Background & Summary

As critical ecosystems, grasslands hold immense ecological and economical significance1,2. They serve as crucial carbon reservoirs, storing approximately 34% of the total organic carbon in terrestrial ecosystems worldwide3. Despite their ecological importance, a substantial portion of grasslands is managed for livestock grazing to meet the increasing human demand for meat and milk4,5. Grazing imposes considerable pressure on grasslands biomass5, affecting ecosystem carbon cycles6 and the surface energy balance7. Grasslands are also at risk of degradation in case of overgrazing8. Moreover, the intensification of grasslands management increases CH4 and N2O emissions from ruminant livestock, manure management, and fertilizer application9. Previous studies have shown that CH4 and N2O emissions from livestock systems contribute to 24% and 6% of their total emissions10,11, respectively. This results in a considerable contribution to climate change due to their high global warming potential9,12. However, the absence of explicit spatial information on grazing intensity greatly limits the assessment of grazing impacts as well as grasslands carbon cycles.

China possesses the world’s second-largest grassland resources, which play a crucial role in supporting livestock production13. Yet, the overall quality of grassland ecosystems in China remains relatively fragile14, persistently affected by escalating global changes and anthropogenic activities2. China is the world’s largest producer and consumer of livestock products. In 2021, the meat consumption reached nearly 100 million tons15, representing 27% of the global total16. Moreover, the proportion of beef and mutton consumption continues to rise15. In the past three decades, China’s livestock population has experienced a vast increase17, exacerbating the irrational use of grasslands. Overgrazing has led to varying degrees of degradation across extensive grasslands in China, posing a threat to the nation’s ecological security and food security2,4. Furthermore, grazing intensity in China is highly uneven due to high spatial heterogeneity in economic development and grassland production18. Therefore, there is an urgent need to generate a reliable, high spatiotemporal resolution grazing intensity dataset for China.

Early information on grazing intensity primarily originated from field surveys, with the position of livestock obtained from Global Positioning System (GPS) devices fitted on livestock19. However, this approach was time-consuming and labor-intensive, making it impractical for extensive geographical areas. In 2007, the Food and Agriculture Organization of the United Nations (FAO) released the first Gridded Livestock of the World (GLW1) at 5 km × 5 km resolution, providing a global livestock density distribution map for 200220. The gridded livestock density was estimated based on the established relationship between subnational livestock density and various environmental factors (e.g., anthropogenic, climatic, and topographical)20,21,22. Specifically, the observed livestock density was calculated using subnational census data and the corresponding administrative unit areas with suitability masking. This observed livestock density was then modelled against environmental factors using multiple linear regression. Finally, the trained model was applied to pixel-level environmental factors to estimate pixel-level livestock density. In 2014, the FAO collected more recent and detailed livestock census data, releasing the updated version of GLW2, providing the global distribution map of cattle, pigs, and chickens, and partial distribution map of ducks in 2006, with a spatial resolution of 0.083° (approximately 10 km)21. In the subsequent years, FAO released GLW3 (for 2010) and GLW4 (for 2015) in succession by replaced multiple linear regression with random forest22. And the global distribution map of cattle, buffalo, horse, sheep, goat, pig, chicken, and duck were obtained22. Despite successive improvements in the algorithm resulting in GLW221, GLW322, and GLW422 datasets, the low spatial and temporal resolution, along with update delays, pose challenges for long-term and regional research. To address these limitations, Zhou et al.23 developed the annual Gridded Dataset of Grazing Intensity (GDGI) maps with a resolution of 100 m based on the FAO’s algorithm, covering the period from 1990 to 2020 for the Qinghai-Tibetan Plateau23. However, the FAO’s algorithm has several limitations, restricting the usability of data based on this algorithm. First, the FAO’s algorithm assumes that the relationship between environmental factors and livestock density is identical at both the administrative and pixel level24. In reality, the average condition at the county level may not be representative of all pixels, especially in counties that contain different climate types. Applying county-level models to the pixel level could lead to biases in the spatial prediction of grazing intensity. To address this issue, Meng et al.18 employed the difference in vegetation status with and without grazing as a weight to calculate pixel-level grazing density. Instead of relying on administrative-level averages, this approach was utilized to map grazing intensity on the Qinghai-Tibetan Plateau from 1982 to 201518. Second, the FAO simplified livestock management systems, failing to distinguish between pasture-based and crop-based livestock. It also assumed that the distribution of all livestock is constrained by environmental conditions. However, humans can modify the environment to accommodate livestock survival, resulting in livestock production, especially intensive farming, having minimal dependence on the surrounding environment25. Therefore, there are significant biases in the FAO’s algorithm for mapping the distribution of intensively livestock system.

In summary, despite several efforts (Table 1), significant uncertainties persist in estimating grazing intensity for China. First, current algorithms failed to distinguish between pasture-based and crop-based livestock. In many semi-pastoral counties of China, a substantial portion of livestock feeds on cereals26, leading to potential overestimation of grazing intensity when not properly differentiated in census data. Second, existing datasets in China primarily focused on the Qinghai-Tibetan Plateau18,27, thereby neglecting the two largest pastoral regions, Inner Mongolia and Xinjiang28. Third, the lack of cross-scale validation undermines the reliability of these datasets, further limiting comprehensive assessment of grazing impacts on grasslands in China.

Table 1 Summary of the dataset produced in this study and other publicly available gridded grazing intensity datasets.
Full size table

Therefore, this study aims to develop a Long-term (1980–2022) High-resolution (1980–2000: 0.1° × 0.1°, 2001–2022: 0.0025° × 0.0025°) Grazing Intensity (LHGI) dataset for the grassland ecosystem of China. The specific objectives are: (1) to differentiate between pasture-based and crop-based livestock at the county level; (2) to generate a publicly available database of long-term, high-resolution gridded grazing intensity maps in China based on county-level pasture-based livestock; and (3) to evaluate the accuracy of the grazing intensity dataset. This dataset serves as crucial foundational data for establishing grassland management scheme, assessing grazing impacts on carbon cycle, and estimating potential greenhouse gas emission mitigation from livestock systems.

Methods

Study area

Grasslands in northern arid regions and the Qinghai-Tibetan Plateau comprise approximately 93% of grasslands in China29, giving rise to the key pastoral regions. This study focused on mapping the grazing intensity across these major pastoral regions, encompassing Inner Mongolia, Xinjiang, Tibet, Qinghai, Gansu, Ningxia, as well as 58 pastoral counties from four municipalities/autonomous prefecture in Sichuan (Fig. 1). The study area is located between 73°30′E–118°49′E and 26°3′N–49°26′N, covering rich natural grassland resources (Fig. 1).

Fig. 1
figure 1

Location of the study area and the distribution of grassland types. The grey solid lines indicate provincial boundary.

Full size image

Framework

This study aimed to generate a Long-term High-resolution Grazing Intensity (LHGI) dataset, which is a component of Terrestrial Ecosystem Disturbance (TED) dataset30,31, named as TED-LHGI. The core assumption of this study is that the difference between the satellite-based (i.e., actual observations) and climate-based (i.e., without grazing and influenced only by climate conditions) growth of grassland can be considered as a disturbance on grassland natural growth caused by human activities32,33. The magnitude of this difference characterizes the degree of disturbance. Grazing intensity was then gridded using the difference as a weighted factor, combined with county-level census livestock data. The methodology involved four steps to map the grazing intensity: (1) estimation of climate-based growth of grassland, (2) estimation of pasture-based and crop-based livestock populations, (3) mapping of grazing intensity, and (4) validation of dataset accuracy and uncertainty analysis.

Estimation of climate-based growth of grassland

In this study, a random forest (RF) regression model was developed to simulate climate-based grassland growth. The growing season averaged Normalized Difference Vegetation Index (NDVIg) was used as the dependent variable to measure grassland growth18,32. The objective was to simulate the climate-based NDVIg for each 0.1° × 0.1°pixel. Although the growing season of China’s grasslands typically spans from May to September34, the period from late August to early September marks the beginning of the mowing, which can cause a significant decline in NDVI. Consequently, we focused on the period from May to August for calculating NDVIg. The simulation process of the climate-based NDVIg involved the following steps:

  1. (1)

    Selection of Predictors: 20 potential predictors, including climate and topography factors35,36, were considered as independent variables and resampled to 0.1° × 0.1° (Table S1). Based on the correlation analysis of all potential predictors, multicollinearity screening was conducted to avoid including highly correlated (|r| > 0.8) predictors simultaneously in the model simulation37. Despite the RF algorithm’s inherent ability to manage multicollinearity, the goal was to find a small set of predictors with explanatory and predictive power to improve model performance. For two factors with a correlation coefficient higher than 0.8, retain the one with the greater explanatory power. Finally, 10 factors were selected for the RF model, including minimum temperature, total precipitation, mean soil moisture, mean soil temperature, and mean relative air humidity (considered during the growing season), as well as downward surface shortwave radiation, direct photosynthetically active radiation, diffuse photosynthetically active radiation (averaged annually), elevation, and slope (Table S1).

  2. (2)

    Selection of Sampling Points: Pixels with no grazing and/or light grazing, where NDVIg is primarily influenced by climate, were selected as sampling points to simulate the climate-based NDVIg33. In this study, it was assumed that no grazing or light grazing areas are uniformly distributed spatially. Within a given window, the pixel with the maximum NDVIg was considered as the no grazing or lightly grazing area33. Specifically, for the period 1980 to 2000, due to the coarse spatial resolution of the available NDVI (0.1° × 0.1°), a 1° × 1° window was used to identify areas with no grazing or light grazing. For the period 2001 to 2022, benefiting from the high resolution of NDVI (0.0025° × 0.0025°), a 0.1° × 0.1° window was used to search for no grazing or light grazing areas (Fig. 2).

    Fig. 2
    figure 2

    Framework for mapping grassland utilization intensity.

    Full size image
  3. (3)

    Model Training and Simulation: The RF model was run annually at the province level or within ecoregions classified by grassland types. For 1980–2000, 75% of the samples were randomly extracted for training the RF model, and the remaining 25% were used to test model accuracy. The trained RF model was applied to predictors to simulate the climate-based NDVIg for each 0.1° × 0.1° pixel (Fig. 2). For 2001–2022, each 0.1° grassland pixel has a sample point. Therefore, a four-fold cross-prediction was applied to simulate the climate-based NDVIg (Fig. 2). For a given province or ecoregions, all samples were randomly divided into four sets, and four RF models were established. For each iteration, three sets of samples were utilized to construct an RF model, which was subsequently employed to estimate NDVIg at the locations of the 4th set of samples (Fig. 2). Furthermore, to enhance the assessment of the RF model’s performance, 25% of the three sets of samples used for model construction were randomly set aside for accuracy validation. Finally, the potential NDVIg for each 0.1° × 0.1° pixel was estimated (Fig. 2).

In this study, NDVIg was converted to aboveground biomass, as numerous studies have demonstrated the non-linear relationship between biomass and NDVI38,39. This means that across various levels of biomass in grassland vegetation, an equal decrease in NDVIg does not necessarily represent an equivalent degree of disturbance. Therefore, in this study, a grassland NDVI-biomass model39 for global grassland ecosystems was used to calculate aboveground biomass based on NDVIg (Eq. 1).

$${Agblive}=52.2\times {{NDVI}}_{g}^{2}+173.9\times {{NDVI}}_{g}$$
(1)

where Agblive (g C m−2) is the aboveground live biomass carbon density. Then, the difference (Diff) between climate-based biomass and satellite-based biomass was calculated as:

$${{Diff}}_{i}={{Agblive}}_{{c\; i}}-{{Agblive}}_{{s\; i}}$$
(2)

where Agblivec i, and Agblives i are the climate-based and satellite-based aboveground live biomass carbon density of the ith pixel, respectively, calculated by the climate-based and satellite-based NDVIg according to Eq. (1), respectively.

Estimation of pasture-based and crop-based livestock populations

In this study, we estimated the number of pasture-based and crop-based livestock for each county based on the potential livestock carrying capacity of grasslands. According to the National Standard for Calculation of Rangeland Carrying Capacity (National standard: NY/T 635–2015)40, the potential livestock carrying capacity was calculated as:

$${SU}_{Pi}=\frac{{Agblive}_{Pi}}{{Agblive}_{S}}$$
(3)
$${{Agblive}}_{Pi}={{Agblive}}_{ci}\times \left(1+{G}_{r}\right)\times {U}_{r}\times {{SH}}_{f}$$
(4)
$${{Agblive}}_{S}=S\times \left(1-{W}_{r}\right)\times {C}_{f}\times D$$
(5)

where SUP i is the potential livestock carrying capacity of the ith pixel, measured in standard sheep unit (SU). AgbliveP i is the available grassland biomass (g C); AgbliveS is the biomass consumed annually by a standard sheep unit. Gr is the grass regeneration rate; Ur is the grass utilization rate; SHf is the standard hay conversion factor; S is the daily hay feed for one standard sheep (1.8 Kg SU−1 day−1); Wr is the water content ratio of standard hay (14%); Cf is the conversion coefficient between biomass and carbon (0.45 g C g−1); and D is the grazing duration within each year (365 days)40.

Gr, Ur, and SHf vary across grassland types. The Vegetation Map of the People’s Republic of China (Table 2) provides the distribution of grassland types, and the National Standard for Calculation of Rangeland Carrying Capacity (National standard: NY/T 635–2015)40 provides the values of Gr, Ur, and SHf for each grassland type (Table S2). Based on these two data, this study obtained the values of Gr, Ur, and SHf for each grassland pixel. Ur provided by the National Standard refers to the proper utilization rate that can simultaneously meet the needs of sustainable grassland production and livestock demand40. We further estimated the actual value of Ur based on the proper Ur and multi-year trends in NDVIg, taking into full consideration the actual use of grasslands (overgrazing, suitable grazing, and light grazing). For a grassland pixel, a negative slope in the NDVIg time series indicates grassland degradation, implying that its actual Ur exceeded the proper Ur. In such cases, the actual Ur was considered to be 20%40 higher than the proper Ur. Conversely, a positive slope suggests the opposite, and the actual Ur was considered to be 20% lower than the proper Ur.

Table 2 Summary of data sources.
Full size table

In this study, grassland resources are assumed to be the preferred choice for livestock feed. For a given county j, the potential livestock carrying capacity (SUP) of all grassland pixels was aggregated to determine the total potential livestock carrying capacity (Eq. 6). This total potential livestock carrying capacity was then compared with the county-level census livestock data to determine the number of pasture-based and crop-based livestock (Eqs. 7, 8).

$${TSU}_{Pj}={\sum }_{i=1}^{n}{SU}_{Pi}$$
(6)
$$\begin{array}{ccc}{{SU}}_{{pas}j} & = & \left\{\begin{array}{cc}{{SU}}_{{cen}j}, & {{SU}}_{{cen}j}{\rm{\le }}{{TSU}}_{Pj}\\ {{TSU}}_{Pj}, & {{SU}}_{{cen}j} > {{TSU}}_{Pj}\end{array}\right.\end{array}$$
(7)
$$\begin{array}{ccc}{{SU}}_{{cro}j} & = & \left\{\begin{array}{rc}0, & {{SU}}_{{cen}j}\le {{TSU}}_{Pj}\\ {{SU}}_{{cen}j}-{{TSU}}_{Pj}, & {{SU}}_{{cen}j} > {{TSU}}_{Pj}\end{array}\right.\end{array}$$
(8)

where TSUP j is the total potential livestock carrying capacity for county j; n is the number of pixels in county j; SUP i is the potential livestock carrying capacity of the ith grass pixel of county j (Eq. 3); SUpas j and SUcro j are the numbers of pasture-based and crop-based livestock for the jth county, respectively; SUcen j is the number of census livestock for county j; TSUP j is the total potential livestock carrying capacity for county j.

Mapping grazing intensity

In this study, pasture-based livestock was allocated from county or municipal level to the pixel level using grassland utilization intensity as a weight (Eq. 9).

$${{SU}}_{{pas}i}={{SU}}_{{pas}j}\times \frac{{{Diff}}_{i}}{\mathop{\sum }\limits_{i=1}^{n}{{Diff}}_{i}}$$
(9)

where SUpas i and SUpas j are the pasture-based livestock for pixel i and county/city j, respectively; Diffi is the difference between climate-based biomass and satellite-based biomass for pixel i; and n is the number of pixels in county j.

Validation of data accuracy and uncertainty analysis

This study involves two aspects of accuracy validation: the RF model and the LHGI dataset. For the RF model, 25% of the samples were randomly set aside for model validation. For the LHGI dataset, we conducted both pixel-level and county-level validation. We collected field grazing intensity data covering 1980–2022 from 73 sites, of which 65 were obtained from the literature and 8 from field survey (Fig. 4a; Table S3). Besides, for each province, in several small-area cities, the total pasture-based livestock was allocated from the municipal level to the pixel level (Fig. S1). The county-level census data in these cities were then used for cross-scale validation.

Subsequently, the relationships between the estimates and the corresponding observations were assessed by linear regression. Then the coefficient of determination (R2), the slope of linear regression (Slope), and the root mean square error (RMSE) were calculated to assess the consistency of estimates and observations. The closer the R2 and Slope are to 1 and the smaller the RMSE, the better the estimates. The R2, Slope, and RMSE are calculated as:

$${R}^{2}=1-\frac{\sum _{i}{\left({x}_{i}-{y}_{i}\right)}^{2}}{\sum _{i}{\left(\bar{{x}_{i}}-{x}_{i}\right)}^{2}}$$
(10)
$${Slope}=\frac{{\sum }_{i=1}^{n}\left({x}_{i}-\bar{x}\right)\left({y}_{i}-\bar{y}\right)}{{\sum }_{i=1}^{n}{\left({x}_{i}-\bar{{x}_{i}}\right)}^{2}}$$
(11)
$${RMSE}=\sqrt{\frac{1}{n}\mathop{\sum }\limits_{i=1}^{n}{\left({x}_{i}-{y}_{i}\right)}^{2}}$$
(12)

where xi and yi are the estimates and observations, respectively, and n is the number of observations.

Due to the absence of county-level census data for certain indicators in specific years and/or provinces, such as the livestock off-take rate, averages from other years were utilized as substitutes. Therefore, we quantified the bias introduced by this method of interpolating data. For each missing indicator, we calculated the standard deviation (SD), and converted it to 95% confidence interval (95% CI) (Eq. 13) to characterize its variation range.

$$E=1.96\times \frac{{SD}}{\sqrt{n}}$$
(13)

where n is the sample size (i.e., the number of years with available data). These variation ranges were then aggregated into the overall variation range of the total livestock numbers according to Eq. 14. Finally, the percentage of the overall variation range relative to the total livestock number was used to characterize the magnitude of uncertainty.

Data Source

Census data of livestock at county level

The county-level livestock stocking numbers at the end of each year were collected from various provincial statistical yearbooks, encompassing cattle (excluding dairy cattle), goats, sheep, as well as large livestock such as horses, donkeys, and camels. For Inner Mongolia, Qinghai, and Sichuan, livestock data was only available at the municipal level after 2016. Furthermore, this study also considered the off-take number and offspring number. The off-take number of major meat animals (beef cattle, goats and sheep) was collected at the county level or calculated based on the municipal-level off-take rate. The number of calves and lambs stocked at the municipal or county level was estimated according to provincial-level census records. From 1980 to 2022, about 30% of the data is missing. For years lacking census data, the average values from other years were used as substitutes. Each type of livestock number was converted into standard sheep unit (SU) using Eq. 1418,40:

$${SU}={N}_{{sheep}}+0.8\times {N}_{{goats}}+5\times {N}_{{cattle}}+6\times {N}_{{large\_liv}}+0.6\times {N}_{{s\_lambs}}+0.48\times {N}_{{g\_lambs}}+1.5\times {N}_{{calves}}$$
(14)

where Nsheep, Ngoats, Ncattle, Nlarge_liv, Ns_lambs, Ng_lambs, and Ncalves are the numbers of sheep, goats, cattle, large livestock, sheep lambs, goats lambs and calves, respectively.

Satellite-based NDVI data

In this study, two NDVI datasets were used (Table 1). For 1980–2000, we utilized the biweekly NDVI from the Global Inventory Monitoring and Modelling Studies (GIMMS) group, derived from the National Oceanic and Atmospheric Administration’s Advanced Very-High Resolution Radiometer (NOAA/AVHRR) Land dataset at ~8 km spatial resolution41,42. We employed the maximum value composite approach to combine the two available NDVI images on a monthly basis. For 2001–2022, the NDVI derived from Moderate-resolution Imaging Spectroradiometer (MODIS) was utilized. The NDVI was calculated based on the surface spectral reflectance for red (620–670 mm) and near-infrared (NIR) (841–876 mm) bands at 250 m resolution. To minimize noise, the maximum value composite method was performed for the derived NDVI values within an 8-day period, and the Whittaker smoother was used to pre-process the initial MODIS-NDVI43. The GIMMS-NDVI was processed in a manner consistent and quantitatively comparable with MODIS, and calibrated within the dynamic range of MODIS-NDVI, therefore, exhibiting good consistency with MODIS-NDVI41.

Map of grassland coverage and types

In this study, the distribution of grasslands was derived from the China Land Cover Dataset generated by Yang et al.29 (Table 1). The dataset contains annual land cover information for China from 1985 to 2021 at a 30 m resolution, achieving an overall accuracy of 79.31%. The vegetation map of the People’s Republic of China was downloaded from the Plant Science Data Center (Table 1). This map illustrates the distribution of 868 basic vegetation classification units in China, from which grassland types can be identified.

Environmental variables

Topography factors were calculated from a digital elevation model with 90 m resolution downloaded from the Resource and Environmental Science Data Platform (Table 1). For temperature, precipitation, and relative humidity, plate smoothing splines were employed to create an interpolated daily dataset with a spatial resolution of 10 km × 10 km44,45. This interpolation utilized data gathered from 735 meteorological stations spanning from 1960 to 2022, obtained from the National Climate Center of the China Meteorological Administration44. Validation against 150 independent sites demonstrated high accuracy of this interpolated dataset44. Due to a lack of soil temperature/moisture and solar radiation observations at these meteorology stations, the soil temperature was derived from the fifth generation ECMWF atmospheric reanalysis (ERA5) daily surface meteorological data46, the soil moisture was download from the website of the Global Land Evaporation Amsterdam Model (GLEM)47, and the radiation factors were extracted from the Climatic Research Unit of the Japanese Reanalysis dataset (CRU-JRA)48. For soil moisture during the period of 2000–2020, a global daily surface soil moisture dataset at 1-km resolution was downloaded from the National Tibetan Plateau Science Data Center49. Finally, all environmental variables were processed to a spatial resolution of 0.1° × 0.1°.

Field grazing intensity data

The field grazing intensity data were collected from literatures and field questionnaires. In total, there are field grazing data from 73 sites covering the period 1980–2022, including 65 from the literatures and 8 from field questionnaires (Fig. 4a,Table S3). Literatures generally provide grazing intensity for a specific location in SU ha−1, which can be directly used in this study. The grazing intensity obtained from survey questionnaires is the ratio of the total number of livestock (SU) to the area of the pasture, as gathered from household surveys.

Data Records

The Long-term and High-resolution Grazing Intensity of grassland in China is available at https://doi.org/10.6084/m9.figshare.2619568450. The dataset is provided in GeoTIFF format with the spatial reference of WGS84 (EPSG: 4326). Each GeoTIFF file represents annual grazing intensity (unit: SU ha−1). These files can be downloaded and viewed using various GIS software program.

Technical Validation

Performance of RF model for simulating climate-based NDVI

The validation of the RF model showed its good performance in estimating potential NDVIg. The RF model demonstrated greater stability in performance across the period of 2001–2022 compared to the period of 1980–2000 (Fig. 3). Averaged over the periods 1980 to 2000 and 2001 to 2022, the slope of the regression line between estimated and observed values across all surveyed provinces/ecoregions ranged from 0.79 to 0.93 (Fig. 3a) and 0.77 to 0.93 (Fig. 3b), respectively. The RMSE ranged from 0.048 to 0.076 (Fig. 3a) and 0.044 to 0.078 (Fig. 3b), respectively, and the R2 was consistently above 0.8 (Fig. 3). Among the provinces/ecoregions, the RF model performed best in Qinghai for 1980–2000 and in Inner Mongolia for 2001–2022, with both the linear regression slope and R2 exceeding 0.9. In both periods, the RF model’s performance in Xinjiang was relatively lower, with the linear regression slope less than 0.8.

Fig. 3
figure 3

Performance of the RF model for each province/ecoregion. (a) The average value from 1980 to 2000, and (b) the average value from 2001 to 2022. ‘Others’ refers to the grasslands types in the 58 pastoral counties of western Sichuan, aside from alpine meadow.

Full size image

Validation of the LHGI dataset at pixel level and county level

We first compared the estimated grazing intensity and field grazing intensity at pixel level. For sites with multi-year estimates, we calculated the average value. Validation across 73 sites demonstrated strong agreement between estimated and field grazing intensity (Fig. 4), with a linear regression slope of 0.84, R2 of 0.78 and RMSE of 0.54 SU ha−1 (Fig. 4b).

Fig. 4
figure 4

Validation of the LHGI dataset at the pixel level. (a) The spatial distribution of the field grazing intensity data; (b) comparison of estimated and field grazing intensity. The grazing intensity distribution is for 2020, the dark dashed line indicates the 1:1 line, and the red solid line indicates the regression.

Full size image

We further compared the estimated and census pasture-based livestock at county level. The validation showed the excellent performance of LHGI in simulating the spatial variation of grazing intensity across all surveyed counties (Fig. 5). Averaged from 1980 to 2000, the R2 between estimated value and census value is 0.61 ± 0.04, the RMSE is 2.91 ± 0.29, and the slope of the linear regression is 0.84 ± 0.03 (Fig. 5a). Compared to 1980–2000, the estimated values showed higher consistency with census values during 2001 to 2022 (Fig. 5), with an averaged R2 of 0.84 ± 0.04 and a slope of the linear regression of 0.93 ± 0.05 (Fig. 5a).

Fig. 5
figure 5

Validation of LHGI dataset at the county level. The performance of the LHGI dataset for (a) 1980–2000 and (d) 2001–2022; County-level comparison of estimated and census sheep unit (SU) of (b) 1986, (c) 1996, (e) 2006, and (f) 2016. The investigated counties refer to those whose livestock were allocated from the municipal level. The dark dashed lines indicate the 1:1 line, and the red solid lines indicate the regression.

Full size image

Spatial pattern of grazing intensity

The spatial pattern of grazing intensity derived from the LHGI dataset remained consistent across various time points from 1980 to 2022 (Fig. 6). Generally, grazing intensity is higher in temperate typical grassland and alpine meadow areas compare to other areas (Figs. 1, 6). Specifically, in the eastern part of the Qinghai-Tibetan Plateau, where alpine meadows prevail, and in eastern and southern Inner Mongolia, as well as Xinjiang, where typical grasslands predominate, the grazing intensity usually exceeds 2 SU ha−1 (Fig. 6). However, in other grassland types such as predominant alpine steppes in Tibet and desert grasslands in western Inner Mongolia, the grazing intensity was notably lower, falling below 0.5 SU ha−1 (Fig. 6).

Fig. 6
figure 6

Spatial patterns of grazing intensity in LHGI maps in (a) 1986, (b) 1996, (c) 2006, and (d) 2016.

Full size image

Spatio-temporal patterns of pasture-based and crop-based livestock

During the study period, the surveyed regions experienced an overall upward trend in pasture-based livestock (Fig. 7a), while the proportion of pasture-based livestock to the total livestock exhibited a declining trend (Fig. 7b). From 1980 to 2022, the total number of pasture-based livestock in surveyed provinces increased approximately 1.6 times, from 226 million sheep unit in 1980 to 355 million sheep unit in 2022, with Inner Mongolia showing the largest increase of 2.2 times (Fig. 7a). However, over the same period, the proportion of total pasture-based livestock to total livestock in the surveyed provinces decreased from 70% to 61%, with decrease magnitude ranging from 7% to 34% across all surveyed provinces, most notably in Ningxia (Fig. 7b). Conversely, in Tibet, all livestock relied on grazing throughout the study period (Fig. 7b).

Fig. 7
figure 7

Long-term changes in (a) pasture-based livestock and (b) the proportion of pasture-based livestock to total livestock from 1980 to 2022 of all surveyed provinces.

Full size image

These trends indicate that the proportion of crop-based livestock continued to increase as livestock numbers rose (Fig. 7). We further investigated the spatial and temporal variation of county-level crop-based livestock. The distribution of crop-based livestock closely correlated with the availability of grassland resources, predominantly in counties with scarce grassland resources, such as western and northeastern Inner Mongolia, southwestern Xinjiang, northwestern and southeastern Gansu, and Panzhihua in Sichuan (Fig. 1; Fig. 8). As the total livestock population rose, grassland areas in more regions became insufficient to sustain grazing, prompting a transition to crop-based livestock farming. This transition was evident in numerous counties across Ningxia, southern Gansu, northern Xinjiang, and Tongliao and Chifeng in Inner Mongolia (Fig. 8).

Fig. 8
figure 8

The proportion of crop-based livestock to total livestock by county in (a) 1986, (b) 1996, (c) 2006, and (d) 2016.

Full size image

Comparison with other datasets

We compared the LHGI dataset with four existing datasets: GLW322, GLW422, the gridded grazing dataset produced by Meng et al. (GI-Meng)18, and the actual livestock capacity (ALCC)51. The five datasets exhibited relatively consistent spatial patterns but significantly differed in accuracy (Fig. 9, 10). Since the GLW3 and GLW4 were produced based on the census livestock data in 2001 and 2015, respectively, the accuracy comparison for 2001 and 2015 was conducted between LHGI and the GLW series datasets (Fig. 9). LHGI demonstrated a much higher spatial capture capability for grazing than GLW3 and GLW4. Specifically, validation at the county level illustrated a significant underestimation of grazing intensity by GLW3 and GLW4, with R2 values between estimated and census values of 0.26 and 0.48, respectively (Fig. 9b,d). Conversely, LHGI showed high consistency with the census value, with R2 above 0.8 and linear regression slope above 0.9 (Fig. 9a,c). The GLW series datasets were mapped based on provincial census data and only considered year-end stocks of cattle and sheep21,22. In contrast, LHGI specifically addressed grazing activities within grasslands, resulting in a more accurate spatial distribution compared to GLW3 and GLW4 (Fig. 9).

Fig. 9
figure 9

Comparisons of the grazing intensity of LHGI and GLW series data. GLW3 and GLW4 denote the Gridded Livestock of the World 3 and Gridded Livestock of the World 4, respectively.

Full size image
Fig. 10
figure 10

Comparisons of grazing intensity spatial patterns and accuracy validation for 2010 on Qinghai-Tibetan Plateau. (a) LHGI; (b) GI-Meng; (c) ALCC; (d) the performance of LHGI, GI-Meng, and ALCC.

Full size image

We further compared the performance of LHGI with GI-Meng and ALCC in grazing intensity on the Qinghai-Tibetan Plateau (Fig. 10). Since GI-Meng and ALCC were also mapped based on county-level census data, all counties of the Qinghai-Tibetan Plateau were used to assess the accuracy. LHGI showed the highest consistency between estimated and census values, with an R2 of 0.96 and a slope of linear regression of 0.99 (Fig. 10d). The accuracy of GI-Meng is slightly inferior to LHGI, with an R2 of 0.86 and a slope of 0.9 (Fig. 10d). However, ALCC significantly overestimated the grazing intensity, with a linear regression slope higher than 4 (Fig. 10d).

Furthermore, compared to the other four datasets, the LHGI dataset offers several advantages simultaneously. First, it provides broader spatial coverage, including all pastoral areas in China, addressing the data gap in the two largest pastoral regions of Inner Mongolia and Xinjiang. Second, it uses more precise census data, considering both livestock off-take and offspring numbers, leading to a more accurate estimate of livestock quantities. Third, it distinguishes between pasture-based and crop-based livestock. This differentiation is crucial, especially in semi-agricultural and semi-pastoral areas, where census data conflate these categories, potentially leading to overestimations of grazing intensity if not properly accounted for. Finally, LHGI offers higher spatial and temporal resolutions (Table 1), making it more suitable for regional and long-term research.

Potential implication of the LHGI dataset

The LHGI dataset can serve as a crucial database for grassland carbon budgets. Grazing prevalently reduces plant biomass, plant and animal abundance, and plant and litter cover, significantly affecting ecosystem structure, function, and composition52. Carbon sequestration is one of the most important functions of grassland ecosystems, which can act as either a carbon sink or a carbon source with respect to the atmosphere53. Annual changes in carbon storage within managed grassland ecosystem have been found to be correlated with carbon removed through grazing54. In China, for instance, light to moderate grazing has been associated with increases in plant biomass and soil carbon stocks, while heavy grazing led to declines across all carbon stocks regardless of grassland ecosystem types6. Therefore, grazing intensity is the key to accurately estimate the carbon budget as well as non-CO2 greenhouse gases of grasslands. However, as the world’s second-largest holder of grassland resources, China faces significant limitations in assessing grazing impacts on the carbon cycle in grasslands due to the absence of a long-term high-resolution grazing intensity dataset. LHGI, therefore, has important implications for grassland carbon budgets estimates as grazing largely affects the carbon budget of grassland ecosystems.

Furthermore, LHGI can provide a valuable reference for grazing management. Grazing is a direct and critical indicator of anthropogenic disturbance on grasslands, and rational grazing management is crucial for sustainable utilization of grasslands. As an important carbon reservoir, grasslands store more carbon than forests globally due to their expansive coverage55, and are assumed to be more resilient to higher temperatures than forests56. However, unreasonable grasslands utilization, especially overgrazing, has significantly reduced productivity, leading to continuous degradation. Previous studies suggest that comprehensive implementation of initiatives such as the “Returning Grazing Land to Grassland” Project in China could potentially increase carbon stock in grassland ecosystems by up to 0.21 Pg−1 yr−1 6. Nonetheless, livestock directly contribute to livelihoods and food security4, especially in developing countries and low-income households6. As a crucial foundation for livestock development, complete grazing prohibition of grasslands is not feasible, because it conflicts with the policy objective of increasing food production. Therefore, it is critical to develop rational grazing practices. LHGI can help identify areas suffering from overgrazing and facilitate the development of more effective grasslands management strategies.

Uncertainties

Although the long-term high-resolution annual grazing intensity distribution map produced in this study effectively captured the spatial variability of grazing intensity, there are still several potential uncertainties. First, for certain indicators without census data in specific years and/or provinces, averages from other years were utilized as substitutes. The bias introduced by this interpolating method was quantified according to Eq. 13. Across all surveyed counties, the uncertainty ranged from 0 to 15% (Fig. 11). Over 99% of the counties had uncertainty below 10%, with more than 89% below 5%, and nearly 20% below 1% (Fig. 11).

Fig. 11
figure 11

The county-level uncertainty of census data averaged from 1980 to 2022.

Full size image

Second, in this study, we identified grassland pixels with maximum NDVIg within each 0.1° × 0.1° window as primarily influenced by climate (i.e., no or light grazing) to predict climate-based NDVIg. While this approach increases the likelihood of identifying non-grazed or lightly grazed pixels, if there are no such areas within a given window, the identified NDVIg may be lower than that driven purely by climate, resulting in an underestimation of the climate-based NDVIg. Conversely, the ecological projects, fertilization, and irrigation can artificially elevate NDVIg beyond purely climate-driven levels, potentially leading to overestimation of the climate-based NDVIg. In addition, the accuracy in estimating potential NDVIg during 1980–2000 was relatively lower compared to that during 2001–2022 (Figs. 3, 4). During 1980 to 2000, high spatiotemporal resolution NDVI products were unavailable, and the grazing intensity maps were generated at a resolution of 0.1°. The use of coarse spatial resolution (i.e., 0.1°) could result in mixed pixels of land cover, leading to biases in the extraction of grassland ranges. In addition, the 0.1°spatial resolution NDVI resulted in larger areas for samples searching (Methods), contributing to relatively sparse sample points used for training the RF model, thus reducing model performance.

Third, this study employed the grassland NDVI-biomass model to convert NDVI into biomass. While model validation demonstrated the robust and stable performance, showing comparability with other models, it exhibited limitations in both sparse and dense vegetated areas39. Specifically, this model overestimated biomass in sparse areas while underestimating it in dense areas. As a result, pasture-based livestock and grazing density were overestimated and underestimated correspondingly.

Finally, this study failed to consider disturbances other than grazing that could decrease grassland biomass, such as fire, pollution, and wildlife. While grazing remains the dominant influence, overlooking these other disturbances may result in an overestimation of grazing intensity in areas where these disturbances are present. Strengthening research on these disturbances is crucial for improving the accuracy of grazing intensity identification.

These uncertainties arising from the above three aspects collectively contribute to the uncertainty in the gridded grazing intensity estimation. Additionally, we emphasize the importance of using high spatial resolution data to ensure dataset accuracy.