Methane (CH₄) uptake in alpine ecosystems is an important component of the global CH₄ sink. However, large uncertainties remain regarding the magnitude and spatial patterns of CH₄ uptake, owing to its extensive spatial variability, diverse controlling factors, and limited regional-scale observations. Here, we investigated field ecosystem CH₄ uptake along a 3,200-km transect across various alpine grasslands on the Qinghai-Tibetan Plateau (QTP). We found a substantial spatial variation in in-situ CH₄ uptake among alpine grasslands, with the highest rates in drier regions of the mid-western QTP. Soil moisture was the most important factor controlling CH₄ uptake, exhibiting a remarkably low threshold of 6.2 ± 0.1 v/v %. Below this threshold, CH₄ uptake was constrained by soil moisture, moisture-induced nitrogen limitation, and high temperatures. Above this threshold, CH₄ uptake was mainly limited by gas diffusion and low temperatures. By integrating grid predictors with a random forest model trained on 1,851 field measurements encompassing both our observations and a regional synthesis across the QTP, we estimated a regional CH₄ uptake of 0.88 ± 0.020 Tg CH₄ yr^-1 from all alpine grasslands on the QTP. This higher estimate, primarily driven by alpine steppes, was significantly greater than current regional estimates from global CH₄ models. Our findings highlight the importance of the CH₄ sink in dry alpine ecosystems characterized by low soil moisture, suggesting that the contribution of the CH₄ sink in drylands may have been substantially underestimated in the current global CH₄ budget.

Study area and sampling sites

This study was conducted on the Qinghai-Tibetan Plateau (QTP), the highest and largest plateau on the Earth, with an average altitude of over 4,000 m above sea level and a broad area of 2.5 × 10⁶ km^2 . During July-August 2020, 58 sites were visited once across the QTP along a precipitation transect (mean annual precipitation (MAP): 56−712 mm yr ^-1, mean annual temperature (MAT): −7−5℃) of 3,200 km from northwest to southeast, covering all alpine grassland types. The grassland type at each site was defined according to the Chinese Vegetation map and through observations of the vegetation composition in the field.

Field measurements and sampling

To monitor in-situ CH₄ fluxes, we randomly selected three to six 1 m × 1 m quadrats within a 10 m × 10 m plot at each site. CH₄ fluxes were measured for 3−6 replicates within each quadrat using an opaque chamber during the daytime between 9:00 a.m. and 12:00 a.m. Specifically, in each quadrat, a cylindrical polyvinyl chloride (PVC) collar (20 cm diameter, 10 cm height) was inserted 2 to 3 cm into the soil and connected to an opaque chamber (covering area: 276.27 cm², height: 44.5 cm). CH₄ concentrations in the chamber were measured over 3 minutes using a Los Gatos Research, Inc. (LGR) ultraportable greenhouse gas analyzer (MGGA, ABB, Canada) while the chamber was closed and formed a closed loop to the gas analyzer. A small fan was fixed inside the chamber to homogenize the inside air. Fluxes of CH₄ were computed by fitting a linear regression model to temporal changes in the chamber CH₄ concentration during the measurement period of 3 minutes, with regression significance at p < 0.05, using the real-time measured chamber volume (which changed with the depth of the collar inserted into soils), air pressure, and temperature. In total, 1233 in-situ flux measurements were obtained in the field. An averaged flux was calculated for each quadrat, resulting in 290 CH₄ fluxes for further data analysis. Soil temperature (ST, °C) and volumetric soil moisture (SM, v/v %) at a depth of 0−7.5 cm were simultaneously recorded adjacent to the PVC collar for each CH₄ flux measurement using a TDR 350 (FieldScout, Spectrum Technologies Inc., USA) with a 7.5 cm probe rod vertically inserted into the soil.

To investigate vegetation characteristics, three out of the 3−6 quadrats within each site were randomly selected at all sites. Plant coverage (PC) and species were estimated in the field, followed by ground-level harvest of plant tissues. Plant roots were randomly collected from the same quadrats at depths of 0−30 cm, where most roots are distributed. Three root cores (7.5 cm diameter) were taken and mixed in each quadrat to measure root biomass. In the laboratory, all plant samples (with roots washed beforehand) were oven-dried at 65°C for 48 hours and weighed to measure plant biomass (PB, which includes above-ground biomass and below-ground biomass at 0−30 cm soil depth). Plant richness (PR) was represented as the number of plant species in each quadrat. Plant coverage was estimated by the projection method.

In the field, soil samples were collected from 0−10 cm depth using a 7.5 cm diameter auger by randomly combining five individual soil cores at each of three selected quadrats. The mixed soil sample was then sieved with a 2-mm mesh after hand picking out rocks and roots and separated into two subsamples, one stored at −20°C for molecular analysis, and the other at 4°C for laboratory incubation and physiochemical analysis. Moreover, to measure soil bulk density (BD), an additional soil sample was collected from the top 10 cm soil depth at each of the three selected quadrats using a soil density ring (100 cm³, 5 cm diameter).

Laboratory incubation

Following the field survey and sampling, soil samples for incubation were immediately transported to the laboratory and stored in 4°C fridges. To assess soil CH₄ oxidation potential, 10 g of fresh soil for each soil sample was weighed into 150 mL glass bottles and pre-incubated for 7 days at 5°C with the field SM intact. After pre-incubation, all bottles were flushed with atmospheric air (_{1.9 ppm CH}4_{) for 5 minutes, and sealed with rubber stoppers to create gas-tight conditions. At the beginning of the incubation, 30 ml of background atmospheric air was injected into each bottle to prevent under-pressure. Then, the soils were incubated in the dark at 5°C for an additional 48 hours, during which a 6 ml gas sample was collected from the headspace every 12 hours using a gas-tight syringe. The CH}4~ concentration for all gas samples was analyzed using a gas chromatograph equipped with a flame ionization detector (GC-7890B, Agilent, Santa Clara, CA, USA) within 24 hours after sampling. Gas fluxes were calculated using a second-order polynomial regression model to assess the changes in CH₄ concentration in the bottle over 48 hours, with the significance of the regressions evaluated at p < 0.05. The slope at 10 hours of the fitted line was applied to calculate the CH₄ oxidation rate. Considering experimental costs, we selected 22 representative sites that cover all types of alpine grasslands, with each located approximately 150 km apart, for the laboratory incubation.

**Soil parameters **

Soil BD samples were oven-dried to a constant weight at 105°C, weighed, and then the ratio of oven-dried soil mass to mental ring volume (100 cm^³) was calculated to represent BD. Soil texture (contents of clay, silt, and sand) was evaluated using a Malvern Mastersizer 2000 (Malvern, UK) after removing soil organic matter and carbonates using 35% H₂O₂ and 10% HCl, respectively. Soil pH was measured in a 1:2.5 soil-water suspension using a pH meter (HM-30G, TOA Co., Japan). Contents of soil total carbon (TC) and total nitrogen (TN) were determined using a Vario EL cube elemental analyzer (Thermo Scientific, Bremen, Germany). Soil inorganic carbon (SIC) contents were quantified using a solid infrared carbon and sulfur analyzer (multi EA4000, Analytic-Jena, Germany). Soil organic carbon (SOC) was the difference between TC and SIC. Soil C/N was the ratio of SOC to TN. Soil available nitrogen (AN) was analyzed by the alkaline-hydrolysis diffusion method. Soil available phosphorus (AP) was determined by measuring phosphate using the ammonium molybdate spectrophotometric method following extraction with 0.03 M ammonium fluoride and 0.025 M hydrochloric acid.

In addition, absolute abundances of methanotrophs were quantified by qPCR (quantitative Polymerase Chain Reaction) on the pmoA gene (particulate methane monooxygenase gene, the functional marker for detecting aerobic methanotrophs) using primer sets of A189f (5′-GGNGACTGGGACTTCTGG-3′) and mb661r (5′-CCGGMGCAACGT-CYTTACC-3′) according to the method used in our previous study. Again, to consider experimental costs, BD, soil texture, SIC, and pmoA abundance analysis were conducted only on soil samples of the selected 22 representative sites.

Regional synthesis and estimates from existing global CH₄ models

To obtain more data representing both temporal (diurnal and seasonal) and spatial variations of CH₄ uptake for model training and the estimation of regional CH₄ uptake, we conducted a regional synthesis of in-situ CH₄ flux across all uplands from QTP. We searched the Web of Science (http:// isiknowledge.com), Google Scholar (https://scholar.google.com), and the China National Knowledge Infrastructure (http://www.cnki.net) for published studies of CH₄ fluxes from alpine upland grasslands on the QTP, using the keywords ‘CH₄’ or ‘methane’ and ‘Tibetan Plateau’. Data were included only when chamber-measured in-situ CH₄ flux and the simultaneously recorded SM and ST (0−10 cm soil depth) were all available. Overall, we compiled a database comprising 1,561 observations from 21 publications, distributed among 20 sites, encompassing in situ CH₄ uptake throughout different seasons from 2006 to 2021 (Supplementary Data). The regional synthesis dataset also considered the diurnal dynamics by including the daily-averaged in-situ CH₄ flux at the annual scale from an alpine meadow at Hongyuan station (32°48' N, 102°58' E), where the CH₄ flux was measured by auto-chambers with hourly measurement frequency. Combining the regional synthesis and our transect observations, the whole dataset exhibited ST ranging from −18.0°C to 45.1°C, effectively representing the temperature dynamics on the QTP. We also recorded corresponding information from the literature on vegetation types, location (longitude and latitude), MAP (mm), MAT (°C), AN (mg kg^-1), and measurement dates.

Furthermore, we retrieved existing regional CH₄ uptake estimates on the QTP from literature that evaluated regional CH₄ and global CH₄ uptake models.^^ For global CH₄ uptake models, global maps of CH₄ uptake were first clipped using the QTP boundary. The regional CH₄ uptake and CH₄ uptake for each vegetation type were then calculated using zonal statistics in QGIS (3.20.0-Odense), based on the vegetation map of QTP.

Gridded climate, vegetation, and soil datasets

Various remote-sensing datasets were gathered and analyzed using Google Earth Engine (GEE) to obtain climate (MAP and MAT) and plant data (Enhanced Vegetation Index (EVI)) for site information and regional predictions. To derive the climatic and plant information for each sampling location, we calculated MAT and MAP using ECMWF Reanalysis v5 (ERA5) air temperature and precipitation data from 1981 to 2020. We then reanalyzed the mean air temperature during the growing degree days (MTGDD3, defined as days with mean air temperature > 3°C) and freezing degree days (MTFDD, defined as days with mean air temperature < 0°C) using the ERA5 air temperature dataset for the same period. Using images with the date closest to the field observation date (0−7 days before or after the field observation date), we obtained EVI^^ for all field observations collected along our transect and from the literature, using a dataset of 16-day Moderate Resolution Imaging Spectroradiometer (MODIS) EVI.

To conduct the regional estimation of CH₄ uptake, datasets with high spatial and temporal resolution spanning a long timescale (2001−2021) were selected and reanalyzed as predictor datasets. Specifically, the SM input was chosen from a long-term, high temporal resolution dataset: surface SM with a spatial resolution of 0.25° (SMsurf, 7 cm soil depth) from the Global Land Evaporation Amsterdam Model (GLEAM). This dataset showed a good relationship with the in-situ surface SM from both our transect and regional synthesis. The ST input was a 1-km dataset from 2001 to 2021 converted from the 1-km MODIS Terra Land Surface Temperature (LST)^ ^using the linear relationship between our field-observed ST and LST-obtained surface temperature. For precipitation, we chose a 1-km monthly precipitation dataset for China. The EVI dataset was the 1-km MODIS EVI. The available nitrogen (AN) dataset was derived from a published China soil dataset of 1-km spatial resolution. Due to the varying time scales, all datasets were monthly resampled before being imported into the model. The SM dataset was further resampled to the 1-km resolution to ensure that all predictors were at the same spatial resolution, and units of each dataset were transformed to remain consistent with the observed dataset.

Model predictions

To estimate regional CH₄ uptake, 1-km grid CH₄ uptake rates on the QTP were calculated monthly using the random forest model and predictor datasets from 2001 to 2021. First, a random forest model was developed based on our transect data using the randomForest R package. Second, to avoid overfitting, random forest auto-selection was employed to simplify the model using the rfUtilities R package. Five variables, including climate, vegetation, and soil condition, were selected: MAP, EVI, ST, SM, and AN. Third, considering the geographical and temporal dynamics of CH₄ uptake across the QTP, we rebuilt the random forest model using the selected variables based on the database encompassing both our transect observations and the regional synthesis. The “leave one out” cross-validation was used to evaluate the prediction performance of the model. We calculated a bias as an average of the absolute error between predictions and actual observations, used Pearson correlation (R) to identify the strength of the linear relationship between observed and predicted fluxes, and calculated root mean squared error (RMSE) to estimate model errors. To assess potential prediction bias introduced by differences (e.g., slight variations in soil moisture measurement depths: 0−7.5 cm for transect vs 0−10 cm for regional synthesis) between our transect observations and the regional synthesis dataset, we compared the predicted CH₄ fluxes for our transect sites using the simplified model trained exclusively on our transect dataset and the rebuilt model trained on both our transect and the regional synthesis datasets. Finally, the rebuilt random forest model was applied to the predictor datasets, producing monthly step grid CH₄ uptake rates at 1-km resolution from 2001 to 2021.

To quantify the uncertainty in our random forest model and the corresponding CH₄ uptake, we employed a repeated random resampling procedure. Through bootstrapping, we generated 200 datasets from the original flux data, each containing an identical number of observations. These 200 datasets were then utilized to produce 200 individual predictions using the random forest model, and the prediction uncertainty was quantified using the prediction interval (PI; 95th percentile–5th percentile). Subsequently, the uncertainty of the estimated annual CH₄ uptake was determined by calculating the range of annual CH₄ uptake across the 200 predictions.

Using the monthly predictions, we then calculated the grid annual mean and total CH₄ uptake as well as the uncertainties for each year. Vegetation-type-based statistics were then derived using zonal statistics in QGIS (version 3.20.0-Odense), based on the vegetation map. Additionally, the differences between our estimation and current global CH₄ models for uplands on the QTP and for each vegetation type were assessed using the same vegetation map.