Clockwise or Counterclockwise: Hysteresis in the Response of Ecosystem Respiration to Temperature

ABSTRACT

Ecosystem respiration (ER) often responds asymmetrically to seasonal warming and cooling, producing thermal hysteresis. Yet, its global patterns and drivers remain unknown. Here, we used daily ER data from 324 eddy‐covariance sites worldwide and revealed two contrasting hysteresis patterns in the temperature response of ER: clockwise hysteresis (ER is higher during the warming than cooling season) at 173 sites and counterclockwise hysteresis (ER is higher during the cooling than warming season) at 151 sites. The direction and magnitude of ER hysteresis were largely governed by the seasonal timing mismatch between precipitation and vegetation‐derived substrate availability. Higher substrate availability during the warming season enhanced ER and promoted clockwise hysteresis, whereas moisture limitation suppressed ER during the warming season and caused counterclockwise hysteresis. These findings provide robust evidence that ER is a complex process responding to the availability of carbon substrate and regulated differentially by seasonal temperature and precipitation. Our results highlight that the hysteresis of terrestrial carbon release is a key aspect of the global carbon cycle, with important implications for understanding and projecting land carbon‐climate feedbacks.

Ecosystem respiration (ER) exhibits pronounced thermal hysteresis, responding asymmetrically to seasonal warming and cooling. Two distinct patterns emerge: clockwise hysteresis (higher ER during warming than cooling) and counterclockwise hysteresis (higher ER during cooling than warming). The direction and magnitude of this hysteresis are primarily driven by seasonal timing mismatches between precipitation and substrate availability. Elevated substrate supply during the warming season enhances ER, inducing clockwise hysteresis, while moisture limitation during the warming season suppresses ER, leading to counterclockwise hysteresis. These findings underscore that the timing of ecosystem processes is a critical driver of carbon‐cycle hysteresis and future climate‐carbon feedbacks.

1. Introduction

Ecosystem respiration (ER) is the second largest carbon flux between the biosphere and atmosphere, playing a major role in the global carbon cycle and contributing to the rise of atmospheric carbon dioxide (CO₂) concentrations (Heimann and Reichstein 2008; Migliavacca et al. 2021). Temperature is widely recognized as a primary driver of ER rates (Luo and Zhou 2006; Yvon‐Durocher et al. 2012; Johnston et al. 2021). Substantial evidence indicates that rising temperature can strongly stimulate ER through increasing autotrophic respiration (from plant roots and rhizosphere microorganisms) (Bond‐Lamberty and Thomson 2010; Maes et al. 2024) and heterotrophic respiration (from microbial decomposition of litter and soil organic matter) when soil moisture availability is optimal (Davidson and Janssens 2006; Karhu et al. 2014). This may trigger positive carbon‐climate feedbacks and accelerate planetary warming (Chen et al. 2023; Maes et al. 2024). Given that a substantial portion of Earth's carbon is stored in terrestrial ecosystems, understanding the dynamics of the temperature response of ER is critical to accurately quantify the land carbon‐climate feedbacks in the context of aggravating climate warming (Mahecha et al. 2010; Niu et al. 2024).

In most terrestrial ecosystem models, the response of ER to both temperature increase (warming) and decrease (cooling) is represented as the same exponential function (Heimann and Reichstein 2008; Bond‐Lamberty and Thomson 2010). However, growing evidence suggests that the temperature response of ER cannot be fully characterized by a simple function (Carey et al. 2016; Chen et al. 2023; Niu et al. 2024), as ER and temperature often exhibit pronounced hysteresis (Zhang et al. 2018; Ataka et al. 2020; Chu et al. 2023). Specifically, if ER increases faster during the warming than the cooling season, this will lead to a clockwise hysteresis (Ataka et al. 2020; Chu et al. 2023). In contrast, the reverse pattern results in a counterclockwise hysteresis with higher ER during the cooling than warming season (Li et al. 2011; Zhang et al. 2018). Although local studies have widely reported pronounced hysteresis in ER and its components (Li et al. 2011; Zhang et al. 2018; Ataka et al. 2020), the global spatial distribution of ER hysteresis and its underlying mechanisms remain unknown, which hinders our ability to improve projections of the carbon cycle and carbon‐climate feedbacks.

The hysteresis in the ER response to temperature is largely regulated by physiological and environmental factors, particularly substrate availability and soil moisture (Kuzyakov and Gavrichkova 2010; Zhang et al. 2018; Liang et al. 2024). For example, ER relies primarily on vegetation photosynthesis to directly support autotrophic respiration and indirectly as the substrate supplying heterotrophic respiration (Collalti et al. 2020). The transport of photosynthates required for autotrophic respiration via the phloem can be delayed by a couple of days (Kuzyakov and Gavrichkova 2010; Chu et al. 2023). Rapid litterfall decomposition in warming seasons can boost microbial and root activities (Kirschbaum 2013; Ataka et al. 2020), which may reinforce ER hysteresis.

Seasonal distribution of precipitation can also modulate ER hysteresis by affecting photosynthesis (Reich et al. 2018; Zhang et al. 2018) and litter and soil organic matter decomposition (Luo and Zhou 2006; Liang et al. 2024). During the warming season, sufficient precipitation can enhance vegetation photosynthesis (Reich et al. 2018; Zhang et al. 2018), photosynthate transport (Kuzyakov and Gavrichkova 2010), and litter decomposition (Kirschbaum 2013; Ataka et al. 2020), thereby stimulating ER and may contribute to a clockwise hysteresis. Conversely, if the warming season is combined with drought conditions, rising temperatures can intensify evapotranspiration and exacerbate soil moisture deficits, which constrain ER, likely leading to a counterclockwise hysteresis (Reich et al. 2018; Zhang et al. 2018).

The ongoing climate warming has increased the spatial and seasonal variability of precipitation (Cohen and Pincus 2025; Chen et al. 2025) and altered vegetation phenology (Liu et al. 2025; Terasaki Hart et al. 2025), thereby shifting the seasonal timing of substrate availability (Reich et al. 2018; Liu et al. 2025). Consequently, the growing temporal mismatch between precipitation and substrate supply may amplify ER hysteresis and increase uncertainty in predicting carbon‐climate feedbacks, yet the underlying mechanisms and the driving factors remain nearly unknown at the global scale.

Here, we investigated the global patterns of ER hysteresis and disentangled key factors that influence these patterns. We conducted a global‐scale analysis of daily ER across 324 sites, covering 2536 site‐years of eddy‐covariance measurements (Figure 1a). We quantified the hysteresis as the average seasonal difference (∆ER) between the warming and cooling seasons for mean daily ER values by dividing the year into two halves (warming and cooling seasons) based on the minimum and maximum daily mean air temperatures (Figure S1, more details in Section 2) (Niu et al. 2011; Liu et al. 2018). We discovered widespread hysteresis in the global ER response to seasonal temperature changes, comprising clockwise (53%) and counterclockwise (47%) hysteresis patterns. The seasonal mismatch between precipitation and substrate availability largely determined ER hysteresis. We further generated the first global map of ER hysteresis based on site‐level observations. By exploring the global patterns of ER hysteresis and its underlying mechanisms, this study provides valuable insights for benchmarking ER responses to seasonal temperature changes in Earth System Models and improving predictions of the terrestrial carbon cycle trajectory under future climate change.

FIGURE 1.

Distribution of sites in the datasets and response patterns of ecosystem respiration (ER) to seasonal dynamics of air temperature (T _air). (a) Global map showing the geographic distribution of 324 sites from the AmeriFlux, FLUXNET2015, and ICOS datasets. We classified hysteresis as the difference between the mean daily ER during the warming and the cooling seasons (∆ER). Two patterns of ER responses to seasonal warming and cooling: (b) ER responds more to warming than cooling season (clockwise hysteresis, ∆ER > 0, 173 sites); (c) ER responds less to warming than cooling season (counterclockwise hysteresis, ∆ER < 0, 151 sites). Arrows show the direction of hysteresis loops of seasonal changes. For the warming or cooling season, the observed ER data were averaged across all sites, with values binned at 1°C intervals (see Section 2).

2. Materials and Methods

2.1. Datasets and Related Variables

The daily eddy‐covariance carbon fluxes and meteorological data were sourced from FLUXNET2015, Integrated Carbon Observation System (ICOS), and AmeriFlux (https://fluxnet.org/). These data have been carefully controlled, filtered, and gap‐filled by consistent methods across all datasets (Pastorello et al. 2020). Data from identical sites in the FLUXNET2015 dataset were updated with more recent data from ICOS and AmeriFlux. In total, 324 individual sites with 2536 site‐years of eddy‐covariance data were used in this study (Table S1), covering a wide range of geographic area (37.43° S to 74.47° N, Figure 1a).

We used ER data (g C m⁻² d⁻¹, RECO_DT_VUT_REF) derived from the daytime flux‐partitioning method because it provides more reliable estimates for analyzes of CO₂ fluxes (Keenan et al. 2019). Daily mean meteorological parameters, including air temperature (T _air, °C), soil temperature (TS, °C), soil water content (SWC, %), photosynthetically active radiation (PAR, μmol m⁻² s⁻¹), vapor pressure deficit (VPD, hPa), seasonal precipitation (SP, mm), and atmospheric CO₂ mole fraction (CO₂, μmol CO₂ mol⁻¹), were obtained from ancillary data associated with eddy‐covariance measurements of ER. To ensure data quality, flux observations were retained only when more than 50% of the values were directly measured or reliably gap‐filled (TA_F_QC > 0.5) (Wang et al. 2024). To reduce uncertainty in data autocorrelation due to the carbon‐flux partitioning method (Tramontana et al. 2020; Niu et al. 2024), we also included ER data (g C m⁻² d⁻¹, RECO_NT_VUT_REF) from the nighttime method and compared them with those derived from the daytime method. In addition to these meteorological variables, vegetation structure, specifically the leaf area index (LAI), was extracted from global datasets for each site based on its geographic coordinates. The LAI dataset can be accessed at https://doi.org/10.3974/geodb.2023.10.03.V1. For sensitivity analyzes, we compiled additional vegetation‐related proxies, including solar‐induced fluorescence (SIF), the maximum rate of carboxylation (Vcmax), and the normalized difference vegetation index (NDVI). These proxies were used to evaluate whether the inferred vegetation pathway to the variation in ER hysteresis was sensitive to the use of LAI as the main indicator of substrate‐supply capacity. The data sources and temporal resolutions of these products are summarized in Table S2.

2.2. Temperature Response Curve of Ecosystem Respiration and Seasonal Hysteresis

We divided each annual cycle into warming and cooling seasons based on the maximum ($T_{\mathrm{air}}^{\max }$) and minimum ($T_{\mathrm{air}}^{\min }$) daily air temperatures (T _air). For each site, we first constructed a site‐specific, multi‐year mean annual cycle by averaging daily ER, T _air, and other environmental variables across all available site‐years for each day of year (DOY). Site‐years affected by climatic extremes, such as severe droughts or heatwaves, were not excluded if they met the standard quality control for flux data. Consequently, extreme‐event years were retained in the analysis, while multi‐year averaging DOY reduced the influence of any single extreme year on the estimated site‐level hysteresis.

Using the site‐specific multi‐year mean annual cycle, we smoothed daily mean T _air (Figure S1a) using a 5‐day moving window and a 1‐day step to ensure the presence of a single $T_{\mathrm{air}}^{\min }$. We then replicated the 1‐year data to create a 2‐year dataset, where the interval between the two $T_{\mathrm{air}}^{\min }$/$T_{\mathrm{air}}^{\max }$ represents the new daily T _air dynamics (Figure S1b). Finally, the two seasons were defined based on the $T_{\mathrm{air}}^{\min }$/$T_{\mathrm{air}}^{\max }$, with the warming season corresponding to the period before $T_{\mathrm{air}}^{\max }$ and the cooling season corresponding to the period following $T_{\mathrm{air}}^{\max }$ (Figure S1c). These methods have been commonly used in previous studies (Niu et al. 2011; Liu et al. 2018). Differences in ER between warming and cooling seasons, as separated by smoothed temperatures, are consistent with the original data (Figure S2).

The daily dynamics of ER and other factors were divided into two periods based on the days of warming and cooling seasons, as defined earlier. We quantified the magnitude of hysteresis as the mean seasonal difference in daily ER between the warming and cooling seasons (∆ER, Figure S1d), following the method used in the previous study (Niu et al. 2011). A similar analysis was performed for key controlling factors, including LAI, SWC, TS, PAR, VPD, SP, and CO₂, to identify the drivers that contribute to the hysteresis of ER response to T _air on a seasonal scale. We evaluated the uncertainty associated with daytime partitioning of eddy‐covariance CO₂ fluxes and found that the slope between ΔER derived from the daytime and nighttime methods did not differ significantly from 1 (Figure S3).

2.3. Variable Selection for Hierarchical Partitioning and Structural Equation Modeling

We initially evaluated a comprehensive set of environmental and climatic predictors, including seasonal differences, extremes, and seasonal ranges (Table S2). Because ∆ER was calculated from daily‐scale data, variables available only at coarse temporal resolutions, such as annual values, were excluded to ensure temporal consistency. To minimize redundancy among ecologically related predictors, we selected representative, non‐redundant variables while retaining those that captured distinct ecological processes or temporal resolutions. We initially considered multiple vegetation‐related proxies for substrate‐supply capacity, including ∆LAI, ∆GPP, ∆SIF, ∆Vcmax, and ∆NDVI. ∆GPP was excluded because both GPP and ER are derived from the partitioning of eddy‐covariance CO₂ fluxes, potentially introducing statistical non‐independence when ∆GPP is used to explain variation in ∆ER (Pastorello et al. 2020). Although ∆SIF and ∆Vcmax more directly represent photosynthetic activity and physiological capacity, respectively, ∆LAI provides a satellite‐observed structural and phenological proxy for seasonally integrated substrate‐supply capacity (Reichstein et al. 2003; Davidson et al. 2006). We therefore retained ∆LAI in the main model and used ∆SIF and ∆Vcmax as alternative functional proxies in sensitivity analyzes. Pairwise correlation analysis and principal component analysis (PCA) were further conducted to evaluate whether these vegetation proxies represented a common substrate‐supply capacity (Figures S4 and S5).

Following the initial screening, we statistically evaluated the remaining predictors. First, multicollinearity was assessed by calculating variance inflation factors (VIFs) for the full dataset and separately for sites exhibiting clockwise and counterclockwise hysteresis. Predictors with VIF values below 3 were considered free from serious multicollinearity (Zuur et al. 2010), and the VIF results are reported in Table S3. Second, hierarchical partitioning analysis was conducted using the glmm.hp R package (Lai et al. 2022) to assess the relative importance of individual predictors in explaining variation in ER hysteresis. Third, general linear regression models were applied to examine the relationship between ∆ER and its potential drivers. Finally, we developed structural equation models (SEMs) to evaluate the factors directly and indirectly regulating ER hysteresis and to quantify their relative contributions based on standardized total effects. The SEM structure was constructed based on mechanistic interpretability and model parsimony rather than solely on predictor ranks derived from hierarchical partitioning. SEM analyzes were performed using AMOS 24.0 (AMOS Development Corporation, Chicago, IL, USA) with the maximum likelihood estimation method. Overall model goodness‐of‐fit was assessed using the χ ² test and associated p values, the comparative fit index (CFI), and the root mean square error of approximation (RMSEA) (Lavallee et al. 2024).

2.4. Mapping the Global Distribution of Ecosystem Respiration Hysteresis

To predict spatial variations in ER hysteresis, we used site‐level ∆ER as the dependent variable and six variables (∆LAI, ∆TS, ∆SP, ∆PAR, ∆SWC, and ∆VPD) derived from SEM as independent variables in a random forest (RF) model. For each run, the sample dataset was resampled with replacement (bootstrap) and then split into training (70%) and testing (30%) subsets. Random forests were implemented using 500 regression trees (ntree = 500). To control model complexity, the minimum leaf size was optimized via 10‐fold cross‐validation within each bootstrap sample, selecting the parameter that minimized the validation root mean squared error (RMSE). Model performance was quantified using RMSE and the coefficient of determination (R ²), which reached 0.88 on the held‐out testing subset (Figure S6).

Each fitted model was then applied to gridded global climate predictors. Grid cells with more than one missing predictor variable were masked from prediction. To derive robust global estimates, the 100 ensemble predictions were aggregated at each grid cell: the ensemble mean was used as the best estimate of ∆ER, while the ensemble standard deviation (SD) characterized predictive uncertainty. The resulting maps of mean ∆ER and SD (720 × 1440 grid cells) provide spatially explicit predictions together with their associated uncertainty.

However, purely data‐based algorithms such as random forest can yield high uncertainty when extrapolating beyond the environmental conditions represented in the training dataset. To evaluate the spatial reliability of our predictions and avoid unjustified extrapolation, we delineated the model's area of applicability (AOA) using the R package CAST (Meyer and Pebesma 2021). The AOA explicitly defines the multidimensional environmental space where the model can be applied reliably, such that the model performance estimated during cross‐validation is expected to remain valid (AOA = 1). By contrast, grid cells with predictor values falling outside the parameter space covered by the training sites (AOA = 0) were identified as outside the applicability domain and were therefore masked in our final global maps of ΔER.

To obtain global data for these predictors, we used the ERA5 daily dataset (the fifth‐generation reanalysis of the European Center for Medium‐Range Weather Forecasts) to retrieve T _air (https://cds.climate.copernicus.eu/datasets). SP and VPD data can be obtained at https://crudata.uea.ac.uk/cru/data/hrg/, while PAR, TS, and SWC data can be accessed at https://data.tpdc.ac.cn/zh‐hans/data. All datasets were provided at a spatial resolution of 0.25° × 0.25°, based on daily data availability during 2014–2018.

3. Results

3.1. Global Patterns of Ecosystem Respiration Hysteresis

Our analysis revealed widespread hysteresis in ER response to seasonal temperature changes across 324 sites globally (Figure 1a). Two distinct ER hysteresis patterns emerged. The first, clockwise hysteresis (ΔER > 0), marked by higher ER during the warming than the cooling season, was observed at 53% (173 sites) of the sites (Figure 1b). The second, counterclockwise hysteresis (∆ER < 0), characterized by lower ER during the warming season compared to the cooling season, was detected at 47% (151 sites) of the studied sites (Figure 1c). These two contrasting but common hysteresis patterns demonstrate that the ER response to temperature is not uniform but rather has seasonal asymmetric patterns across global biomes.

3.2. Drivers of Ecosystem Respiration Hysteresis

Our hierarchical partitioning analysis demonstrated that the hysteresis in LAI (∆LAI) was the primary driver of clockwise hysteresis. In contrast, the hysteresis in seasonal precipitation (∆SP) emerged as the dominant driver of counterclockwise hysteresis (Figure 2a,c). ΔLAI and ΔSP explained 26% and 39% of the variation in the clockwise and counterclockwise hysteresis, respectively (Figure 2b,d). Structural equation model (SEM) confirmed the pathways and interactions by which ΔLAI and ΔSP dominantly controlled the clockwise and counterclockwise ER hysteresis, respectively (Figure 3a,b). In addition, hysteresis in soil temperature (ΔTS) also contributed to the clockwise hysteresis, while hysteresis in vapor pressure deficit (ΔVPD) accounted for the counterclockwise hysteresis.

FIGURE 2.

Main predictors of clockwise and counterclockwise hysteresis (∆ER) in the response of ecosystem respiration to temperature. (a) Individual contributions of controlling factors in the case of clockwise hysteresis (∆ER > 0), with hysteresis in LAI (∆LAI) as the primary driver. (b) Regression between ∆ER and ∆LAI. (c) Individual contributions of controlling factors in the case of counterclockwise hysteresis (∆ER < 0), with hysteresis in seasonal precipitation (∆SP) as the main influencing factor. (d) Regression between ∆ER and ∆SP. ∆LAI, ∆SP, ∆TS, ∆VPD, ∆SWC, ∆PAR and ∆CO₂ are the differences between the warming and cooling seasons of leaf area index (LAI), seasonal precipitation (SP), soil temperature (TS), vapor pressure deficit (VPD), soil water content (SWC), photosynthetically active radiation (PAR), and atmospheric CO₂ mole fraction (CO₂).

FIGURE 3.

Mechanisms affecting ecosystem respiration hysteresis (∆ER). Structural equation models (SEM, a and b) describes direct and indirect pathways of biotic and abiotic drivers influencing clockwise (∆ER > 0) and counterclockwise hysteresis (∆ER < 0). Red and blue arrows indicate positive and negative relationships, respectively. Numbers adjacent to the arrows are standardized path coefficients, and the arrow width corresponds to the strength of the association. Solid arrows denote significant relationships (p < 0.05), while dashed lines indicate non‐significant relationships (p > 0.05). The R ² values indicate the proportion of variance explained by the model. (c) Schematic diagram illustrating potential mechanisms driving hysteresis direction in ER (ΔER) response to seasonal temperature changes. The arrow indicates the loop direction. Clockwise hysteresis (ER is higher during the warming than cooling season at the same temperatures): ∆ER > 0; Counterclockwise hysteresis (ER is lower during the warming than cooling season at the same temperatures): ∆ER < 0; WA: Water availability; SA: Substrate availability; LAI: Leaf area index; SP: Seasonal precipitation; PAR: Photosynthetically active radiation; VPD: Vapor pressure deficit; SWC: Soil water content.

Specifically, compared with the cooling season, higher LAI during the warming season significantly stimulated ER, thereby inducing clockwise hysteresis (Figure 3c). Conversely, lower precipitation during the warming season reduced water availability, suppressing ER and leading to counterclockwise hysteresis. Overall, the direction and magnitude of ER hysteresis are primarily governed by seasonal mismatches between precipitation and substrate availability.

Sensitivity analyzes using alternative vegetation proxies further supported the robustness of the vegetation‐related substrate‐supply pathway. ∆LAI, ∆SIF, and ∆Vcmax were positively correlated and loaded strongly and positively on the first principal component, which explained 79.3% of their total variance (Figures S4 and S5). This common vegetation axis was positively associated with ∆ER, indicating that ∆LAI captures a coherent seasonal vegetation signal related to substrate supply and consistent with functional indicators of photosynthetic activity and physiological capacity.

3.3. Estimation of Global Ecosystem Respiration Hysteresis

Using the predictors from the SEM analysis, we applied a random forest model to map ER hysteresis at the global scale. The model explained more than 88% of the spatial variation in ER hysteresis (Figure S6). Furthermore, the area of applicability (AOA) analysis demonstrated that 74.8% of the global land surface fell within the reliable environmental space represented by the training data (AOA = 1), indicating strong model reliability in capturing the global pattern. The resulting map showed that ΔER ranged from −1.7 to 1.0 g C m⁻² d⁻¹ globally (Figure 4), with distinct differences across climatic zones. The clockwise hysteresis dominated in the temperate continental climate zones of North America and Europe. In contrast, the counterclockwise hysteresis prevailed in savanna regions, including central and southern Africa, northern Australia, and South America, largely consistent with the hysteresis pattern of SP in response to temperature (ΔSP, Figure S7). Notably, regions with greater uncertainty in predicted ER hysteresis, where the standard deviation (SD) across 100 model iterations exceeded the global average of 0.2 g C m⁻² d⁻¹ (Figure S8a), were spatially consistent with areas identified by the area of applicability method as outside the applicability domain (AOA = 0) (Figure S8b).

FIGURE 4.

Global pattern of hysteresis in the response of ecosystem respiration to temperature (∆ER, g C m⁻² d⁻¹). ∆ER was calculated with a random forest model using biotic and abiotic variables for 2014–2018. Gray areas indicate regions outside the model's area of applicability (AOA = 0), representing environmental conditions where predictions are considered unreliable due to extrapolation beyond the training data (see Section 2).

4. Discussion

This study provides robust empirical evidence for the widespread hysteresis in ER responses to seasonal warming and cooling, based on eddy‐covariance data from 324 globally distributed sites. Our findings reveal two common but distinct hysteresis patterns: clockwise and counterclockwise hysteresis. The directions and magnitude of ER hysteresis are regulated by the seasonal mismatch between precipitation (affecting soil water content) and photosynthesis, defining substrate availability for microorganisms. To the best of our knowledge, this study represents the most comprehensive efforts to identify the hysteresis in ER responses to seasonal temperature changes, offering new insights into how these hysteresis patterns influence the trajectory of global respiration and land carbon‐climate feedbacks in a warming world.

Seasonal variation in substrate availability for roots and soil microorganisms appears to be a key driver of hysteresis in ER and its components. Autotrophic respiration, which accounts for approximately 30%–60% of global ER (Qin et al. 2024), is strongly coupled to photosynthetic activity and the supply of recently assimilated carbon (Kuzyakov and Gavrichkova 2010; Collalti et al. 2020). Because photosynthesis is closely linked to photosynthetically active radiation (PAR) (Niu et al. 2011; Zhao et al. 2022), seasonal peaks in photosynthetic carbon supply often precede the annual temperature maximum. This temporal decoupling between substrate availability and temperature may induce hysteresis in autotrophic respiration (Collalti et al. 2020). Vegetation phenology further determines the seasonal timing of carbon assimilation and substrate transfer belowground (Reich et al. 2018; Liu et al. 2025). Climate warming has induced spatially heterogeneous phenological responses, including earlier vegetation activity across many regions but delayed green‐up or canopy development in some high‐latitude or high‐elevation ecosystems (Huang et al. 2023). These phenological shifts may alter the temporal alignment between substrate supply and seasonal temperature dynamics, thereby changing the magnitude or even the direction of ER hysteresis.

Heterotrophic respiration also has hysteresis, partly due to seasonal changes in microbial biomass and a decline in substrate availability following the rapid decomposition of fresh litter, especially during the cooling season when substrate replenishment is limited (Kirschbaum 2013; Ataka et al. 2020). In addition, delays in the phloem transport of photosynthates to tree roots (2–5 days) and their subsequent release into the soil for microbial use may contribute to hysteresis in both autotrophic and heterotrophic respiration (Davidson et al. 2006; Kuzyakov and Gavrichkova 2010; Gavrichkova and Kuzyakov 2017). Therefore, seasonal variation in substrate availability, particularly its temporal mismatch with temperature, likely regulates the hysteresis of ER components and contributes to the observed ER hysteresis under seasonal temperature changes (Hu et al. 2016; Li et al. 2017; Ataka et al. 2020).

Although leaf area index (LAI) is not a direct physiological measure of photosynthesis (Reichstein et al. 2003; Davidson et al. 2006), the consistency among ΔLAI, ΔSIF, and ΔVcmax suggests that the identified vegetation pathway represents an integrated substrate‐supply signal rather than a LAI‐specific effect (Figure S5). Seasonal vegetation development and substrate supply, for which LAI serves as a structural and phenological proxy, are strongly influenced by temperature (Chen et al. 2019; Zhang et al. 2025) and precipitation (Reich et al. 2018; Collalti et al. 2020). PAR was also consistently higher during the warming season than during the cooling season at nearly all sites (Figure S9). In regions where water availability is not limiting, favorable hydrothermal conditions promoted substrate supply and enhanced ER during the warming season, thereby contributing to clockwise hysteresis (Figures 3c and S7). Several mechanisms may explain this pattern. First, adequate moisture conditions accelerate vegetation structural development (e.g., increased LAI), thereby increasing photosynthetic substrate availability and inducing clockwise hysteresis in autotrophic respiration, which may dominate ER during the warming season (Figure S10a–f). Second, favorable hydrothermal conditions stimulate microbial and enzymatic activities, accelerate the decomposition of macromolecular organic compounds in soil (Davidson and Janssens 2006; Ataka et al. 2020), and increase the availability of labile substrates (Kirschbaum 2013), all of which raise heterotrophic respiration and contribute to clockwise hysteresis. Third, greater soil‐water film thickness facilitates microbial and enzymatic diffusion and accelerates root nutrient uptake (Davidson et al. 2006), further stimulating ER and reinforcing clockwise hysteresis.

Under moisture‐limited conditions, however, precipitation plays the dominant role in regulating ER hysteresis by influencing both substrate availability and microbial access to soil substrates, thereby modulating ER responses to temperature and promoting counterclockwise hysteresis. During the warming season, for instance, declining precipitation combined with rising temperatures can increase vapor pressure deficit (VPD) and intensify drought stress (Zhang et al. 2024). These conditions constrain vegetation growth and induce stomatal closure (Reich et al. 2018; Wang et al. 2023), thereby reducing photosynthetic carbon assimilation and substrate supply and ultimately suppressing ER, particularly autotrophic respiration (Figure S10g–i). Simultaneously, drought reduces soil‐water film thickness and restricts the diffusion of dissolved organic carbon and other soluble substrates (Davidson et al. 2006; Holz et al. 2018), reducing substrate accessibility to soil microorganisms and suppressing heterotrophic respiration. Subsequent increases in precipitation during the cooling season may stimulate rapid respiration pulses through rewetting‐induced substrate mobilization and priming effects (Metz et al. 2023), resulting in higher ER at a given temperature than during the warming season and thus generating counterclockwise hysteresis. Collectively, the temporal mismatch between seasonal precipitation and substrate availability largely determines the direction and magnitude of ER hysteresis (Figures 3c and S10). Including all seven candidate predictors in the SEMs increased the proportion of variance explained in ΔER by only 0.01 for both clockwise and counterclockwise hysteresis (Table S4 and Figure S11). This marginal improvement indicates that our selected predictors captured the predominant controls and that the inferred importance of seasonal precipitation and substrate availability was robust to the inclusion of additional candidate variables.

Temperature and precipitation vary substantially across both space and time worldwide (Johnson et al. 2018), which can shape ER hysteresis and induce uncertainty into projections of land carbon‐climate feedbacks. By mapping the global distribution of ER hysteresis, we identified when and where ecosystem carbon emissions may be particularly sensitive to climate warming. Savanna climate zones predominantly exhibited counterclockwise hysteresis, driven by higher precipitation during the cooling than the warming season (Figure S7). In these regions, further warming during the warming season may not proportionally enhance ER because respiration is constrained by water and substrate availability, potentially attenuating the positive feedback between respiration and climate warming. By contrast, in regions exhibiting clockwise hysteresis, relatively abundant precipitation and substrate availability during the warming season may strengthen the respiratory response to rising temperature and amplify this feedback. Global warming may also increase interannual variability in precipitation regimes (Cohen and Pincus 2025; Chen et al. 2025) and vegetation phenology (Liu et al. 2025; Terasaki Hart et al. 2025). Such changes could directly or indirectly alter the timing and magnitude of substrate supply, thereby increasing uncertainty in predicting ER hysteresis. Nevertheless, the strong agreement between predicted and observed hysteresis (R ² = 0.88) suggests that our model reliably captured both its direction and magnitude across 74.8% of the global regions (Figures 1a and 4). However, the extent to which changes in precipitation and phenology alter substrate availability and, consequently, the interannual stability of ER hysteresis remains unclear and warrants further investigations.

In addition to these mechanistic complexities, uncertainties inherent in the underlying data sources must also be considered. Eddy‐covariance measurements are subject to non‐negligible random errors and uncertainties arising from gap filling (Tramontana et al. 2020; Niu et al. 2024). These uncertainties may obscure hysteresis signals at sites where the estimated magnitude is exceptionally small (ΔER ≈0). To assess this possibility, we performed sensitivity analyzes using both statistical and empirical thresholds to exclude sites characterized by weak or negligible hysteresis and potentially dominated by measurement noise (see Supporting Methods). Hierarchical partitioning of the filtered datasets consistently identified ΔLAI and ΔSP as the predominant predictors of clockwise and counterclockwise hysteresis, respectively, confirming the robustness of the inferred environmental controls (Figure S12). Notably, retaining sites with near‐zero ΔER maintained higher overall explanatory power, suggesting that these sites may represent ecologically meaningful transitional states along the continuous environmental gradients rather than measurement artifacts alone. Furthermore, our AOA analysis identified ecoclimatic regions in which additional observations are needed to improve understanding of seasonal ER dynamics and their environmental controls.

Despite the prevalence of ER hysteresis in the observational dataset, this pattern was not adequately reproduced by the CABLE dynamic global vegetation model (DGVM). Comparison of eddy‐covariance observations with CABLE simulations revealed that the model failed to capture the ER hysteresis evident in the observations (Figure S13). This discrepancy may partly arise from the model's inability to account for interactions and spatio‐temporal heterogeneity among ER drivers (e.g., precipitation, vegetation structure, and photosynthetic activity). It may also stem from inconsistencies between the apparent temperature sensitivity (Q₁₀) function used in CABLE and the actual temperature response of ER components (Haaf et al. 2021; Niu et al. 2024). Future generations of DGVMs should therefore represent not only the regulatory roles of precipitation and substrate availability in driving ER hysteresis, but also their seasonal phase relationships with temperature. Explicitly accounting for temporal mismatches among temperature, water availability, and substrate supply may improve the representation of ER hysteresis and reduce uncertainty in projections of land carbon‐climate feedbacks. Such seasonal disequilibrium may become increasingly important under climate change as precipitation regimes and vegetation phenology continue to shift.

In summary, this study provides the first global‐scale evidence of widespread hysteresis in the ER response to seasonal temperature changes. We identified two distinct hysteresis patterns whose direction and magnitude were primarily associated with seasonal timing mismatches between precipitation‐driven water availability and vegetation‐derived substrate supply. This temporal decoupling demonstrates that ecosystem carbon release is inherently non‐stationary and reveals an overlooked mechanism governing terrestrial carbon‐cycle dynamics.

Continued climate warming is expected to intensify precipitation variability and alter vegetation phenology, potentially amplifying temporal mismatches between water availability and substrate supply and increasing uncertainty in projections of land carbon‐climate feedbacks. However, these hysteretic processes remain inadequately represented in current dynamic global vegetation models. Future model development should explicitly account for the seasonal phase relationships among temperature, hydroclimatic conditions, and biotic factors. Incorporating these temporal interactions will improve the representation of ER dynamics and strengthen projections of terrestrial carbon cycling under a changing climate.

Author Contributions

Houkun Chu: writing – original draft, writing – review and editing, data curation, methodology. Lìyǐn L. Liáng: writing – review and editing, supervision, methodology. Miko U. F. Kirschbaum: methodology, supervision, writing – review and editing. Ashley Ballantyne: supervision, writing – review and editing. Qin Zhang: supervision, writing – review and editing. Weinan Chen: methodology, data curation. Jianyang Xia: supervision, writing – review and editing. Qinyu Zheng: methodology, data curation. Kailiang Yu: supervision, writing – review and editing. Zheng Fu: supervision, writing – review and editing. Dashuan Tian: supervision, writing – review and editing. Ronglei Zhou: methodology, data curation. Song Wang: methodology, supervision, writing – review and editing. Jiaqiang Liao: methodology, data curation. Yakov Kuzyakov: supervision, writing – review and editing. Shuli Niu: funding acquisition, methodology, supervision, writing – review and editing. Guirui Yu: supervision, writing – review and editing. Jinsong Wang: funding acquisition, writing – review and editing, supervision, methodology.

Funding

This work was supported by National Natural Science Foundation of China (Grants 32241035, 32588202, 32401384 and 32501476), “Kezhen‐Bingwei” Young Talents (Grant 2022RC004) and China Postdoctoral Science Foundation (Grants GZC20241692 and 2025M772580).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Division between warming and cooling seasons. (a), Original data for air temperature (T _air) by Julian day. (b), The moving average of daily T _air ensures a unique $T_{\mathrm{air}}^{\min }$, and 1 year of data is then replicated to create 2 years of data, (c), With the interval between the two $T_{\mathrm{air}}^{\min }$ representing the new daily T _air dynamics. The years were divided into two parts based on the $T_{\mathrm{air}}^{\max }$: warming and cooling seasons. (d), Ecosystem respiration (ER) during the warming and cooling seasons has a pronounced hysteresis pattern. The data were obtained from sites in Klingenberg, Germany (DE‐Kli).

Figure S2: Agreement between calculated hysteresis in the response of ecosystem respiration (∆ER) to smoothed and original temperatures. The dashed line represents the 1:1 fitting line. Shaded areas indicate 95% confidence intervals.

Figure S3: Agreement between ecosystem respiration hysteresis (ΔER) estimated from daytime and nighttime partitioning methods. The dashed line represents the 1:1 fitting line. Shaded areas indicate 95% confidence intervals.

Figure S4: Pairwise correlations among vegetation‐related proxies across ER hysteresis patterns. Pairwise correlations among ΔGPP, ΔSIF, ΔVcmax, ΔNDVI, and ΔLAI were evaluated for all sites, sites exhibiting clockwise hysteresis (ΔER > 0), and sites exhibiting counterclockwise hysteresis (ΔER < 0). Values in the upper triangle represent Pearson correlation coefficients. Circles in the lower triangle indicate the direction and significance of the correlations, with color denoting the sign and magnitude of each correlation. Asterisks indicate statistical significance, with *** denoting p < 0.001.

Figure S5: Principal component analysis (PCA) of ΔLAI, ΔSIF, and ΔVcmax and their integration into a common vegetation axis. (a), PCA biplot showing the distribution of sites along the first two principal components. Points represent an individual eddy‐covariance site and is colored according to the direction of ER hysteresis. Red points indicate sites exhibiting clockwise hysteresis (ΔER > 0), whereas blue points denote sites exhibiting counterclockwise hysteresis (ΔER < 0). Arrows represent the loading vectors of ΔLAI, ΔSIF, and ΔVcmax, and ellipses indicate the distributions of sites within the two hysteresis groups. (b), Relationship between vegetation PC1 and ΔER. The solid line represents the fitted linear regression, and the gray band indicates the 95% confidence interval.

Figure S6: Distribution of the coefficient of determination (R ²) across 100 training runs of the random forest model.

Figure S7: Global pattern of hysteresis in the response of seasonal precipitation to temperature (∆SP, mm d⁻¹). ∆SP was calculated using the dataset for 2014–2018.

Figure S8: Uncertainty and area of applicability for the global estimation of hysteresis in the temperature response of ecosystem respiration (∆ER, g C m⁻² d⁻¹). (a) Spatial distribution of prediction uncertainty, expressed as the standard deviation (SD) across 100 iterations of the random forest model. (b) Global area of applicability (AOA), indicating the spatial reliability of model predictions. Regions within the model's reliable environmental space (AOA = 1) account for 74.8% of the global land surface, whereas regions requiring extrapolation beyond the training data (AOA = 0; shown in gray) account for 25.2% and are masked in the main map.

Figure S9: The dependence of photosynthetically active radiation (PAR) (mean ± SE) on annual temperature across all sites. The histogram shows the distribution of the magnitude of hysteresis in the response of PAR to temperature (∆PAR). Arrows show the direction of hysteresis loops of seasonal changes.

Figure S10: The leaf area index (LAI) and ecosystem respiration (ER) response to temperature depending on seasonal precipitation (SP). The temperature response hysteresis loops are shown via SP mapping, with the arrow indicating the direction of the loop. The data following each site name represent the mean annual precipitation.

Figure S11: Sensitivity SEMs including all seven candidate predictors after variable screening. Full seven‐predictor structural equation models were constructed for (a) clockwise hysteresis (ΔER > 0) and (b) counterclockwise hysteresis (ΔER < 0) to test whether excluding some candidate predictors from the main SEMs affected the mechanistic interpretation. The seven predictors included ΔLAI, ΔSWC, ΔSP, ΔPAR, ΔVPD, ΔTS, and ΔCO₂. Red and blue arrows indicate positive and negative relationships, respectively. Numbers adjacent to arrows are standardized path coefficients, and arrow width represents the strength of the relationship.

Figure S12: Relative importance of environmental predictors of ER hysteresis in sensitivity analyzes after excluding sites with weak or near‐zero hysteresis signals (∆ER ≈0). Panels show results after filtering using a statistical t‐test (a, b) and a 5% relative threshold (c, d) for clockwise (∆ER > 0; a, c) and counterclockwise (∆ER < 0; b, d) hysteresis, respectively. ∆LAI and ∆SP remain the dominant predictors of clockwise and counterclockwise hysteresis, respectively, confirming that the main findings are robust to the potential influence of measurement noise.

Figure S13: Comparison of observed and CABLE modeled ER response to temperature. (a–d) The four selected sites represent four pairs of functional comparisons for the temperature dependence of ER between the Dynamic Global Vegetation Model CABLE (hollow markers) and observed data (solid markers). At the sites reviewed in this study, the ER response to temperature exhibited predominantly clockwise hysteresis (52%, b, d), whereas the CABLE model primarily simulated counterclockwise hysteresis (62%, a, b). Moreover, the observed and modeled ER responses to seasonal warming and cooling showed opposite directions at 40% of the sites (b, c). Even when the observed and modeled ER responses aligned in direction, the magnitude of ER responses to seasonal temperature changes had substantial divergence between model and observation (a, d).

Table S1: Information of 324 sites in this study.

Table S2: Candidate environmental predictors and data sources evaluated in this study.

Table S3: Variance inflation factors (VIFs) of candidate predictors.

Table S4: Comparison between parsimonious SEMs and full seven‐predictor SEMs.

Data Availability Statement

Data from FLUXNET2015, Integrated Carbon Observation System (ICOS), and AmeriFlux can be downloaded at https://fluxnet.org/. In the current prediction of ER hysteresis patterns, air temperature (T _air) was obtained from the ERA5 daily dataset (the fifth‐generation reanalysis of the European Centre for Medium‐Range Weather Forecasts, https://cds.climate.copernicus.eu/datasets). Seasonal precipitation (SP) and vapor pressure deficit (VPD) data can be obtained from https://crudata.uea.ac.uk/cru/data/hrg/, while photosynthetically active radiation (PAR), soil temperature (TS), and soil water content (SWC) data can be accessed at https://data.tpdc.ac.cn/zh‐hans/data. LAI data can be accessed at https://doi.org/10.3974/geodb.2023.10.03.V1.