Improved GPP modeling reveals rising precipitation influence in typical grasslands of Northern China’s Hulun Lake Basin

Abstract

The Hulun Lake Basin in northern China harbors extensive temperate grasslands, whose carbon uptake and ecosystem functioning are effectively represented by gross primary productivity (GPP). However, the long-term response of grassland GPP to regional climate change remains insufficiently understood, due to the limited ability of global models to capture local ecosystem variability. This study improved GPP estimation in the Hulun Lake Basin by integrating additional grassland flux observations from China into a Random Forest model (M_CF). We subsequently examined the spatiotemporal dynamics of GPP and identified the dominant climatic drivers and their long-term trends under changing climate conditions. Results show that the M_CF model significantly outperformed the model trained solely on FLUXNET 2015 Tier 2 data in both accuracy and trend patterns of grassland GPP. A 23-year M_CF simulation in the Hulun Lake Basin showed a clear west–east GPP gradient, averaging 676.36 g C m^− 2 yr^− 1 and totaling 70.13 Tg C yr^− 1. Most of the region exhibited increasing GPP trends, with an average growth rate of 6.85 g C m^− 2 yr^− 1, resulting in an annual total increase of approximately 0.45 Tg C yr^− 1. SHAP analysis demonstrated that shifts in precipitation and temperature are reshaping the dominant factors contributing to GPP increases. Structural equation modeling further revealed that precipitation enhanced GPP through both direct and indirect pathways, with the direct effect accounting for 63.6% of the total effect, while the remaining 36.4% was mediated primarily through NIRv. In particular, over large areas of the central-eastern basin, increasing precipitation has gradually replaced shortwave radiation as the primary driver of GPP enhancement. This study provides new insights into how temperate grassland GPP responds to climate change, offering a scientific basis for region-specific ecosystem management and policy decisions.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s13021-026-00491-9.

Introduction

As a vital component of global terrestrial ecosystems, temperate grasslands play a critical role in maintaining regional and even global ecological security through essential ecosystem services such as carbon sequestration [23]. Gross Primary Productivity (GPP), a key metric for ecosystem carbon budgets, reflects the functional performance of grasslands [47, 69]. Accelerated climate change, marked by warming and shifts in precipitation [30], is disrupting the relationships between grassland structure and function [29]. Thus, deciphering the spatiotemporal dynamics of climate-driven GPP and their response mechanisms becomes imperative to project carbon budgets and optimize adaptive management.

The estimation of GPP primarily relies on process-based models and remote sensing-driven light use efficiency models [9, 67]. Recent advances in remote sensing and machine learning (ML) have expanded the capacity for spatially explicit terrestrial carbon quantification by integrating satellite observations with environmental information across heterogeneous landscapes [14]. For GPP estimation, ML approaches can integrate multi-source data and capture nonlinear interactions among meteorological variables [6, 68]. Furthermore, novel vegetation indices such as near-infrared reflectance of vegetation (NIRv) and kernel NDVI (kNDVI), which better correlate with photosynthetic activity [3, 6], have been increasingly adopted as critical inputs to ML frameworks for improving the model performance.

However, the performance of large-scale productivity models depends on how well regional ecological and climatic constraints are represented. Current ML models are trained predominantly on global flux datasets, such as FLUXNET2015, to achieve cross-ecosystem generality [44]. Such training may not fully represent the structural complexity and region-specific ecological constraints of grassland ecosystems. McCallum et al. [34] noted that globally applied satellite-based production efficiency models may likewise overlook region-specific physiological and meteorological conditions. They highlighted the value of regional approaches within broader monitoring frameworks. In China, grasslands are extensive and regulated by a distinctive monsoon climate and complex topographic gradients [28, 49], leading to pronounced regional variability in the spatiotemporal dynamics of GPP. Consequently, the transferability of globally trained ML models to the temperate grasslands of northern China remains uncertain.

China’s grassland GPP has increased significantly over the past two decades, with the strongest gains occurring in northern grasslands [21, 58]. The Hulun Lake Basin lies within the Hulunbuir grassland region of Inner Mongolia and contains a representative mosaic of temperate grasslands and wetlands. Beyond this representativeness, the basin forms part of a critical ecological security barrier in northern China [20]. Its ecosystems support biodiversity, regulate climate, conserve water, and limit desertification [17]. The basin also spans pronounced hydroclimatic and productivity gradients, making it a valuable setting for examining how changing water availability reshapes grassland carbon uptake. Under ongoing climate change, interactions between shifting hydroclimatic conditions and local ecological constraints may contribute to divergent productivity responses within the basin. Several parts of the region have recently shown annual GPP increases exceeding 10 g C m^− 2 yr^− 1 [31]. However, strong spatial heterogeneity persists, and some degraded grasslands remain characterized by low productivity [59].

Climate change is a major driver of GPP in grassland ecosystems, which are highly sensitive to climatic factors such as temperature, precipitation, and solar radiation [48]. Global studies show that in arid and semi-arid regions, GPP variability is mainly controlled by water availability, while in colder grasslands, temperature is the dominant factor affecting photosynthesis [53, 55]. These climatic changes affect GPP both directly, by influencing photosynthesis, and indirectly, by altering soil moisture and vegetation phenology. Consequently, GPP responses to climate change are complex and exhibit marked spatial heterogeneity across regions.

Northern China has experienced marked warming and increasing precipitation in recent decades, particularly across temperate grasslands [32, 70]. These climatic shifts may alter not only GPP but also the relative importance of its climatic controls. However, whether these controls have changed over time, and how vegetation activity mediates climate–GPP relationships, remain unclear.

To address these gaps, we developed a machine learning-based GPP model adapted for temperate grasslands in China by incorporating regional flux observations. We applied this model to reconstruct GPP across the Hulun Lake Basin from 2001 to 2023 and examine its spatiotemporal dynamics and changing climatic drivers. We further investigated the direct and vegetation-mediated relationships between climate and GPP. By connecting regionally improved GPP estimation with long-term shifts in climatic regulation, this study advances understanding of carbon uptake in cold, semiarid temperate grasslands. It also provides a scientific basis for adaptive ecosystem management in northern China.

Materials and methods

Study area

The Hulun Lake Basin (47°00′–50°15′ N, 115°30′–122°28′ E) lies in northeastern Inner Mongolia on the western slope of the Greater Khingan Mountains (Fig. 1). It borders Mongolia to the west and south and Russia to the northwest, spanning ~ 94,000 km². The terrain descends from eastern mountains to central and western plains. Temperate grassland dominates the region’s ecosystems, covering ~ 68% of the land and serving as a key ecological buffer in northern China. The region experiences a temperate continental climate, with long, cold winters and short, warm summers. Annual temperatures range from − 1.5 °C to 1.5 °C, and precipitation ranges from 150 to 400 mm.

Fig. 1.

Overview of the study area. (a) Location of Inner Mongolia within china, (b) Location of Hulun Lake Basin, and (c) Digital elevation model (DEM) map of Hulun Lake Basin. (Administrative division data is sourced from the National Geomatics Center of China; DEM data is obtained from the Terra Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Version 3.)

Data and preprocessing

Table 1 presents all the datasets used in this study. Detailed descriptions and preprocessing steps for each dataset are provided below.

Table 1.

List of data properties and sources used in the paper

Data	Data Source	Spatial resolution	Temporal resolution	Remark	Citation
VI	MODIS MOD09GA	500 m	Daily	QA mask	Şenel et al. [43]
GPP from eddy covariance	FLUXNET 2015 Tier 2 dataset	Site	Monthly	GPP_DT_VUT_75	Pastorello [37]
	China FLUX	Site	Monthly	-	Yu et al. [61]
Meteorological data	TerraClimate	4638.3 m	Monthly	-	Abatzoglou et al. [1]
GPP products	PML-v2.2a	500 m	8 Days	-	Zhang et al. [73]
	GLASS-GPP	500 m	Yearly	-	Yuan et al. [63]; Zheng et al. [75]
	MODIS MOD17A2HGF	500 m	8 Days	-	Running and Zhao [40]

Eddy covariance data

The FLUXNET 2015 Tier 2 dataset (https://fluxnet.org/), derived via the eddy covariance technique, encompasses data from 212 regional flux network sites worldwide, spanning various terrestrial ecosystem types such as grasslands, deciduous broadleaf forests, and evergreen broadleaf forests. This comprehensive dataset offers meteorological variables across multiple temporal scales—hourly, daily, and monthly—alongside critical ecosystem carbon flux measurements such as GPP and net ecosystem exchange (NEE) [37]. For this study, we selected grassland sites featuring long-term observations (more than 8 years) from this dataset as shown in Table 2. And its monthly GPP product (GPP_DT_VUT_75) was utilized.

Table 2.

List of GRA sites used in this study

Purpose	Dataset	Site	Long, Lat	Site year	Remark
Model training	FLUXNET1	AT_Neu	11.3175, 47.1167	2002–2012	Site-month samples were randomly split into 80% for training and 20% for testing
		DE-Gri	13.5126, 50.9500	2004–2014
		GL-ZaH	-20.5503, 74.4733	2000–2014
		IT-MBo	11.0458, 46.0147	2003–2013
		RU-Sam	126.4958, 72.3738	2002–2014
		US-Var	-120.9508, 38.4133	2000–2014
		CN-Cng	123.5092, 44.5934	2007–2010
	ChinaFLUX	Damxung	91.066, 30.497	2004–2008
		Xilingol	116.40, 43.325	2004–2006
Independent validation in China sites	ChinaFLUX	Damxung	91.066, 30.497	2009–2010	-
		Xilingol	116.40, 43.325	2007–2008
Independent validation across global sites	FLUXNET2	AU-Dap	131.3181, -14.0633	2007–2013	Time spans are less than 10 years; therefore, these sites were used to validate model generalization across more sites.
		CH-Cha	8.4104, 47.2102	2005–2014
		CH-Fru	8.5378, 47.1158	2005–2014
		CH-Oe1	7.7319, 47.2858	2002–2008
		CZ-BK2	18.5429, 49.4944	2006–2012
		IT-Tor	7.5781, 45.8444	2008–2014
		NL-Hor	5.0713, 52.2403	2005–2011
		US-IB2	-88.241, 41.8406	2004–2011
		US-SRG	-110.8277, 31.7894	2008–2014

In the FLUXNET dataset, both the duration of observations and the number of sites covering Chinese grasslands are relatively limited. To address this, we incorporated grassland ecosystem observation data from ChinaFLUX (https://chinaflux.org), a long-term ecological monitoring platform established under the China Ecosystem Research Network (CERN) [61]. We selected data from two northern China grassland sites—Damxung and Xilingol—which have relatively long observation periods of 7 and 5 years respectively, and high-quality datasets [11]. In this study, GPP was calculated based on NEE and Re provided by the stations, using Eq. (1).

GPP=Re - NEE 1

Detailed information for all grassland flux sites used in this study is summarized in Table 2. For clarity, the FLUXNET sites were divided into two groups according to observation length, regional representativeness, and study purpose. FLUXNET¹ comprised sites with relatively long observation records (generally > 10 years), together with the Chinese FLUXNET site (CN-Cng), and was used for model training along with part of the ChinaFLUX observations. Because of the limited availability of long-term Chinese grassland flux data, the ChinaFLUX records were further divided by year: some years were used for model training, whereas the remaining years were reserved for independent validation to assess model performance for GPP estimation in China. The remaining FLUXNET sites with shorter observation periods were grouped as FLUXNET² and were not involved in model training; instead, they were used for independent validation to evaluate the model’s transferability across global grassland sites (Fig. 2).

Fig. 2.

Spatial distribution of grassland flux sites used in this study. Details and definitions of FLUXNET¹ and FLUXNET² are provided in Table 2

NIRv

This study utilized the MODIS surface reflectance product MOD09GA.061 obtained via the Google Earth Engine (GEE) platform to derive NIRv. The product is based on multispectral observations from the Moderate Resolution Imaging Spectroradiometer (MODIS) instrument aboard the Terra satellite and undergoes rigorous atmospheric correction to provide daily surface reflectance data. MOD09GA.061 features spatial resolutions of 500 m and 1 km and is accompanied by quality assurance (QA) data designed to filter out contaminants such as clouds and shadows effectively. Its strengths in temporal continuity, extensive spatial coverage, and precise radiometric calibration have established it as a widely adopted dataset worldwide for extracting vegetation indices and monitoring ecosystem dynamics [43].

The formula used to calculate NIRv is presented below:

2

In Eq. (2), NIR and RED denote the reflectance values of the near-infrared and red bands, respectively, as derived from the MOD09GA.061 product with 500 m spatial resolution.

The monthly mean NIRv values were extracted by averaging pixels within a 500 m buffer zone around each flux tower site, corresponding to the observation months. Additionally, monthly average NIRv values covering the entire study region from 2001 to 2023 were extracted to support regional-scale GPP estimation.

Meteorological data

Meteorological variables utilized in this study were obtained from the TerraClimate dataset (https://www.climatologylab.org/). This dataset integrates WorldClim’s high-resolution global climate data with the Climate Research Unit’s TS4.0 (CRU Ts4.0) and the 55-year Japanese atmospheric reanalysis (JRA-55), employing a climate-aware interpolation technique to generate a high-resolution monthly dataset (~ 4.6 km) well-suited for climate change studies [1]. It offers extensive temporal coverage and a wide range of meteorological variables characterized by robust spatiotemporal consistency and continuity. We accessed TerraClimate data via the GEE platform to extract meteorological variables corresponding to flux tower sites and the study area spanning 2001 to 2023. For regional-scale modeling, the TerraClimate data were resampled to match the spatial resolution of MODIS products.

Methods

Machine learning model development

In this study, a GPP estimation model was constructed using the random forest (RF) algorithm. As a decision tree ensemble learning technique, RF exhibits powerful nonlinear modeling capabilities and strong resilience to overfitting. It effectively manages heterogeneous multi-dimensional data and enhances predictive stability and accuracy through the integration of numerous decision trees, leading to its widespread use in GPP estimation [8, 18].

In selecting input features, this study considered that ecosystem GPP is influenced by multiple interacting factors [10, 62]. Consequently, we incorporated NIRv alongside several key meteorological variables as inputs to the model. Vegetation indices, which are critical indicators of plant growth status, play a pivotal role in estimating GPP. Among these, NIRv more accurately captures vegetation growth and photosynthetic activity and demonstrates a strong correlation with grassland GPP [65]. The meteorological variables include shortwave radiation (Srad), precipitation (Pre), temperature (Temp), and vapor pressure deficit (VPD). Srad supplies the fundamental energy required for photosynthesis; Pre controls soil moisture availability and serves as a primary limiting factor for GPP in water-limited grassland ecosystems [38, 56]; VPD reflects atmospheric dryness, directly influencing stomatal behavior and transpiration intensity, thereby modulating carbon assimilation processes. These meteorological datasets were obtained from TerraClimate.

Based on the aforementioned feature datasets, an initial GPP estimation model (M_F) was developed using grassland sites from the FLUXNET network. To enhance the model’s regional applicability, ChinaFLUX observations were subsequently integrated, resulting in an improved version (M_CF) with better representation of Chinese grassland conditions. This approach was used to evaluate the extent to which incorporating Chinese site observations enhances the accuracy of GPP simulations in the study region.

For model training, 80% of the dataset from FLUXNET¹ and ChinaFLUX was allocated randomly for training and the remaining 20% for testing to ensure robust predictive performance. In addition, observations from the Damxung site (2009–2010) and the Xilingol site (2007–2008) were excluded from model training and reserved for independent validation. This allowed for a direct comparison of the performance between the M_CF and M_F models on these specific sites, providing a robust assessment of how incorporating Chinese site data influences the accuracy of GPP estimation in China’ grasslands.

To further evaluate the generalization ability of M_CF and benchmark its performance against existing GPP products, we conducted an additional external validation using nine independent GRA FLUXNET sites that were completely excluded in model training. For each site, monthly NIRv and TerraClimate variables were extracted following the same preprocessing workflow used for the training data, and monthly GPP was predicted using the trained M_CF model. Annual cumulative GPP from M_CF, PML-v2.2a, GLASS-GPP, and MODIS-GPP was extracted for each independent flux tower site and compared with site-observed annual GPP. The model’s performance was assessed using two key metrics: the coefficient of determination (R²) and root mean square error (RMSE). R² quantifies how well the predicted values fit the observed data, whereas RMSE indicates the absolute size of the prediction errors.

Regional-scale estimation of GPP

The M_CF model was applied to estimate grassland GPP in the Hulun Lake Basin. Input features consistent with those used during model training—including 500 m resolution NIRv and meteorological variables such as Srad, Pre, Temp, and VPD—were used to generate monthly GPP estimates.

Analysis of regional GPP trends

To comprehensively analyze the spatiotemporal variation trends of GPP in the study region from 2001 to 2023, this research applied the Theil-Sen median trend estimator alongside the non-parametric Mann-Kendall (MK) test to quantify trend magnitude and assess statistical significance [15, 16]. The Sen’s Slope estimator provides a robust estimate of the rate of change in a time series, where the slope β denotes the average annual rate of change; β > 0 signifies an upward trend, while β < 0 represents a downward trend. The formula for its calculation is given below:

3

In the Eq. 3, x_i and x_j denote the specific observed values within the time series.

The MK trend test is a non-parametric approach that does not assume normality of data and is robust against missing values and outliers. It is commonly used in conjunction with Sen’s Slope to assess the statistical significance of trends in long-term time series data. The computational procedures and classification criteria for trends are detailed in Eqs. (4–6) and Table 3.

4

5

6

Table 3.

Mann-Kendall test trend categories

β	Z	Trend features
β > 0	2.58 < Z	Extremely significant increase (ESI)
	1.96 < Z ≤ 2.58	Significant increase (SI)
	1.65 < Z ≤ 1.96	Slightly significant increase (SSI)
	Z ≤ 1.65	Non-significant (NS)
β = 0
β < 0
	1.65 < Z	Significant decrease (SD)

In the equations, S represents the test statistic, while x_i and x_j denote specific GPP values within the time series. The sign function sgn() determines whether the difference between two GPP values is positive, zero, or negative. Z_mk is the standardized statistic used in the MK test to assess trend significance. The significance level is set at α = 0.05, corresponding to a two-tailed critical value of Z_1−α/2 = ± 1.96. The trend is considered statistically significant when the absolute value of Z exceeds 1.65, 1.96, and 2.58, corresponding to 90%, 95%, and 99% confidence levels, respectively. To enhance visual clarity and highlight the primary regional dynamics, the detailed trend classifications were aggregated. Specifically, all non-significant changes (increase, decrease, and no change) were combined into a single “Non-significant (NS)” category, and all levels of significant decreases were merged into “Significant decrease (SD)”.

Regional-scale input variables were applied to the model on a pixel-by-pixel basis to produce spatial distribution and temporal trend maps of GPP in the study area spanning 2001 to 2023.

Analysis of driving factors of GPP

To analyze the key drivers and interactions influencing GPP, the RF-SHAP method was applied to assess the relative importance of input variables. SHAP analysis maintains the original model structure while quantifying the marginal contribution of each input feature to individual prediction outcomes, offering both global explanatory power and localized temporal insight. In recent years, SHAP has gained widespread use in the ecological remote sensing community [7, 33, 41, 60]. Accordingly, we applied SHAP to identify the key drivers controlling GPP dynamics in the study area.

To elucidate the spatial heterogeneity and temporal dynamics of driving factor importance at the regional scale, this study conducted pixel-level SHAP analysis by calculating SHAP values for each driver variable at each pixel. In combination with the MK trend test, we further examined the temporal variation and trend of each variable’s contribution to GPP.

To investigate the spatiotemporal shifts in dominant drivers of GPP over the past 23 years, two representative time intervals—2001–2005 and 2021–2023—were selected. For each period, the multiple-year SHAP values and spatial distribution patterns of key meteorological variables were computed. Comparative analyses were then conducted to highlight changes in the primary drivers of GPP between the early and recent stages of the study period.

Partial least squares structural equation modeling (PLS-SEM) was further applied to disentangle the direct and indirect effects of climatic variables on GPP. This method is widely used for exploring complex pathway relationships and identifying potential mediating mechanisms among multiple interacting variables [57]. In this framework, Srad, Pre, Temp, and VPD were specified as exogenous variables, NIRv was treated as a mediator, and GPP was defined as the endogenous response variable. This model structure allowed climatic variables to influence GPP both directly and indirectly through NIRv. Model performance was evaluated using the coefficient of determination (R²), Stone–Geisser’s Q² for predictive relevance, and the goodness-of-fit (GOF) index [12, 64].

Results

Model performance of simulated GPP in China’s Grasslands

Trained using a combination of FLUXNET2015 and ChinaFLUX datasets, M_CF exhibited strong performance both on training and testing sets. As illustrated in Fig. 3, it achieved a R² of 0.95 and a RMSE of 0.84 g C m^− 2 d^− 1.

Fig. 3.

The performance of M_CF on training and test sets

To rigorously evaluate the model’s extrapolation capability, independent validation was performed using observational data from two Chinese grassland sites: the Xilingol site for the period 2007–2008 (Fig. 4) and the Damxung site for the period 2009–2010 (Fig. S1). Crucially, data from this period were strictly held out and completely excluded from both the training and testing datasets. This approach allows us to objectively assess performance differences between the M_F and M_CF models on unseen temporal data (Fig. 4). When trained solely on FLUXNET 2015 data, M_F failed to accurately simulate GPP at the Xilingol site, particularly underestimating peak values, resulting in a low R² of only − 0.24. After additional Chinese grassland flux sites were incorporated, M_CF showed marked improvement, with R² increasing to 0.68 (Fig. 4b). Furthermore, M_CF successfully captured the seasonal dynamics of GPP (Fig. 4c), in contrast to M_F, which failed to reproduce the observed temporal patterns. To further confirm the model’s long-term robustness, we expanded our evaluation to encompass the entire available time series for the Chinese sites. Over the complete temporal coverage, the M_CF model maintained high predictive accuracy, achieving R² values of 0.64 at the Xilingol site (Fig. S2) and 0.87 at the Damxung site (Fig. S3), significantly outperforming the baseline M_F model. These results demonstrate that the inclusion of localized Chinese grassland observations substantially enhances the model’s performance and confidence for simulating long-term GPP dynamics in China’s temperate grasslands.

Fig. 4.

Performance comparison of the M_F and M_CF models at the Xilingol site (2007–2008). (a) and (b) are scatter plots of simulated versus observed GPP for M_F and M_CF, respectively; (c) is the comparison of the seasonal temporal patterns of GPP simulated by M_F and M_CF

To further assess the transferability of the model, we conducted an independent validation using nine additional grassland FLUXNET sites that were not used in model training. For each site, monthly GPP was first predicted by the M_CF model and then aggregated to annual cumulative GPP for validation. The results showed that M_CF maintained good predictive performance across these independent sites, with an annual-scale R² of 0.51 and an RMSE of 502.28 g C m^− 2 yr^− 1. These results indicate that the improved model retained robust generalization ability.

We further compared M_CF with several widely used GPP products at the independent validation sites based on annual GPP. Among all products, M_CF exhibited the highest agreement with site-observed annual GPP (R² = 0.51), followed by PML-v2 GPP (R² = 0.39), GLASS GPP (R² = 0.32), and MODIS GPP (R² = -0.15) (Fig. 5). This comparison further highlights the superior performance of M_CF for estimating grassland GPP at independent sites.

Fig. 5.

Comparison of M_CF and existing GPP products against observed annual GPP at independent validation sites

Trends in annual total GPP

The interannual variation in grassland ecosystem GPP in the Hulun Lake Basin from 2001 to 2023, as simulated by the M_CF model, is shown in Fig. 6. Over this period, the total annual GPP increased from approximately 64.39 Tg C in 2001 to 74.62 Tg C in 2023, with an average annual increase of 0.45 Tg C yr^− 1, indicating an overall upward trend in carbon assimilation by the grassland ecosystem. Prior to 2008, the increase was relatively minor, with the annual GPP rising by only 0.01 Tg C yr^− 1 on average. After 2008, however, the trend became significantly stronger, with the average annual increase accelerating to 0.45 Tg C yr^− 1.

Fig. 6.

Interannual changes in total GPP in the Hulun Lake Basin from 2001 to 2023

The observed increase in GPP is associated with regional changes in vegetation and climatic conditions (Fig. 7). NIRv showed a significant upward trend, suggesting a recovery in grassland productivity (Fig. 7a). Climatic variables—including Temp, Pre, Srad, and VPD—exhibited a certain trend of change, although none of these changes were statistically significant (Fig. 7b − e). Nevertheless, the increases in Temp and Pre should not be ignored. Relative to the first five years of the study period, mean Temp and Pre during 2021–2023 increased by 0.45 ℃ and 6.68 mm, respectively. However, pronounced interannual fluctuations likely diminished the statistical detectability of these trends.

Fig. 7.

Interannual variations in environmental factors of NIRv (a), Srad (b), Pre (c), Temp (d), and Vpd (e) in the Hulun Lake Basin from 2001 to 2023

During the study period, annual GPP ranged from 60.99 to 81 Tg C yr^− 1, indicating substantial interannual variability. A sharp decline in GPP was observed from 2013 to 2016, with values dropping from 76.69 Tg C yr^− 1 to 63.37 Tg C yr^− 1. This level was notably lower than the average during 2008 − 2012, which was approximately 70 Tg C yr^− 1. This decline was likely driven by extreme drought conditions induced by the strong El Niño event of 2015 − 2016, as reflected in the anomalous patterns of Pre and Temp during that period (Fig. 7c and d). Since 2016, GPP has shown a steady recovery, reaching its peak in 2018. This process was accompanied by increases in both Temp and Pre. Therefore, the climatic shifts of Temp and Pre may have contributed to the observed rise in GPP.

Spatial patterns of GPP

The spatial distribution of average GPP in the grassland ecosystems of the Hulun Lake Basin during 2001 − 2023 is shown in Fig. 8. GPP exhibited a distinct spatial gradient, with lower values in the west and higher values in the east. Based on the average GPP over the 23-year period (Fig. 8f), the eastern grasslands generally showed higher productivity, ranging from 600 to 1000 g C m^− 2 yr^− 1. In contrast, GPP in the western region, particularly areas surrounding Hulun Lake, was typically below 600 g C m^− 2 yr^− 1. The overall regional average was approximately 676 g C m^− 2 yr^− 1.

Fig. 8.

Temporal changes in the spatial distribution of GPP in the Hulun Lake Basin. (a), (b), (c), (d), and (e) show the average GPP distributions for the periods 2001 − 2005, 2006 − 2010, 2011 − 2015, 2016 − 2020, and 2021 − 2023, respectively. (f) presents the overall average for 2001 − 2023

GPP exhibited a consistent increasing trend across the study area over the 23-year period, with regional variations in the rate of increase. The average GPP increased by 198.16 g C m^− 2 yr^− 1 between 2001 and 2005 (Fig. 8a) and 2021–2023 (Fig. 8e) in the eastern region near the Greater Khingan Mountains. In contrast, grasslands in the western region, particularly near Hulun Lake, experienced a smaller increase of 127.68 g C m^− 2 yr^− 1 over the same period.

Based on estimates from the Sen’s slope method (Fig. 9), the most rapid increase was observed in regions adjacent to the Greater Khingan Mountains, with rates exceeding 10 g C m^− 2 yr^− 1. This increase was extremely significant. The northern part of the Hulun Lake region exhibited a moderate increase of approximately 7–10 g C m^− 2 yr^− 1, with significance levels generally ranging from significant to slight significant. The southern region of Hulun Lake showed the slowest rate of change, typically less than 5 g C m^− 2 yr^− 1, and most of the increases in this area were not statistically significant.

Fig. 9.

Spatial distribution of GPP trends in the Hulun Lake Basin from 2001 to 2023. (a) shows the Sen’s slope (g C m^− 2 yr^− 1), while (b) illustrates trend significance based on the MK test. Trends are classified into extremely significant increase (ESI), significant increase (SI), Slightly significant increase (SSI), Non-significant (NS) and Significant decrease (SD)

Driving factors of GPP

By computing SHAP values, we identified the relative importance of environmental factors in simulating GPP as shown in Fig. 10.

Fig. 10.

Annual SHAP values of NIRv (a) and environmental factors of Srad (b), Pre (c), Temp (d), in the Hulun Lake Basin

From 2001 to 2023, NIRv exhibited the highest SHAP value among all factors and showed a slight upward trend. This highlights its critical role in GPP estimation (Fig. 10a). Among the meteorological factors, the SHAP values ranked from low to high as VPD (Fig. S4), Pre, Temp, and Srad, all of which were significantly lower than that of NIRv (Fig. 10b-d, respectively). The SHAP value trends for Temp and VPD remained relatively stable, whereas those for Srad and Pre exhibited a gradual increase, suggesting a strengthening role in GPP estimation over time.

To further investigate the spatial dynamics of meteorological drivers over the past 23 years, this study employed the SHAP values of Pre, Temp, and Srad to construct a ternary phase diagram, as shown in Fig. 11.

Fig. 11.

Spatial distribution of annual SHAP values of environmental drivers in the Hulun Lake Basin (a: 2001, b: 2023)

During the period 2001–2005, the northwestern part of the basin (west of Hulun Lake) was primarily influenced by Srad, with a notable secondary contribution from Pre. In portions of the southwestern region, Temp emerged as another key driver alongside Srad. In the central area, the contributions of Pre, Temp, and Srad were relatively balanced, indicating that GPP was shaped by multiple factors without a single dominant driver. Toward the eastern margin, adjacent to the Greater Khingan Mountains, there was a gradual transition to joint dominance by Srad and Pre.

As shown in Fig. 11, when comparing the periods 2020 − 2023 to 2001 − 2005, the relative contribution of Pre and Temp increased substantially in specific sub-regions. Within the northwest region, Pre emerged as a considerably stronger driver. In the southwestern region, the area primarily influenced by Temp expanded noticeably. In the central region, Pre became the dominant driver, with its zone of influence extending further eastward.

To characterize the spatial expansion of precipitation influence, we quantified shifts in relative importance using the difference in mean SHAP values (ΔSHAP) between 2020 and 2023 and 2001–2005. The analysis revealed a massive expansion of precipitation-constrained regions (Fig. 12). The driving force of Pre was enhanced across 74.61% of the basin, predominantly indicating areas where water availability has progressively overshadowed temperature and radiation as the primary interannual constraint. In contrast, the areas showing an enhanced influence of Srad and Temp were highly localized, accounting for only 15.12% and 10.28% of the total area, respectively.

Fig. 12.

Spatial distribution of regions with enhanced climatic drivers. Enhanced precipitation driving force was defined for pixels satisfying all of the following criteria: ΔSHAP_Pre > 0, ΔSHAP_Pre > ΔSHAP_Temp, and ΔSHAP_Srad ≤ 0. Enhanced temperature driving force was defined using the same logic. Enhanced solar radiation driving force was defined as regions with ΔSHAP_Srad > 0

To further disentangle the direct and indirect effects of climatic variables on GPP, we conducted a PLS-SEM analysis with NIRv as a mediating variable between climate and GPP (Fig. 13). The model showed strong explanatory and predictive performance, with R² values of 0.79 for NIRv and 0.92 for GPP, Q² values of 0.77 and 0.91, respectively, and a GOF of 0.92. NIRv had a strong positive effect on GPP (0.723), confirming its key mediating role. Among the climatic variables, precipitation showed the strongest positive total effect, with both direct and indirect pathways contributing positively. In contrast, temperature and solar radiation exhibited contrasting direct and indirect effects, resulting in weak net total effects, while VPD showed only minor influences overall. These results indicate that climatic effects on GPP were partly mediated by NIRv, with precipitation exerting the strongest overall influence.

Fig. 13.

PLS-SEM of direct and indirect climatic effects on GPP mediated by NIRv. Arrows denote causal paths among climatic variables, NIRv, and GPP. Red arrows indicate positive effects, whereas blue arrows indicate negative effects. Line thickness is proportional to the absolute standardized path coefficient, with thicker line representing stronger effects. Numbers next to lines are standardized path coefficients

Overall, the spatial distribution of meteorological driver importance experienced marked changes over the past 23 years, among which the enhanced role of precipitation was especially prominent.

Discussion

The importance of China grassland fluxes observation for GPP simulation in Hulun lake basin

According to simulations from process-based and remote sensing–driven models, grassland GPP in China generally ranges from 300 to 800 g C m^− 2 yr^− 1 [22, 24]. Our results show that the mean GPP of the grassland in Hulun Lake Basin from 2001 to 2023 was 660 ± 180 g C m^− 2 yr^− 1, which falls within this range. Considerable discrepancies exist among different models in simulating China’s grassland GPP. The largest divergences occur in relatively dry and cold regions, such as the central and eastern grasslands of Inner Mongolia, where the grassland of Hulun lake basin is situated [22].

Our M_F model trained on the Fluxnet2015 dataset exhibited non-negligible errors (R² less than 0) in simulating GPP for the Xilingol typical temperate grassland in Inner Mongolia, with particularly poor simulation of seasonal GPP variations (Fig. 4). M_F performance further deteriorated at the drier and colder alpine meadow site in Damxung (Fig. S1). Previous studies have demonstrated similar challenges in this region, where satellite-based models showed an average relative uncertainty of 49% when validated against flux observations in arid climates [24]. These simulation uncertainties are attributed to the unique compound constraints imposed by coupled water availability and temperature fluctuations on vegetation growth under complex climatic conditions [19]. Targeted optimizations addressing these region-specific bioclimatic interactions are required to enhance model reliability for analyzing spatiotemporal GPP dynamics and driving mechanisms in such ecosystems.

The performance of machine learning models in GPP modeling is highly dependent on the feature distribution, regional coverage, and variable representativeness of training data [42]. The limited number of Chinese grassland flux observation sites in the FLUXNET 2015 dataset has resulted in suboptimal GPP simulations in arid/cold regions. After incorporating additional flux observation data from arid/cold grasslands in northern China, our M_CF model demonstrated significantly improved simulation accuracy for both temperate grasslands (Fig. 4) and alpine grasslands (Fig. S1). The improved performance of M_CF was further supported by the external validation using nine additional grassland FLUXNET sites. Compared with existing GPP products, M_CF achieved the highest agreement with site-observed annual GPP. This finding indicates that integrating region-relevant grassland flux observations with NIRv and climatic variables enhanced the model’s ability to capture interannual variations in grassland GPP. In contrast, existing large-scale GPP products are developed for broader applications and may be less effective in representing the environmental constraints and phenological characteristics of temperate grasslands, particularly in water-limited and climatically heterogeneous regions. However, the model still exhibits limitations in capturing extreme high GPP values. Future work should consider increasing the number of high-value samples or optimizing the RF algorithm to enhance its fitting ability for extreme values.

4.2 Sustained increases and fluctuations of GPP

From 2001 to 2023, the grassland GPP in the Hulun Lake Basin exhibited a sustained increasing trend, consistent with previously reported global terrestrial ecosystem GPP growth trends [5, 52]. Simulations based on process-based and remote sensing-driven models suggest that the GPP growth rate for Chinese grasslands ranges from 0.68 to 3.13 g C m^− 2 yr^− 1 [22], while the growth rate in the Hulun Lake Basin is higher, ranging from 4 to 12 g C m^− 2 yr^− 1 [11]. Our findings indicate an average GPP increase of 6.85 g C m^− 2 yr^− 1 for the temperate grasslands in the Hulun Lake Basin, with 5.39 g C m^− 2 yr^− 1 near Hulun Lake and 8.13 g C m^− 2 yr^− 1 near the Greater Khingan Mountains, aligning well with prior estimates.

The increase in GPP in the Hulun Lake basin grasslands is attributed to the ongoing vegetation recovery in recent years, which is reflected in changes in vegetation indices. From 1982 to 2016, NDVI in Hulun Lake basin rose continuously, with growth rates among the highest in Inner Mongolia’s temperate grasslands [67], although a browning trend was observed from 2000 to 2019 [26]. The 2015/16 strong El Niño suppressed central Inner Mongolian precipitation, inducing drought and a pronounced NDVI decline in temperate grasslands [27, 66]. We find that this event reduced NIRv in 2016 to its lowest level since 2001, with only slight recovery in 2017, and that extreme climate in 2023 again caused a significant NIRv decrease (Fig. 7a). These findings indicate that, although temperate grassland vegetation is highly sensitive to environmental constraints and subject to acute short-term impacts from extreme climate, it exhibits long-term recovery [54].

Greening trends have shifted in some Chinese grassland regions. On the Tibetan Plateau, warming initially promoted greening, but insufficient moisture has led to recent browning [2]. In contrast, Hulun Lake basin, located in northern China’s drylands, has seen increased precipitation in recent years, alleviating water limitations [25] and indicating continued potential for future greening.

The change of driving factors on GPP of grassland in Hulun lake basin

Based on our findings, NIRv exhibited stronger importance than temperature and precipitation in driving grassland GPP variations. This contrasts with previous NDVI-based studies reporting weaker contributions of vegetation indices compared to climatic factors [74]. The intrinsic correlation differences between vegetation indices (NIRv vs. NDVI) and GPP may influence the assessment of their driving effects on GPP dynamics. Global flux tower-based analyses revealed stronger correlations between NIRv and GPP than NDVI in grassland ecosystems [36, 51], with NIRv demonstrating enhanced capability in tracking temporal GPP variations [65]. These distinctive characteristics of NIRv could explain its greater explanatory power over environmental factors in our study.

GPP was strongly influenced by climatic variability [39, 51]. Our study highlights radiation as the dominant climatic factor affecting grassland GPP in the Hulun Lake basin. While temperature has generally been identified as the primary driver of GPP variations in Inner Mongolian grasslands [35], radiation exerts stronger control over GPP in typical temperate grasslands of northeastern Inner Mongolia, particularly within the Hulun Lake Basin adjacent to the Greater Khingan Mountains [13]. This spatial pattern aligns with our findings (Fig. 11).

Machine learning approaches integrated with SHAP analysis have demonstrated the significant contribution of temperature to grassland GPP dynamics in previous studies [76]. However, in cold-arid regions, precipitation deficits constrain the sensitivity of GPP to temperature increases and may even dominate GPP variations [13, 32]. Furthermore, rising global atmospheric CO₂ concentrations enhance GPP through fertilization effects [50], though water limitations can attenuate this CO₂-driven enhancement [46, 72]. These findings align with experimental CO₂ enrichment studies conducted in Inner Mongolian grasslands [45, 77].

Our results reveal a 50.2% increase in precipitation over the past two decades in the Hulun Lake Basin, with spatial heterogeneity driven by monsoon trajectories and topographic effects—most pronounced in areas adjacent to the Greater Khingan Mountains. Improved precipitation conditions likely enhanced GPP not only by directly alleviating water limitation, but also by indirectly promoting vegetation activity, as reflected by NIRv. This mechanism helps explain why the eastern part of the basin became the most significant contributor to GPP enhancement within the study area. A similar positive effect of precipitation on ecosystem carbon uptake has also been reported in a cold, semi-arid alpine meadow, where precipitation was identified as an important control on growing-season NEE [71]. Given projected increasing precipitation trends in northern China [4], the regulatory role of water availability in GPP dynamics may further intensify in the future.

Conclusion

This study developed a RF model to estimate grassland GPP in the Hulun Lake Basin by integrating flux data from FLUXNET and ChinaFLUX, aiming to reveal long-term GPP responses to regional climate change. Results show that incorporating region-specific Chinese flux tower data significantly improves GPP simulation accuracy. From 2001 to 2023, annual GPP showed a notable upward trend, despite considerable interannual variability linked to extreme climatic events. Spatially, GPP exhibited a distinct west–east gradient, with higher values near the Greater Khingan Mountains and lower levels around Hulun Lake and the southwestern basin. SHAP analysis identified Srad as the dominant climatic driver in the early years, while the influence of Pre and Temp has grown over the past two decades. In particular, increased precipitation in the eastern basin has improved moisture availability, potentially amplifying vegetation responses to temperature and CO₂, and has emerged as a major driver of GPP dynamics in that region. These findings underscore the importance of localized data for accurate GPP modeling and offer valuable indicators for assessing climate impacts on grassland ecosystems. Future work should improve model responsiveness to extreme events and better capture climate-driven variability in GPP to support adaptive ecosystem management under ongoing climate change.

Author contributions

Tong Sun: Conceptualization; Methodology; Validation; Investigation; Formal analysis; Data Curation; Writing - Original Draft; Writing - Review & Editing; Visualization; Funding acquisition; NanShan: Conceptualization; Methodology; Validation; Investigation; Data Curation; Writing - Original Draft; Writing - Review & Editing; Supervision; Funding acquisition; Xuehe Lu: Conceptualization; Methodology; Validation; Formal analysis; Data Curation; Writing - Original Draft; Writing - Review & Editing; Visualization; Bingshuai Cao: Validation; Data Curation; Visualization; Wenjing Li: Validation; Data Curation; Visualization; Qian Zhang: Validation; Data Curation; Visualization; Funding acquisition; Qing Huang: Validation; Methodology ; Visualization; Haidong Zhang: Validation; Data Curation; Formal analysis; Qianghao Zeng: Validation; Data Curation; Visualization; Suwan Chen: Validation; Data Curation; Visualization;

Funding

This research was jointly funded by the First Phase of the Project of Strengthening the Scientific and Technological Research Capacity of Hulun Lake Natrue Reserve (HSZCS-C-F-210094), the Ecological Security Investigation and Assessment Project of Hulun Lake (HSZCS-G-F-210059), the Postgraduate Research & Practice Innovation Program of Jiangsu Province (SJCX25_1876), and the Taishan Scholar Project (tsqn202306210).

Data availability

All relevant data supporting the findings of this study are available on request.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

The authors all give consent for publication.

Competing interests

The authors declare no competing interests.