2.1 Site-Scale Snow Manipulation Experiment

A snow manipulation experiment was conducted at a temperate grassland at the Inner Mongolia Grassland Ecosystem Research Station (IMGERS, 43°38′ N, 116°42′ E, 1200 m above sea level) of the Chinese Academy of Sciences. The region experiences a semiarid continental climate, with a mean annual temperature (MAT) of 0.9°C and mean annual precipitation (MAP) of 334 mm (Chen et al. 2019), approximately 10% of which occurs during wintertime (Li et al. 2020; Zuo et al. 2019). Over the past four decades, mean snow depth has increased across Inner Mongolia (Huang et al. 2016; Peng et al. 2010), including at our study site (Li et al. 2020). The soil at the study site was classified as Calcic Chernozems, with a loamy-sand texture. The grassland plant community was dominated by perennials, including Leymus chinensis (Trin.) Tzvel., Stipa krylovii Roshev., Artemisia frigida Willd. Sp. Pl., and Potentilla chinensis Ser.

The field experiment employed a randomized block design to account for shading, with two treatments: ambient snow and deepened snow (Figure S1a,b). Starting in 2013, a polyethylene mesh snow fence, measuring 1.25 m (height) × 100 m (length), was installed perpendicular to the prevailing northwest wind during wintertime to increase snow accumulation in the deepened snow plots. The snow fence was removed in late March of the following year. Three replicate plots, each measuring 8 × 4 m, were established along the snow fence for each treatment (Figure S1c). A buffer zone (> 20 m) was designed between ambient snow and deepened snow to minimize edge effects caused by the snow fence (Figure S1c).

2.2 Measurement of In Situ N₂O Fluxes

From June 2018 to May 2019, surface N₂O fluxes were measured using an automatic chamber system (Figure S1c,d). This system consisted of six opaque, automated gas flux chambers (28.5 cm diameter, 16.2 cm height; SC-21 N1, LICA, Beijing, China) connected to a multiplexer (SF3500 Soil Greenhouse Gas Flux Monitoring System, LICA, Beijing, China). The multiplexer enabled dynamic chamber deployment and directed gas samples to an off-axis integrated cavity output spectrometer (LGR-913-0055, ABB Inc., Quebec, Canada) for high-precision measurement of greenhouse gas concentrations. The chambers were randomly assigned to the ambient plots (n = 3) and deepened snow plots (n = 3; Figure S1c). Each chamber had a fixed polyvinyl chloride collar (PVC, 19.2 cm inner diameter, 10 cm height) which was inserted 3 cm into the soil. During wintertime, to make the automated chambers work effectively, snow was carefully excavated around the chambers (Figure S1c). This reduced the effective snowpack depth above the chambers, but was necessary to achieve a proper seal and maintain data quality. We acknowledged that this localized disturbance might influence the microenvironment around the chambers. However, due to the flat topography of the site, snowmelt in the soil was well connected, ensuring that soil moisture beneath the chambers remained similar to the surrounding area during the freeze–thaw period. Additionally, as the same protocol was consistently applied across all treatments and plots, comparisons between treatments remained valid.

Where V represents the chamber volume (cm³), and P_av, W_av, and T_av indicate the average air pressure (kPa), average water mole fraction (mmol mol⁻¹), and average air temperature (°C), respectively. R is the ideal gas constant (8.314 J mol⁻¹ K⁻¹), and S is the surface area (cm²) covered by the chamber. dC'/dt (μmol⁻¹ s⁻¹) is the slope of linear regression of dry N₂O concentrations (C′) over time, which are corrected by the water vapor content. M is the molecular mass (g mol⁻¹) of N atoms of N₂O. The annual flux was estimated based on the trapezoidal interpolation between measured fluxes and sampling intervals (e.g., Luo et al. 2022; Matson et al. 2017).

2.3 Regional-Scale Sampling

Twelve study sites were selected along a 1500-km southwest–northeast transect across Inner Mongolia in northern China (Figure S1e). The transect spans longitudes from 107°42′ E to 122°30′ E and latitudes from 38°54′ N to 50°12′ N, with MAT ranging from −2°C to 8°C, and MAP varying from 201 to 488 mm (Fick and Hijmans 2017). The aridity index (AI), defined as the ratio of precipitation to potential evapotranspiration, was obtained from WorldClim2 as well (Fick and Hijmans 2017), and varied from 0.13 to 0.55, with higher values indicating moister soil conditions. Vegetation types along the transect, from west to east, included desert steppe, typical steppe, and meadow steppe. The soils, primarily arid, sandy, brown loessials rich in calcium, were classified as Kastanozem in the FAO classification system (Cheng et al. 2009).

At each site, a 20 × 20 m plot was established, with four 1 × 0.5 m subplots randomly selected and spaced 5 m apart. To assess the role of plant function on the spatial variability of freeze–thaw-induced N₂O emissions and water retention capacity, aboveground biomass (AGB) and belowground biomass (BGB) were measured in each subplot. AGB was measured by harvesting all plant material, while BGB was determined by excavating roots using soil cores (7 cm in diameter and 5 cm in depth). Additionally, 0–5 cm soil samples were randomly collected by a soil corer (3 cm diameter) in each subplot and immediately sieved through a 2-mm sieve. These soil samples were divided into two parts: one was air-dried for physiochemical analysis, and the other was stored at −80°C for DNA extraction. Furthermore, after removal of the aboveground biomass, four intact soil cores (7 cm in diameter and 5 cm in depth) were collected from each subplot, capped at both ends to limit disturbance during transport, and stored at −20°C for a further incubation experiment. The intact soil cores were not sieved or otherwise treated to preserve natural soil structure and minimize potential losses of N₂O emissions. This ensured that gas diffusion and microbial processes remained consistent with in situ conditions, reducing the risk of introducing artifacts for the subsequent experiment.

2.4 Microcosm Freeze–Thaw Experiment

To estimate how changes in winter snow depth could influence freeze–thaw-induced N₂O emissions at the regional scale, we conducted a microcosm incubation experiment using intact soil cores collected from the 12 study sites. For each site, we established four snow depth treatments: 0, 8, 16, and 28 cm, based on historical snow depth data in this region (Li et al. 2020; Ying et al. 2019). Each treatment had four replicates. Due to practical challenges in collecting and preserving natural snow for laboratory use, crushed ice was utilized as a surrogate. Importantly, as snow transitions into ice during the melting process in the field, the use of crushed ice provided a reasonable approximation of snowmelt dynamics. We used the measured average snow water equivalent (0.16 ± 0.03) by Li et al. (2020) at IMGERS in 2017 to convert the four snow depths into corresponding water equivalents, that is, 0, 49, 98, and 172 g. The snow meltwater was then frozen to produce crushed ice to simulate the gradual melting of snow. Hereafter, we referred to crushed ice as “simulated snow.” To accommodate the simulated snow, a PVC tube (6 cm height, 7 cm diameter) was attached to each soil core (Figure S1f), and waterproof tape was used to secure the joint to prevent water leakage during the simulated snow melting. To minimize the artifacts during the freeze–thaw period, we wrapped 2-cm-thick insulation material around the sidewall of each soil core to reduce the extreme temperature fluctuations that may occur (Henry 2007). Each soil core was placed on a support above a 200 mL jar to collect the leachate during thawing. This setup resulted in 192 microcosms (4 treatments × 4 replicates × 12 sites).

Prior to the start of the experiment, the intact soil cores were placed into an incubator at 5°C for 3 weeks to stabilize N₂O fluxes. To better approximate natural conditions, the frequency and duration of freeze–thaw cycles were set based on the previously mentioned snow manipulation experiment at IMGERS (Jia et al. 2022). Three periods were designated (Figure S2): (1) a freezing period, during which soil samples were immediately frozen at −10°C for 1 week after snow application; (2) a subsequent freeze–thaw period, featuring 12 h of daylight at +7°C and 12 h of night at −10°C for 1 week; and (3) a thawing period, during which soil samples were thawed at +7°C for 10 days. The freeze–thaw and thawing periods were repeated, with the duration reduced to 3 and 5 days, respectively. This cycle was designed to mirror the temperature variations observed in the field. Soil temperature at each depth in incubators was recorded using iButton temperature data loggers (DS1922L, Maxim Dallas Semiconductor Corp., USA) during the incubation period (Figure S2).

Where dc/dt is the slope of the linear regression for N₂O concentrations (c, mg m⁻³) over time (t, h). V indicates the volume of the headspace (m³) of glass jar. P and T are atmospheric pressure (Pa) and air temperature (°C) in the glass jar, respectively. R is the ideal gas constant (8.314 J mol⁻¹ K⁻¹). M is the mole mass (g mol⁻¹) of N atoms of N₂O, and A_core is the surface area of the soil core (m²).

2.5 Analysis of Soil Properties

General soil characteristics (pH, total N, SOC, silt, sand, and clay content) in the top 5 cm were determined according to standard methods (Table S1). Briefly, air-dried soil subsamples sieved through a 2-mm sieve were used to measure soil pH at a soil-to-water ratio of 1:2.5. For determination of soil C and N concentrations, 30 mg of soil was ground by ball mill (Retsch MM400, Haan, Germany) and treated with hydrochloric acid (HCl), followed by analysis using an element analyzer (Vario EL III; Elementar Analysensysteme GmbH, Hanau, Germany). Another air-dried soil subsample was fractionated into sand, silt, and clay using the laser diffraction method (Mastersizer 3000, Malvern Panalytical Inc., UK) (Faé et al. 2019). All particle size analysis results were described as a percentage of the oven-dried soil by weight.

After completing the microcosm incubation experiment, we immediately sampled soil cores and sieved the soil through a 2-mm sieve. Each soil sample was split into three subsamples. The gravimetric water content was determined by oven drying the soil sample at 105°C for 24 h and expressed as WFPS using the bulk density for each site and the mineral soil particle density of 2.65 g cm⁻³. Microbial biomass carbon (MBC) and nitrogen (MBN) were determined using the chloroform fumigation–extraction method (Brookes et al. 1985). About 10 g of fresh soil was fumigated for 24 h, followed by extraction with 0.5 M K₂SO₄. MBC and MBN were calculated as the difference in the extractable C and total extractable N between the paired fumigated and unfumigated soils, divided by k_C and k_N factors of 0.45 and 0.68, respectively (Shen et al. 1984). Extractable organic C and extractable N concentrations were measured using a total organic carbon analyzer (TOC-L-CPN, Shimadzu, Japan). Mineral N (NH₄⁺, NO₃⁻) and dissolved organic C (DOC) concentrations were measured by 2 M KCl using continuous flow injection colorimetry (SEAL Analytical AA3, SEAL Analytical GmbH, Norderstedt, Germany) and a total organic carbon analyzer, respectively.

2.6 DNA Extraction and SmartChip Analysis

DNA was extracted from the soil samples using the FastDNA Spin Kit (MP Biomedicals, USA) according to the manufacturer's instructions. The quality of DNA was assessed by ultraviolet absorbance (ND1000, NanoDrop, Thermo Fisher Scientific, USA). DNA concentration was measured using a QuantiFluor dsDNA Kit (Promega, USA).

High-throughput qPCR-based SmartChip technique (Su et al. 2022; Zheng et al. 2018; Zhu et al. 2017) was used to quantify the abundances of denitrifying genes (narG, napA, nirS, nirK, and nosZ) and nitrifying genes (amoA and amoB) in each soil sample. The choice of primers is listed in Table S2. The bacterial 16S rRNA (515 F/907 R) gene was used as the reference. The mixing system contained 3.1 μL of DNA sample, 24.8 μL of mix enzyme (LightCycler SYBR Green I), and 3.1 μL of denitrifying and nitrifying gene primers. The homogenates were transferred into a SmartChip using a MultiSample NanoDispenser and then quantified with a real-time PCR system (WaferGen, Biosystems, USA). The thermal cycle of qPCR conditions was initial denaturation at 95°C (10 min), followed by 40 cycles of denaturation at 95°C (30 s), annealing at 58°C (30 s), and extension at 72°C (30 s). A threshold cycle of less than 31 was used for later analysis. To ensure reproducibility, each qPCR reaction was conducted in three replicates. Relative and absolute gene copy numbers of a denitrifying gene, a nitrifying gene, and the 16S rRNA gene were calculated as follows (Zhu et al. 2017):

Relative gene copy number RG_Fun = $$$ {10}^{\left(31- threshold\ cycle\right)/3.33} $$$,

$$$ \mathrm{Absolute}\ \mathrm{gene}\ \mathrm{copy}\ \mathrm{number}\ {AG}_{Fun}=\frac{AG_{16S}}{RG_{16S}}\times {RG}_{Fun} $$$

Where AG_Fun and RG_Fun indicate absolute and relative functional gene abundance, respectively. AG_16S and RG_16S indicate absolute and relative 16S rRNA gene copy number, respectively.

2.7 Bacterial 16S rRNA Gene Sequencing and Data Processing

We used the primers 338F (5′-ACTCCTACGGGAGGCAGCA-3′) and 806R (5′-GGACTACHVGGGTWTCTAAT-3′) to amplify the variable V3V4 region of the 16S rRNA gene for identification of bacterial communities (Mori et al. 2014). The gene was amplified, purified, quantified, pooled, and sequenced on an Illumina Novaseq6000 platform. For the demultiplexed paired-end 16S sequences, bioinformatic processing was performed using QIIME 2 version 2022.8 (Bolyen et al. 2019). In brief, primers were trimmed using Cutadapt (Martin 2011). Sequences were then denoised, dereplicated, and chimera filtered using DADA2 (Callahan et al. 2016). Reads were trimmed to retain only high-quality sequences based on the quality plots. Subsequently, filtered sequences were grouped into amplicon sequence variants (ASVs). Taxonomy was assigned to the curated ASVs using q2-feature-classifier against the SILVA release 138 reference database trained for the specific gene region (Bokulich et al. 2018; Quast et al. 2013). We removed chloroplasts and mitochondria ASVs, as well as singletons and doubletons. We used the functional annotation tool of the prokaryotic taxa (FAPROTAX) database to analyze the functional groups of bacteria in the soil (Louca et al. 2016). The bacterial ASVs were compared with dataset obtained by FAPROTAX (script version 1.2).

2.8 Fluorometric Enzyme Assays

The activities of six soil hydrolytic enzymes (α-glucosidase, β-glucosidase, β-xylosidase, cellobiohydrolase, leucine aminopeptidase, and N-acetyl-β-D-glucosaminidase) targeting C and N substrates in soils were measured using a modified method described by German et al. (2011). Briefly, fresh soil samples (2 g) were homogenized in 120 mL of 25 mM maleate buffer with pH 6.0 using a hand blender. The homogenates (125 μL) were mixed with 125 μL fluorometric substrate in microplates and incubated in the dark for 4 h at 25°C. Each assay microplate contained homogenate control, substrate control, and standard (4-methylumbelliferone). Following incubation, fluorescence was measured at 365 nm excitation and 450 nm emission by a fluorometer (BioTek Synergy H1 microplate reader, Winooski, VT, USA). Enzyme activities were calculated from these fluorescence values and expressed as the rate of substrate converted in nmol g¹ dry soil h⁻¹. To determine the maximum reaction velocity (V_max), enzyme activities and substrate concentrations were fitted to the Michaelis–Menten model using nonlinear regression analysis. It is noteworthy that the experimental V_max is an apparent kinetic parameter, not the actual value. The apparent V_max indicates the maximum potential enzyme activity and enzyme concentration (Peng et al. 2023). The V_max of organic C- and N-degrading enzymes was expressed as the sum of V_max for α-glucosidase, β-glucosidase, β-xylosidase, and cellobiohydrolase, and the sum of V_max for leucine aminopeptidase activity and β-N-acetylglucosaminidase, respectively.

2.9 Statistical Analysis

Shapiro–Wilk's test and Levene's test were used to assess the normality and homogeneity of variance for N₂O fluxes measured both in situ at the IMGERS site and in the microcosm experiment. Log₁₀ transformations were performed where necessary to meet model assumptions. Linear mixed-effect models were used to assess the differences between/among snow treatments, with snow treatments as fixed effects and sampling day and replicates as random effects (Crawley 2012). Relationships between net N₂O fluxes and tested variables were explored using a linear regression model for each treatment across all sites.

A multimodal inferencing approach (Anderson and Burnham 2002) was used to determine the relative importance of soil and microbial variables in explaining variation in net N₂O fluxes at different simulated snow depths across the 12 study sites. Variables with high collinearity (variance inflation factor, VIF > 5) were excluded through collinearity diagnostics. Multiple generalized linear regression models were generated for all possible combinations of variables using the dredge function in the MuMIn package. The best-supported model for each simulated snow depth was selected based on the lowest Akaike information criterion corrected (AICc) value. The relative importance of each candidate variable in the best-supported model was assessed using proportional marginal variance decomposition (pmvd) through the calc. relimp function in the relaimpo package (Feldman 2005).

Principal component analysis (PCA) was applied to microbial variables to reduce redundancy (Figure S3) and aggregate their effects on net N₂O fluxes. Microbial attributes were expressed as an aggregated value derived from the first principal component (PC1) scores, with higher PC1 scores indicating greater microbial function (i.e., greater microbial biomass, enzyme activities, and functional genes, Figure S4). We then fit predictions of net N₂O fluxes using WFPS and the new aggregated microbial attributes (PC1 score) as predictors. To evaluate the model's term, we performed model comparisons (Table S3) by separately fitting (i) a linear model with linear interaction terms for the two predictors and (ii) a nonlinear model with the same linear predictors plus a quadratic term for WFPS (WFPS²). The nonlinear model for net N₂O fluxes had a lower AIC value (Table S3), indicating better model performance and supporting the inclusion of quadratic WFPS (WFPS²). To integrate the effects of WFPS and microbial attributes, we used the fitted models to predict net N₂O fluxes (Table S4). Two-dimensional contour plots were used to visualize how freeze–thaw-induced N₂O fluxes varied simultaneously with WFPS and microbial attributes, with observed net N₂O fluxes overlaid.

Model performance was evaluated by comparing modeled and observed net N₂O fluxes, assessing their agreement by goodness-of-fit measures. The modeled values and observed net N₂O fluxes in the microcosm experiment are close to the 1:1 reference line (Figure S5a). The goodness of fit was then summarized by linear regression. The modeled and observed net N₂O fluxes are highly correlated (R² = 0.68, root mean square error = 0.34, Figure S5a). This evaluation verifies that the model can perform well for data prediction. Residual diagnostics were conducted to validate model assumptions. The normality assumption of the model residuals is verified by Figure S5b, supporting log₁₀-transformed net N₂O fluxes. Then we used a broken-line regression model (segmented method) in the segmented package (Muggeo 2008) to identify possible thresholds of WFPS, and thus visualize the significant interaction effects of WFPS and microbial attributes on net N₂O fluxes.

Redundancy analysis (RDA) was conducted using the vegan package to evaluate the effects of climate, soil properties, and plant on microbial attributes among the three types of steppes and their individual effects based on the rdacca.hp package (Lai et al. 2022). All statistical analyses were conducted in R 4.3.0, and significance was considered at p < 0.05.

3.1 Snow Depth Drives Freeze–Thaw-Induced N₂O Emissions

Our high-frequency in situ measurements showed that deeper snow significantly increased N₂O emissions compared to ambient snow throughout the year (p < 0.03; Figure 1a,b). Specifically, deeper snow increased freeze–thaw-induced N₂O emissions (p = 0.008; Figure S6). The maximum pulse of N₂O emissions under deeper snow (252.4 μg N m⁻² h⁻¹) was nearly ninefold of that in ambient snow (29.5 μg N m⁻² h⁻¹; Figure S6). Remarkably, N₂O emissions during the brief 46-day freeze–thaw period accounted for 45% and 57% of annual fluxes in deepened and ambient snow, respectively (Figure 1b).

Across all soil collected, freeze–thaw-induced N₂O emissions nonlinearly increased with simulated snow depths (p < 0.02; Figure 1c; Figure S7a). Notably, significant spatial variability was observed between sampling sites in N₂O emissions with simulated snow depth, ranging from 3.4 ± 0.2 to 1184.1 ± 243.9 μg N m⁻² h⁻¹ (Figure 1d). Emissions were lowest in arid steppes (3.7 ± 0.1~37.6 ± 4.7 μg N m⁻² h⁻¹), followed by typical steppes (5.7 ± 0.5~94.9 ± 29.4 μg N m⁻² h⁻¹), and were highest in meadow steppes (20.1 ± 10.3~302.2 ± 67.5 μg N m⁻² h⁻¹) across all simulated snow depths (p < 0.03; Figure S7b). A similar pattern with N₂O emissions was observed between WFPS and simulated snow depths (p < 0.03; Figure S8).

3.2 Controls of Ex Situ Net N₂O Fluxes

At each simulated snow depth, net N₂O fluxes were negatively correlated with MAT, pH, and sand content (Figure 2a,c,f). In contrast, positive correlations were observed with AI, SOC, BGB, WFPS, and C and N availabilities (Figure 2b,d,e,g–i), with particularly strong correlations between net N₂O fluxes and WFPS. Substrate availability and V_max of organic C- and N-degrading enzymes were increased with WFPS (Figure S9).

Net N₂O fluxes were positively correlated with MBC and MBN, V_max of organic C- and N-degrading related enzymes, amoB gene abundance of ammonia-oxidizing bacteria involved in nitrification, and the ratio of N₂O production to reduction genes (ProRed: N₂O production genes consisted of narG- and napA-encoding nitrate reductase, and the N₂O reduction gene consisted of nosZ-encoding N₂O reductase), although the strength of these correlations varied across simulated snow depths (Figure 3a–e).

3.3 Relative Importance of Controls of Ex Situ Net N₂O Fluxes

Under snow-free conditions, WFPS emerged as the overwhelmingly important predictor, with NO₃⁻-N, SOC, ProRed, and BGB collectively explaining 86% of total variation in net N₂O flux (Figure 4a–c). Under shallow snow conditions, soil and microbial factors together explained 68% of their total variation, of which abiotic factors accounted for over 50% (Figure 4b,d). In contrast, microbial factors, particularly V_max of N-related enzymes and bacterial amoB gene abundance, dominated N₂O emissions at 16 and 28 cm of simulated snow depth (Figure 4b,e,f). Notably, the contribution of soil physiochemical properties to the explained variance decreased from 80% to 24% with the increase in simulated snow depth (Figure 4b). Specifically, the importance of WFPS diminished sharply from 65% to being excluded from the model, while microbial predictors rose from 17% to 60% (Figure 4b).

Two thresholds for WFPS, 43% and 66%, were identified as key transitions for N₂O emissions (Figure 5a). A significant interaction was observed between WFPS and microbial attributes (Figure 5b,c), aggregated values of microbial predictors based on PCA (Figure S3). Specifically, in Phase 1 (WFPS < 43%), soil moisture was the overwhelming factor triggering N₂O emission. In Phase 2 (43% ≤ WFPS ≤ 66%), both microbial attributes and soil moisture jointly control emissions (Phase 2). Finally, in Phase 3 (WFPS > 66%), microbial attributes acted as the sole constraints on N₂O emissions (Figure 5a,b,d).

3.4 Factors Driving Hotspots of N₂O Emission Across Ecoregions

Given the joint control of N₂O pulses by WFPS and microbial attributes, we further analyzed the factors shaping their spatial patterns to identify potential N₂O hotspots. In addition to simulated snow depths, WFPS was strongly influenced by BGB and SOC (Figure 6a). The spatial heterogeneity of microbial attributes was mainly shaped by MAT, pH, BGB, SOC, and sand content (Figure 6b; Table S5), collectively accounting for 35% of the variance. Thus, in cold and humid grassland ecosystems with relatively low pH, high root biomass and microbial activity promoted the formation of hotspots for freeze–thaw-induced emissions, where N₂O pulses were 1–2 orders of magnitude higher than those in arid steppes.

4.1 Controls of Freeze–Thaw N₂O Emission Hot Moments

Freeze–thaw N₂O pulses are driven by both the physical release of trapped N₂O and microbial processes (Risk et al. 2013). While we could not directly quantify the relative importance of the two concurrent processes, previous studies suggest that physical release is less significant than microbial production during freeze–thaw period (Risk et al. 2014; Teepe et al. 2001). This was evidenced by our findings that N₂O emissions were significantly accelerated with simulated snow depths (Figures 1c and 5) despite similar soil temperatures during the freeze period among these treatments (Figure S2). When the soil is dry, the ice crystals form without disrupting the soil matrix (Bullock et al. 1988), potentially trapping N₂O due to the limited gas diffusion; in contrast, the presence of snowmelt water can facilitate the freeze–thaw expansion and contraction of ice, destabilizing soil aggregates and releasing nutrients from plant and microbial detritus within them (Bullock et al. 1988; Ruan and Robertson 2017). This could improve nutrient accessibility for nitrifiers and denitrifiers, thereby driving de novo N₂O production (Wang et al. 2023). Thus, while physical release occurs during initial thawing, microbial processes, which remain active at low temperatures, sustain N₂O production, particularly during the thaw phase when conditions favor denitrification (Risk et al. 2013).

Additionally, the soil's physicochemical structure, particularly the arrangement of organic residues within soil pores, could play a critical role in triggering N₂O emission “hot moments” during freeze–thaw cycles (Groffman et al. 2009; Kim et al. 2022; Kravchenko et al. 2017). As snowmelt is absorbed by organic residues such as plant litter or organic particles, localized anaerobic hotspots conducive to denitrification are created (Kim et al. 2022; Kravchenko et al. 2017). High moisture levels within organic residues not only facilitated microbial metabolisms, leading to oxygen consumption, but also restricted oxygen inflow. Consequently, organic residues served as hubs for multiple processes (i.e., decomposition, nitrification, and denitrification), triggering substantial N₂O emissions (Butterbach-Bahl et al. 2013). Indeed, our results showed that freeze–thaw-induced N₂O emissions increased significantly with SOC and belowground plant C input (Figure 2d,e). Additionally, as simulated snow depth increased, WFPS also rose (Figure S8), facilitating the diffusion of substrates to surviving microbes within soil pores (Sorensen et al. 2016). This rise in WFPS could reactivate dormant microbes and enhance enzyme activities related to C and N cycling (Figure S9), thereby promoting greater N₂O emissions.

Furthermore, we found that changes in the proportion of producers (narG/napA) and consumers (nosZ) of N₂O played a critical role in determining N₂O emissions at all simulated snow depths (Figure 3a,e). This emphasized the critical role of nitrate-reduction denitrifiers, instead of nitrite-reduction denitrifiers (nirS- and nirK-type denitrifiers), in regulating N₂O emissions during freeze–thaw cycles in arid and semiarid grasslands. While most molecular studies of bacterial denitrification focus on nitrite reductase (Bothe et al. 2000), which catalyzes the first key step that results in the formation of a gaseous intermediate (Groffman et al. 2006), the role of nitrate reductase should not be ignored, especially in arid regions (Krichels et al. 2023). NarG denitrifiers can survive and remain active in arid environments, potentially favoring NO₃⁻ reduction to N₂O upon wetting (Krichels et al. 2023). Under oxygen-limited conditions, NO₃⁻ reduction may serve as an alternative pathway for energy conservation (Wang et al. 2023). Collectively, the development of anaerobic microenvironments, combined with moisture-induced variation in soil C and N substrates, enzyme activities, and specific soil functional genes, synergistically influenced freeze–thaw N₂O pulses.

4.2 Interactive Effects of WFPS and Microbes on Freeze–Thaw N₂O Emissions

Our analysis revealed that abiotic and biotic factors interacted hierarchically to regulate freeze–thaw N₂O emissions, with their relative importance shifting across different conditions (Figure 4), supporting our second hypothesis. This shift in control patterns was strongly dependent on WFPS thresholds (Figure 5), aligning with a global study that emphasizes moisture thresholds as pivotal in determining ecological processes (Zhang et al. 2023). Below 43% WFPS (Phase 1), typical of snow-free soil conditions or soils with low capacity to retain snowmelt, WFPS was the most decisive environmental trigger for freeze–thaw N₂O emissions (Figure 4a). Low soil water content enhances oxygen availability in soil (Wei et al. 2023), making the conditions unfavorable for denitrifying activities and anaerobic denitrification (Wu et al. 2014; Yao et al. 2010), consequently constraining N₂O emissions.

As WFPS continued to increase to 43%–66% with greater simulated snow depths (Phase 2), we noticed an abrupt decline in the importance of WFPS, whereas the role of microbial properties became more pronounced (Figures 4 and 5), suggesting that once anaerobic conditions were established, soil microbial properties became paramount. At WFPS levels exceeding 66% (Phase 3), close to the optimal 70%–80% range for denitrification (Davidson et al. 2000), microbial control became dominant. These high soil moisture conditions created strongly anaerobic conditions with low gas diffusivity, facilitating N₂O reduction to N₂ (Davidson et al. 2000). In the two phases, soils with enhanced microbial N cycling functions, particularly those with higher N-cycling enzyme activities (Figure S9), greatly influenced the magnitude of the N₂O pulses (Figure 4e,f). These findings emphasize the complex interplay between soil moisture and microbial attributes in regulating N₂O. Specifically, maintaining higher soil moisture levels was a prerequisite for controlling soil microbial activities and triggering freeze–thaw N₂O emission “hot moments,” a pattern also reflected in the Cannon model proposed by Zhang et al. (2024).

4.3 Characteristics of Hotspots for Freeze–Thaw-Induced N₂O Emissions

The climate and soil properties determined plant productions as well as microbial properties (Figure 6b), together controlling the spatial patterns of N₂O emissions during the freeze–thaw period across different ecoregions. Compared to warm and dry steppes, cold and humid meadows harbored higher microbial biomass, N enzyme kinetics, and greater abundance of denitrifying genes (Figure S4), all of which indicated enhanced microbial N cycling. In these regions, higher root production not only enhances root C inputs (Phillips et al. 2011) but also increases WFPS by promoting soil porosity through root growth and soil aggregation (Figure 6a). This combination of high WFPS and greater plant C input (Figure 6b) facilitates the formation of plant residue-induced microscale hotspots of N₂O production and emissions (Kravchenko et al. 2017). Consequently, these microscale hotspots aggregated into large-scale hotspots, contributing to 4 to 21-fold greater freeze–thaw-induced N₂O emissions in cold and humid grasslands than warm and dry steppes, which supported our third hypothesis. In addition, deepened snow also promotes aboveground biomass production (Deng et al. 2023), which can accumulate a thick snow cover (Essery and Pomeroy 2004; Wolf et al. 2010), further enhancing N₂O emissions during the freeze–thaw period.

Soil properties, particularly soil pH and texture, were also key modifiers for spatial variability of freeze–thaw N₂O emissions (Figure 2c,f). The relationship between N₂O and pH is multifaceted, as pH regulates both N-related enzyme activity (Sinsabaugh et al. 2008) and denitrifier community composition (Liu et al. 2014; Qiu et al. 2024). Consistent with the global study by Sinsabaugh et al. (2008), our results showed that N-related enzyme activity decreased with increasing pH from 5 to 9 (Figure 6b). This suggested that low pH enhanced N mineralization by increasing enzyme activities, thereby improving substrate availability and promoting N₂O emissions. Furthermore, pH can indirectly affect N₂O emissions by shaping the denitrifier community (Qiu et al. 2024), thus regulating the spatial variability of freeze–thaw N₂O emissions.

Soil texture determines SOC formation and water-holding capacity (Rawls et al. 2003), thereby shaping N₂O emissions (Figure 6). As noted by Kravchenko et al. (2017), the interaction between soil water and organic residues within soil pores is a critical driver for the occurrence and activities of N₂O hotspots (Figure 6b). Specifically, sandy soils in our study area, typically found in high-pH, warm regions, have large soil pores that result in rapid drainage and limited substrate retention, constraining microbial processes and reducing N₂O production. By contrast, clay-rich soils, generally found in low pH and cold regions, exhibited enhanced moisture retention, accumulation of organic residue, and favorable enzymatic conditions, fostering an environment conducive to the formation of N₂O emission hotspots (Kim et al. 2022; Kravchenko et al. 2017). Therefore, these textural and climatic interactions jointly shaped the spatial variability of freeze–thaw N₂O emissions under climate change.

4.4 Implications for N₂O Modeling and Emission Accounting

Understanding the mechanisms that control the occurrence and the magnitude of freeze–thaw induced N₂O “hot moments” is crucial for reducing uncertainty in N₂O accounting (Butterbach-Bahl and Wolf 2017; Wagner-Riddle et al. 2017). To our knowledge, this is the first study to combine in situ high frequency N₂O automated measurements with cross-ecoregion soil core incubations to explore the effects of snow regime shift on freeze–thaw-induced N₂O emissions. Our findings revealed a shift in the controlling factors for N₂O emissions, from WFPS, which governed the establishment of anaerobic conditions, to microbial constraints with increasing simulated snow depth. The hierarchical control exerted by WFPS and microbial properties explained much of the high temporal and spatial variability of N₂O fluxes during freeze–thaw periods (Figure 5).

Building on this framework, we highlighted the potential to refine the parameterization in process-based models, enabling them to better capture the dynamic responses of freeze–thaw-induced N₂O fluxes to changing environmental conditions. For example, integrating WFPS thresholds to categorize N₂O emission drivers at different WFPS levels into model algorithms could enhance their predictive accuracy. Since WFPS is relatively easy to measure, and key microbial attributes driving N₂O emissions can be estimated using climate and soil physicochemical properties, these insights could streamline the identification of N₂O hotspots during freeze–thaw periods. Incorporating the hierarchical interactions between WFPS and microbial properties into predictive models would provide better tools for accurate N₂O accounting and management of N₂O emissions in response to climate change.

While our study provides valuable insights into freeze–thaw N₂O pulses, several limitations must be acknowledged. Controlled simulations with standardized temperature fluctuations and truncated snow cover duration may introduce uncertainties when extrapolating findings to natural field conditions. Additionally, the limited sampling sites may restrict the generalization of findings across broader climatic and ecological gradients. Future studies should focus on multisite experiments that simultaneously monitor key variables (e.g., soil temperature, moisture, and microbial community) to better capture the complexity of freeze–thaw processes. Moreover, advanced techniques are needed to overcome the challenges of field monitoring during freeze–thaw periods. High-frequency automated chambers, coupled with isotopic tracing and integrated sensor networks measuring soil thermal and hydric conditions, can offer a comprehensive, multidimensional dataset, improving both N₂O emission quantification and our understanding of the underlying mechanisms. This is crucial for improving N₂O accounting, especially given that more than 50% of the continental land areas in the Northern Hemisphere experience freeze–thaw cycling (Chen et al. 2022).