databases for LULCC data input variables (Klein Goldewijk et al., 2017; Hurtt et al., 2020; Sitch et al., 2024). However, absolute agreement with all observations remains unattainable for any global product. While these LULCC datasets are common across models, their implementation varies according to the land use processes and functionalities inherent to individual DGVMs. In this study, we employ Table 1 to describe differences in the performance of individual DGVMs, which were adapted from Sitch et al. (2024). Meanwhile, the historical land use estimates within the datasets exhibit regional variations and limitations of spatial resolution accuracy, which may result in inaccurate simulations of responses to land use changes due to discrepancies in the data (Prestele et al., 2016). The globally applicable gridded dataset necessitates simplifications of patterns and historical trends in deforestation, cropland, and pasture management, yet these simplifications inherently introduce uncertainties (Prestele et al., 2016; Arneth et al., 2017). These limitations may contribute to the model's inadequate representation of anthropogenic factors affecting regional vegetation LAI and GPP trends in China, as the models struggle to accurately capture the impacts of human activities on forested land, grassland, and cropland.

Our analysis highlights several critical limitations in the current DGVMs regarding their ability to accurately represent vegetation responses to environmental drivers and anthropogenic changes. These limitations arise primarily from differences in how individual DGVMs parameterize and simulate critical ecological and physiological processes. For example, Teckentrup et al. (2021) demonstrated significant divergences among DGVMs, particularly in their approaches to modeling responses to elevated atmospheric CO2 concentrations and nutrient cycle. Among these physiological processes, plant carbon assimilation mechanisms such as stomatal conductance and maximum carboxylation velocity (Vcmax) play pivotal roles. Previous studies found that stomatal functioning and Vcmax are related to elevated atmospheric CO2 and photosynthesis acclimation, potentially impacting the estimation of vegetation dynamics in DGVMs (Rezende et al., 2016). Many DGVMs currently simplify or inadequately represent these physiological processes. For instance, several DGVMs estimate Vcmax at the canopy scale rather than the physiologically more appropriate leaf scale, leading to systematic underestimations. Furthermore, nutrient cycling modules integrated into DGVMs often reduce the sensitivity of vegetation growth to elevated CO2 concentrations (Smith et al., 2014; Zaehle, 2013; Meyerholt et al., 2020). Collectively, these simplifications constitute structural deficiencies that substantially affect DGVM accuracy. To elucidate mechanistic discrepancies in vegetation trend simulations, we synthesize findings from prior studies to summarize the operational frameworks of 14 models across six critical parameters: the presence of stomatal conductance, Vcmax-related leaf nitrogen content, dynamic natural PFT coverage, and nutrient cycle (Teckentrup et al., 2021; Rezende et al., 2016; Lian et al., 2021; Friedlingstein et al., 2022a; Sitch et al., 2024). A systematic comparison of these parameterizations is presented in Table 2.

Our results clearly indicate that no single DGVM among the evaluated 14 consistently performs well across all aspects of vegetation dynamics (e.g., trend vs. seasonal variations, LAI vs. GPP) throughout China (Figs. 2 and 3, Figs. 7 and 8). For example, while DLEM performs relatively better in capturing regional GPP trends in certain locations, its successful performance is spatially limited (covering less than 50 % of the study area) and remains poor in capturing LAI trends (Figs. 2e and 3e). Similar discrepancies exist for other models regarding their ability to simulate different vegetation variables. Thus, we emphasize the effectiveness and importance of utilizing MME approaches, which collectively capture model strengths and weaknesses, thereby providing more robust and comprehensive assessments.

In contrast to the vegetation trend, the models simulate the seasonal cycle of LAI and GPP excellently both overall and spatially (Figs. 6–8). The simulated GPP also demonstrates high seasonal cycle correlations with observed GPP in sites (Fig. S12). Seasonal variations in precipitation, radiation, and temperature play a crucial role in influencing the seasonal dynamics of vegetation (Hou et al., 2022; Duan et al., 2022; Wu et al., 2015). While temperature is often considered the primary factor controlling plant phenology, the complex interactions between temperature and other climatic factors cannot be overlooked (Flynn and Wolkovich, 2018). Radiation works in concert with temperature to regulate the timing of defoliation, meet vegetation growth thresholds, and influence phenological events (Flynn and Wolkovich, 2018; Chuine, 2010), while precipitation affects plant phenology primarily through its indirect influence on the thermal requirements of these events (Fu et al., 2014). The findings that observed LAI and CSIF exhibit similar seasonal fluctuations in response to the three key climate factors suggest that the model may effectively capture the response to seasonality of vegetation in China (Figs. S19 and S20 in the Supplement). In this paper, we verified the seasonal cyclic correlation of simulated LAI and GPP with climate factors through CCF (Figs. 9–11) and demonstrated that the simulated LAI and GPP are equipped with realistic seasonal cycles due to the simulated vegetation's ability to respond to cyclic climatic variations affecting its seasonality. Given that the Farquhar model and its variants, which underpin the photosynthesis module of the DGVM, provide a detailed and unified understanding of photosynthesis (Farquhar et al., 1980), DGVMs are capable of reasonably simulating dynamics of photosynthesis and vegetation change (Zhao et al., 2022). To further improve simulation accuracy, researchers have developed more mechanistic models that account for vegetation growth influenced by seasonal climate variations (Piao et al., 2019).

Moreover, we observed that the simulated LAI response to seasonal climate factors typically lags behind the simulated GPP response, which aligns with the analysis by Zhao et al. (2022). This discrepancy arises because the model simulates carbon allocation to canopy structure in response to seasonal climate changes more slowly than what occurs in reality (Du et al., 2020). As with the above research, studies on peak seasonal values and phenological greening dynamics hold significant research importance, and future work will strengthen research efforts in this thematic area.

Remotely sensed observations used in the study might be compromised by various factors such as atmospheric disturbances and sensor errors, leading to noise, blurring, and distortion. Even with the use of the latest reprocessed MODIS LAI data and advanced spatiotemporal filtering techniques, it is impossible to completely eliminate the potential impacts of cloud cover, aerosol pollution, and sensor degradation (Yuan et al., 2011; Zou et al., 2023). These limitations mean that satellite-based estimates may inherently underpredict values, requiring careful consideration when comparing model data to ensure appropriate metrics and accurate site-based comparisons for model evaluation (De Kauwe et al., 2016). The quality of the MODIS LAI data used in this study is particularly low in non-vegetated areas but significantly better in vegetated regions (Fig. S21 in the Supplement). Additionally, in northern China, pixel quality is severely affected from December to February due to snow accumulation and other factors. Consequently, the pixel quality of the remotely sensed observations themselves represents the primary source of uncertainty in the experiments conducted in this study.