A Drier Year Markedly Enhances Methane, but Not Carbon Dioxide, Emissions in a Mediterranean Reservoir

ABSTRACT

Reservoirs are increasingly recognized as important sources of greenhouse gases (GHGs), yet the temporal scales and environmental drivers governing CO₂ and CH₄ flux variability remain poorly constrained. This is particularly the case for Mediterranean systems, which are subject to strong hydrological variability. Here, we present and analyze year‐round, high‐frequency eddy covariance measurements of CO₂ and CH₄ fluxes in a Mediterranean reservoir over two years with contrasting hydrological conditions, explicitly evaluating variability at diurnal, multiday, and seasonal timescales. The reservoir acted as a net source of both gases in both years, but CH₄ fluxes exhibited substantially greater temporal variability than CO₂ fluxes. The seasonal component dominated CH₄ flux variability (> 75%), with annual areal fluxes and reservoir‐wide emissions 44% higher during the drier year due to an earlier onset and longer duration of the high‐emission summer period. Near‐sediment temperature and water‐column depth were the primary predictors of seasonal CH₄ emissions, reflecting temperature controls on methanogenesis and hydrological regulation of ebullition. In contrast, CO₂ flux variability was dominated by diel processes, with wind speed emerging as the main predictor of short‐term emissions, while water depth and surface temperature regulated seasonal variability. For both gases, chlorophyll‐a was an important predictor at multiday timescales. Our results demonstrate that the temporal variability of GHG fluxes from Mediterranean reservoirs is highly sensitive to hydrologically mediated changes in lake water levels, particularly for methane. Under climate change, GHG emissions could increase significantly in Mediterranean regions, and potentially in other seasonally dry regions worldwide, as a result of projected trends towards stronger eutrophication and increases in drought frequency and duration. Our study underscores the value of long‐term, high‐frequency observations to improve emission estimates and to better represent reservoir processes in regional and global carbon cycle assessments.

We monitored greenhouse gas emissions from a Mediterranean reservoir over 2 years that differed markedly in water availability. Methane emissions were much higher during the drier year, largely because lower water levels and warmer bottom waters favored methane production and release, while carbon dioxide emissions were less affected. Our results highlight the sensitivity of reservoir methane emissions to drought and their potential to increase under future climate change.

1. Introduction

Inland freshwater ecosystems, despite occupying a small fraction of the Earth's surface (Messager et al. 2016), play an outsized role in the global carbon cycle (Tranvik et al. 2009), contributing substantially to global atmospheric CO₂ and CH₄ emissions (Rosentreter et al. 2021; DelSontro et al. 2018; Raymond et al. 2013). In lakes and reservoirs, although CO₂ emission rates are roughly four times higher than CH₄ emission rates, CH₄ is the dominant greenhouse gas (GHG) in terms of climate forcing due to its much higher global warming potential, accounting for approximately 75% of total CO₂‐equivalent emissions in these lentic systems (DelSontro et al. 2018; Deemer et al. 2016). Recent global bottom‐up assessments indicate that more than 40% of global methane emissions originate from aquatic ecosystems, with lakes and reservoirs alone accounting for over 10% (Rosentreter et al. 2021). Large uncertainty is still inherent in these estimates, as observational data coverage is insufficient—especially with respect to episodic fluxes, the complete annual cycle, and spatial coverage (e.g., Rodríguez‐Velasco et al. 2024; Johnson et al. 2022, 2021; Deemer and Holgerson 2021; Deemer et al. 2016).

Air–water CO₂ and CH₄ fluxes integrate the effects of a wide range of physical, biological, and, particularly in the case of CO₂, chemical processes operating across multiple temporal scales within a year. CO₂ fluxes reflect the balance between metabolic processes such as net ecosystem production and allochthonous carbon inputs (Lapierre et al. 2013; McDonald et al. 2013; Cole et al. 2007), carbonate equilibrium and calcite precipitation (Many et al. 2024), catchment lithology (León‐Palmero, Morales‐Baquero, and Reche 2020), as well as physical forcing associated with stratification, mixing, and synoptic weather variability (e.g., Reed et al. 2018; Liu et al. 2016). These processes can interact with trophic dynamics, including phytoplankton–zooplankton cycles and the development of algal blooms, further modulating CO₂ exchange at diel to seasonal scales (Ouyang et al. 2017). In contrast, methane emissions are primarily governed by sedimentary methanogenesis under anoxic conditions and by oxidation (Bastviken et al. 2004; Mayr et al. 2020) and/or further production in the water column (Ordóñez et al. 2023; León‐Palmero, Contreras‐Ruiz, et al. 2020; Tang et al. 2014), with fluxes strongly shaped by physical transport and release mechanisms. Seasonal stratification and turnover, ice formation, hydrological variability, and episodic pressure‐ and temperature‐driven ebullition events exert strong control on CH₄ emissions (Ustinov et al. 2025; Jammet et al. 2015; Yvon‐Durocher et al. 2014; Bastviken et al. 2004). Nutrient loading and eutrophication further modulate methane production and release by enhancing labile organic matter supply for methanogenesis (Martínez‐García et al. 2024; Grasset et al. 2018; West et al. 2015).

Beyond intra‐annual variability, another source of uncertainty in estimates of lake and reservoir CO₂ and CH₄ emissions arises from interannual variability in both flux magnitudes and the relative importance of their controlling processes. Year‐to‐year changes in climate forcing, hydrology, and ecosystem state can alter the balance between physical transport, biogeochemical production and consumption, and biological activity, potentially leading to shifts in the dominant drivers of gas exchange. This interannual variability is particularly relevant in the Mediterranean region (roughly 27°–47° N, 15° W–40° E) and more generally, in regions with a Mediterranean climate, which are characterized by a pronounced warm and dry season and a highly uneven distribution of precipitation during the remainder of the year. As a result, lakes and reservoirs in this region commonly experience large seasonal fluctuations in water temperature and water level, with strong implications for stratification, mixing dynamics, and sediment–water interactions. In addition, the Mediterranean climate exhibits marked interannual and multi‐decadal variability in precipitation (e.g., Esteban‐Parra et al. 2022), an increasing trend in evaporative water losses due to rising air temperatures (e.g., Esteban‐Parra et al. 2022), and widespread anthropogenic alterations of catchments. Reservoirs are the dominant lentic ecosystems in this region (e.g., Lehner and Döll 2004), and water‐level drawdown during summer is often intensified to meet irrigation and domestic water demands, further amplifying hydrological variability.

Despite the recognized importance of lakes and reservoirs in global carbon budgets and the substantial intra‐ and interannual variability of CO₂ and CH₄ fluxes, much of our understanding and assessment is still based on discrete measurements of fluxes conducted with flux chambers or gas traps that are deployed intermittently, or indirect gas‐transfer models (e.g., Deemer et al. 2016; Wik et al. 2013; Bastviken et al. 2004). While these approaches have provided valuable insights, they are often limited in temporal resolution, spatial coverage, and their ability to capture episodic or scale‐dependent flux dynamics. In contrast, eddy covariance (EC) systems enable continuous, long‐term, and spatially integrated non‐invasive measurements of air–water gas exchange (e.g., Baldocchi 2014), offering a unique opportunity to quantify carbon fluxes across multiple temporal scales. To date, only a small subset of EC studies have examined the temporal patterns or environmental drivers of CO₂ and/or CH₄ fluxes in lakes or reservoirs over complete annual cycles or across multiple years, whether for CO₂ (Golub et al. 2023; Scholz et al. 2021; Reed et al. 2018; Ouyang et al. 2017; Liu et al. 2016; Mammarella et al. 2015; Shao et al. 2015; Huotari et al. 2011), CH₄ (Waldo et al. 2021; Iwata et al. 2020; Taoka et al. 2020) or both gases (Hounshell et al. 2023; Spank et al. 2023; Eugster et al. 2020; Jammet et al. 2017). Only four of these studies were conducted in temperate reservoirs (Hounshell et al. 2023; Spank et al. 2023; Waldo et al. 2021; Liu et al. 2016), and only two of these (Hounshell et al. 2023; Spank et al. 2023) measured CO₂ and CH₄ concurrently. Moreover, none of these studies explicitly examined reservoir carbon fluxes under hydrologically contrasting years, which are common in Mediterranean climates. Consequently, our understanding of how ecosystem transitions and contrasting or altered hydrological regimes regulate reservoir CO₂ and CH₄ fluxes—and their underlying drivers—remains limited. Here, we use EC measurements from a shallow Mediterranean reservoir spanning two hydrologically contrasting years—a dry year with low precipitation and reduced water storage, and a wet year—to identify the dominant temporal scales of CO₂ and CH₄ flux variability, determine the key environmental predictors at each scale and their interannual shifts in importance, and quantify the influence of hydrological variability on reservoir carbon flux dynamics.

2. Methods

2.1. Study Site

Cubillas Reservoir is a small, shallow, polymictic eutrophic reservoir located in southeastern Spain (37.276° N, −3.671° E). The reservoir is primarily used for irrigation and recreation. It has been in operation since 1956 and drains a 647 km² calcareous, predominantly agricultural watershed (e.g., León‐Palmero, Morales‐Baquero, and Reche 2020), with Cubillas River as the main inflow. The reservoir has a surface area of 1.94 km², and its initial total capacity was 19 Hm³. Current capacity is generally much lower and averaged (reference period 1991–2020) 13 ± 4 Hm³, with water levels at 638 ± 3 m.a.s.l, resulting in annual maximum depths near the dam of 13.6 ± 2.8 m (Figure 1a). Within a year, the reservoir experiences an average water‐level drawdown of 5.4 ± 2.1 m (Figure 1a), with maximum levels typically occurring in April–May and minimum levels at the beginning of each hydrological year (October). For this study, we use data collected in 2022 and 2024, which represent contrasting hydrological conditions. Although both years were warm (positive air temperature anomalies; Figure 1b), Year 2022 (hereafter the dry year) was characterized by a cumulative precipitation of 447 mm (close to the historical average, Figure 1c), a maximum water depth of 8 m, and an average inflow rate of 0.34 m³ s⁻¹. In contrast, 2024 (hereafter the wet year) had a cumulative precipitation of 532 mm (positive anomaly, Figure 1c), a maximum reservoir depth of 9.4 m, and inflow rates of 0.42 m³ s⁻¹. Although 2023 exhibited a larger negative precipitation anomaly than 2022, EC malfunction throughout the entire summer prevented temporal analyses of carbon fluxes for that year.

FIGURE 1.

Historical meteorological and hydrological anomalies for the Cubillas Reservoir from 1980 to 2024. (a) Annual maximum reservoir water depth and annual water‐level drawdown, (b) mean annual air temperature (T _air), and (c) cumulative annual precipitation and number of wet days (i.e., days with > 1 mm of precipitation). Anomalies were calculated relative to the 1991–2020 reference period. Historical air temperature and precipitation data were derived by interpolating all available observations from the State Meteorological Agency (AEMET), the Automatic Hydrological Information System (SAIH), and the Andalusian Agroclimatic Information Network (RIA) meteorological stations, as in Herrero (2007). Maximum reservoir depth was derived from reservoir bathymetry, historical water‐level records, and historical changes in hypsographic curves provided by the Guadalquivir Hydrographic Confederation. Grey shaded areas indicate the 2 years studied.

2.2. Data Collection

We deployed a 5 × 5 m aluminum floating platform at 37.278° N, −3.672° E (Figure S1) in December 2021. The platform is anchored to the reservoir sediment using four 300 kg concrete blocks, deployed at its four vertices and positioned 40–50 m from the platform. At ≈10 m from the platform, each of the four mooring ropes has an additional 30‐kg concrete block, which acts as an additional anchorage during low‐water level conditions and limits the platform's motion and rotation during water level changes. The platform holds an accelerometer (WTGAHRS2, Wit‐Motion) to track platform rotation angles (yaw, pitch, roll). The accelerometer measured at a rate of 1 Hz until March 2023, when it was increased to 10 Hz.

The observation platform was equipped with an EC measurement system installed at a height of 3.34 m. The EC system includes a 3D ultrasonic anemometer (Windmaster, Gill) to measure wind speed and direction, as well as two open‐path infrared gas analyzers to measure CH₄ (LI‐7700, LI‐COR), and CO₂ and water vapor (LI‐7500DS, LI‐COR) in the ambient air. The EC system records data at a rate of 10 Hz. Since installation, the system also provides data on air temperature and relative humidity every minute, and starting on May 5, 2022, it also records incoming and reflected shortwave radiation and incoming and outgoing longwave radiation (CNR4 net radiometer, Kipp & Zonen) with the same frequency. All the flux and meteorological data collected by the EC system were integrated into the SmartFlux system (LI‐COR).

Additional 5‐min meteorological data, including atmospheric pressure, wind speed and direction (2D WindSonic M, Gill), and incoming photosynthetically active radiation PAR (LI‐190R, LI‐COR), were also available at the platform. In addition, one surface mooring and one bottom mooring, both fixed to the platform, provided measurements of other environmental variables. For this study, we used: (i) 1‐min water temperature records from three high‐frequency thermistors built by a local manufacturer (Edrónica), calibrated to a precision of 0.01°C, and located 0.5 and 1 m below the lake surface and 0.5 m above the lake sediment; and (ii) 1‐min records of dissolved oxygen DO concentration collected with optical sensors (TrioOS) at 1 m depth and 0.5 m above the lake sediment. To convert height above the sediment to depth below the surface, we also installed a piezoresistive level probe (36XW, Keller) at a height of 0.5 m above the lake sediment. Hanging at 0.5 m below the lake surface, a fluorometer (nanoFlu, TrioOS) provided data on chlorophyll‐a concentration every 5 min. To avoid photoquenching effects (e.g., Rousso et al. 2021), we only used nighttime values. We further retained only chlorophyll‐a measurements collected after sensor cleaning; before June 2024, cleaning was performed manually on a regular weekly basis, while from June 2024, an automatic wiper enabled daily quality‐controlled values. Hourly data on water level, reservoir volume and inflow and outflow discharges are available from the Guadalquivir Hydrographic Confederation website (https://www.chguadalquivir.es/saih/Inicio.aspx; last access on 29 January 2026).

2.3. EC CH₄ and CO₂ Flux Calculations

We obtained 30‐min CH₄ and CO₂ fluxes from the raw 10 Hz EC data using the EddyPro v.7.0.9 software (LI‐COR). Flux processing followed established international standards and protocols, including spike detection and removal (Vickers and Mahrt 1997), double rotation for sensor tilt correction (Wilczak et al. 2001), covariance maximization with default settings for time‐lag compensation between the anemometer and the gas analyzers, and block‐averaging to extract turbulent fluctuations from the time series. Corrections applied included the Webb–Pearman–Leuning (WPL) density correction for open‐path analyzers (Webb et al. 1980), as well as spectral corrections for high‐frequency (Moncrieff et al. 1997) and low‐frequency (Moncrieff et al. 2004) signal attenuation. For all the fluxes, we calculated quality check flags (Mauder and Foken 2004) to test for developed turbulent conditions and stationarity. We selected only 30‐min fluxes with the best quality flags (flag values = 0). We further filtered fluxes based on the automatic gain control (AGC) values of the open‐path gas analyzers to ensure optical cleanliness. Specifically, we retained periods with AGC ≥ 70% for the CO₂ gas analyzer and AGC ≥ 20% for the CH₄ gas analyzer.

We conducted a second analysis to evaluate platform rotation on EC flux estimate errors using the raw 10 Hz EC data for 2024 for which yaw, pitch and roll rotational angles were also available at 10 Hz. Prior to the flux calculations with EddyPro, we corrected the raw 3D ultrasonic anemometer data for sensor tilt using the measured yaw, pitch and roll angles (e.g., Spank et al. 2023), and we compared the results with the unrotated 30‐min fluxes. Pitch and roll rotation of the platform led to non‐significant differences between rotated and non‐rotated signals (Figure S2), demonstrating that rotational effects along the horizontal x and y axes could be discarded. The wind direction estimated by the EC systems was corrected for yaw angle platform rotation before footprint calculations. We calculated the sampling area of the EC tower with the 2D flux footprint model of Kljun et al. (2015). We retained all 30‐min CO₂ and CH₄ flux measurements for which the wind was coming from directions where 80% of the footprint (80% contour line) was over the water at the lowest water level reached in a given year. Flux measurements were also removed when wind direction indicated potential influence from the dock or from a transient island that emerges in the northern sector of the reservoir during periods of low water level. This translated into removing wind directions between −36° and 50° in 2022 and −25° and 50° in 2024 (Figure S1). After quality control and footprint analysis, we retained 23% of the 30‐min CO₂ and 36% of CH₄ fluxes in 2022, and 25% of CO₂ fluxes and 33% of CH₄ fluxes in 2024. These percentages fall within the range reported for EC measurements over lakes and reservoirs (e.g., Hounshell et al. 2023; Erkkilä et al. 2018). We filled gaps in the 30‐min flux time series for CO₂ and CH₄ using the REddyProc online tool (Wutzler et al. 2018; https://bgc.iwww.mpg.de/5622399/REddyProc). Gaps remaining in the series (less than 1.5%) were filled through linear interpolation. Time series of gas fluxes were averaged into hourly and daily fluxes.

2.4. Wavelet Analysis

We partitioned the original hourly time series of gap‐filled CH₄ fluxes into diffusive and ebullitive components using the wavelet analysis approach proposed by Iwata et al. (2018), based on local scalar similarity and dissimilarity between CH₄ concentration and air temperature, respectively. The hourly time series of gap‐filled CO₂ and CH₄ fluxes (including the partitioned ebullitive and diffusive components) measured by the EC system were decomposed into contributions at different temporal scales using the maximal overlap discrete wavelet transform (MODWT) (e.g., Golub et al. 2023; Knox et al. 2021; Sturtevant et al. 2016). The MODWT was applied using a Symlet wavelet of order 8 (sym8) (Sturtevant et al. 2016). The wavelet coefficients were computed using the MATLAB function modwt, and the corresponding multiresolution analysis (MRA) components were reconstructed using modwtmra. This procedure decomposed the original signal into a set of scale‐dependent detail components and a residual smooth component, whose sum equals the original time series.

Given the hourly sampling interval, each MODWT level j corresponds to a specific range of temporal scales of 2^j h. We aggregated the reconstructed components into three broader time‐scale bands (e.g., Knox et al. 2021): (i) a diurnal component, corresponding to periods between approximately 8 and 32 h; (ii) a multiday component (synoptic to sub‐seasonal variability), corresponding to periods between approximately 3 and 43 days; (iii) a seasonal component, corresponding to periods longer than approximately 43 days. We performed aggregation by summing the reconstructed wavelet components whose characteristic periods fell within each band (Figures S3–S6). We calculated the variance of each aggregated component and expressed it as a fraction of the total variance of the original interpolated time series. In this manner, we could quantify the relative contribution of diurnal, multiday, and seasonal variability to the overall variance of CO₂ and CH₄ fluxes. Due to the redundancy and non‐orthogonality of the MODWT, as well as boundary effects, the sum of variances across multiresolution components does not necessarily equal the total variance; therefore, variance fractions should be interpreted as approximate contributions.

2.5. Attribution Study and Statistical Analysis

To attribute variability at each temporal scale to environmental and biogeochemical predictors, we fitted Random Forest (hereafter RF) regression models independently to the original fluxes and to each decomposed flux component. We used a common set of 15 potential predictor variables (Figure S7; see Section 2.2 for measurement resolution), including air temperature (T _air), near‐surface (T _surf) and bottom water temperature (T _bot), thermal stratification strength (ΔT = T _surf−T _bot), atmospheric pressure (P _atm), water column depth (Depth), the temporal gradient of total pressure (atmospheric plus hydrostatic) (ΔP _tot/Δt), wind speed, bottom‐water oxygen concentration (O_2,bot), surface chlorophyll‐a concentration (Chl‐a), inflow discharge as a surrogate for allochthonous nutrients and carbon inputs, incoming shortwave radiation (SWR), net surface buoyancy flux (B _0,net) as a precursor for convective turbulence, net ecosystem production (NEP) and gross primary production (GPP). NEP and GPP were derived from oxygen concentrations following the diel dissolved oxygen method (Staehr et al. 2010), based on a mass‐balance approach that partitions diel O₂ dynamics into biological production, respiration, and air–water gas exchange. Although Random Forests are relatively robust to predictor collinearity for prediction, strongly correlated variables can bias permutation‐based importance metrics; therefore, we reduced collinearity before fitting the model. We assessed collinearity among predictors using Spearman's rank correlations. Because air temperature and surface and bottom water temperatures were strongly correlated (|R| > 0.8, Figure S8), only bottom temperature was retained for CH₄ flux models and surface temperature for CO₂ flux models based on their expected physical relevance. SWR was also strongly correlated with B _0,net; therefore, B _0,net was retained given the known influence of convection on diffusive gas fluxes (e.g., Eugster et al. 2003; MacIntyre et al. 2010). Precipitation over the reservoir was not included as a predictor variable. Although rainfall events have been linked to short‐term variability in CH₄ ebullitive emissions in some systems (e.g., Niu et al. 2025), rainfall in Mediterranean regions typically occurs outside the summer high‐ebullition season.

We implemented RF models with identical hyperparameters across gases, years, and temporal scales, including a fixed number of trees (300) and permutation‐based out‐of‐bag (OOB) predictor importance. We fitted the models separately for each year to assess interannual variability in driver importance. We evaluated predictor importance using the increase in OOB prediction error following random permutation of each predictor, and we interpreted the results in terms of relative rankings and grouped process contributions rather than absolute importance values. To assess the robustness of predictor importance estimates to the stochastic nature of the algorithm (bootstrap resampling and random predictor selection), each analysis was repeated 50 times using different random seeds. The reported importance values correspond to the mean across runs. NEP and GPP were originally derived at daily resolution from dissolved oxygen measurements and linearly interpolated to the hourly time step to match the temporal resolution of the EC flux data. Chlorophyll‐a concentrations were available at weekly resolution until June 2024 and at daily resolution thereafter and were similarly interpolated to an hourly time step. We applied these interpolations solely to align predictors with the flux time series and did not introduce additional high‐frequency variability; consequently, the influence of these predictors is expected to be expressed primarily at multiday to seasonal time scales. The remaining predictors were collected at sub‐hourly resolution (see Section 2.2) and aggregated to hourly resolution by averaging.

To test for differences between daytime and nighttime conditions, we analyzed the original hourly methane (CH₄) and carbon dioxide (CO₂) fluxes. We defined daytime and nighttime using photosynthetically active radiation (PAR): hours with PAR ≥ 10 μmol photons m⁻² s⁻¹ were classified as daytime, and hours with PAR < 10 μmol photons m⁻² s⁻¹ as nighttime. To account for the temporal autocorrelation of hourly measurements, fluxes were aggregated to daily mean values for daytime and nighttime periods. This produced a dataset with one daytime and one nighttime flux estimate per day. We tested differences between daytime and nighttime fluxes using a linear mixed‐effects model, with daytime/nighttime as a fixed effect and day as a random intercept to account for repeated measurements within days. We fitted models using restricted maximum likelihood (REML). Statistical significance of diel differences was assessed using the fixed‐effect coefficient for day‐night, with α = 0.05. Conclusions were not sensitive to moderate changes in the PAR threshold used to define daytime conditions.

3. Results

3.1. Methane and Carbon Dioxide Emissions at Different Temporal Scales

The reservoir acted as a net source of both CO₂ and CH₄ independently of the year considered (wet/dry). Daily carbon fluxes varied markedly between years, with the greatest interannual differences observed in CH₄ fluxes (Figure 2). During the dry year, CH₄ fluxes averaged 16.83 ± 20.48 mmol m⁻² d⁻¹ and exhibited significant seasonal changes, with minimum fluxes on the order of 10⁻¹ mmol m⁻² d⁻¹ in winter and early spring, and maximum fluxes reaching up to 73 mmol m⁻² d⁻¹ in summer. In the wet year, mean CH₄ fluxes decreased to 11.44 ± 20.11 mmol m⁻² d⁻¹ while maintaining a similar seasonal range of variability. This reduction in average fluxes during the wet year is primarily attributable to a delayed onset of the summer high‐flux period, which began on June 28—almost 2 months later than in the dry year, when it started on May 1 (Figure 2a). As a result, the cumulative annual flux in the dry year (5.9 mol m⁻² year⁻¹) was 44% higher than that in the wet year (4.1 mol m⁻² year⁻¹) (Figure 2b). Assuming that the EC footprint is representative of the entire reservoir, cumulative CH₄ emissions amounted to 8.05 × 10⁶ and 5.6 × 10⁶ mol year⁻¹ during the dry and wet years, respectively (Figure S9). Thus, reservoir‐wide emissions were also 44% higher during the dry year, despite seasonal differences in reservoir surface area of up to 20 ha. Ebullition dominated over diffusion, accounting for 70% and 75% of the cumulative areal fluxes in the dry and wet years, respectively (Figure 2b).

FIGURE 2.

Time and cumulative series of CH₄ and CO₂ fluxes. EC diurnal averaged (a, c) and 30‐min cumulative (b, d) gap‐filled CH₄ (a, b) and CO₂ (c, d) fluxes for the two studied years. Cumulative fluxes in (b) also include the partition of the total flux into ebullition and diffusion. Shaded areas in (a, c) indicate ±1 standard deviation. X‐axis ticks mark the first day of each month.

CO₂ fluxes exhibited weaker seasonality than CH₄ fluxes. During the dry year, CO₂ fluxes averaged 40.95 ± 57.18 mmol m⁻² d⁻¹, with maximum values of 166 mmol m⁻² d⁻¹ occurring in late spring (Figure 2c). In the wet year, the seasonal maximum occurred earlier, in early spring. However, no significant interannual variability was observed, as mean fluxes were similar (41.06 ± 54.15 mmol m⁻² d⁻¹, or 14.9 mol m⁻² year⁻¹; see Figure 2d). Reservoir‐wide cumulative CO₂ emissions amounted to 20.54 × 10⁶ and 22.65 × 10⁶ mol year⁻¹ during the dry and wet years, respectively (Figure S9), indicating that total CO₂ emissions were approximately 10% higher during the wet year due to the larger reservoir surface area.

Seasonality dominates the temporal variability of CH₄ fluxes, with the decomposed seasonal component explaining more than 75% of the variance in the original signal in both years (Figure 3). Diurnal variability is the second most important contributor, accounting for 14% and 9% of the total variance during the dry and wet years, respectively (Figure 3). Variability at the multiday scale explains less than 6% of the total variance in both years. In contrast, diurnal variability dominates the temporal variability of CO₂ fluxes (Figure 3), with the decomposed diurnal component explaining 42% and 48% of the total variance in the original signal during the dry and wet years, respectively. The contribution of the multiday scale differed between years, explaining 27% of the variance in the dry year but only 17% in the wet year. Seasonality accounted for a smaller share, explaining 11%–13% of the total variance.

FIGURE 3.

Contribution of the different temporal scales to the total variance of the hourly gas flux signals. Results derived from the discrete wavelet transform and multiresolution analysis of the hourly gap‐filled CH₄ and CO₂ flux time series.

3.2. Predictor Variables for CH₄ and CO₂ Fluxes

In both years, bottom water temperature (T _bot), the temporal gradient in total pressure (ΔP _tot/Δt), wind speed, and water column depth were the four major predictors of the original CH₄ signal (R ² > 0.95; Figure 4a; Figures S10–S11), in the RF analysis. During the wet year, wind speed surpassed bottom water temperature in importance, probably due to the shorter duration of the summer high‐flux (ebullitive) period compared to that of the dry year (Figure 2a). When the signal was decomposed into its temporal components, T _bot clearly emerged as the dominant driver of seasonality, followed by water column depth (R ² ≥ 0.999; Figure 4b; Figures S10–S11). Multiday variability was well explained by variations in T _bot and water column depth (R ² ≥ 0.970; Figure 4c; Figures S10–S11). Finally, diurnal variability was best predicted by wind speed, T _bot, ΔP _tot/Δt, and additional ancillary variables, with some differences between years (R ² > 0.88; Figure 4d; Figures S10–S11).

FIGURE 4.

Random forest predictor importance for CH₄ (a–d) and CO₂ (e–h) fluxes during the dry and wet years. Random Forest regression models were applied separately to the original flux time series (a, e) and to the decomposed flux components obtained via maximal overlap discrete wavelet transform (MODWT) and multiresolution analysis: seasonal (b, f), multiday (c, g), and diurnal (d, h) signals. Each Random Forest analysis was repeated 50 times using different random seeds to assess model stochasticity. Bars represent mean predictor importance across runs, and black horizontal lines indicate ±1 standard deviation.

For CO₂ fluxes, wind speed was the main RF predictor of both the original CO₂ signal and the decomposed diurnal components in both years (Figure 4e,h), explaining between 40% and 53% of the variability in these signals (Table S1). Seasonal variability was primarily explained by water column depth and surface temperature T _surf (R ² > 0.999; Figure 4f; Figures S12–S13). Water column depth alone explained 75% and 97% of the variability in the seasonal CO₂ flux signals during the dry and wet years, respectively (Table S1). Multiday variability was explained by temporal variations in water column depth, T _surf, Chl‐a concentrations, and GPP (R ² > 0.970). Water column depth, together with Chl‐a, was more influential during the dry year, whereas Chl‐a, GPP, and T _surf better predicted multiday variability during the wet year (Figure 4g).

Bottom water temperature alone explained more than 80% and more than 90% of the variability in the original and seasonal CH₄ flux signals, respectively (Table S1). The sensitivity of methane emissions to temperature can be described using an exponential Arrhenius‐type function (e.g., Yvon‐Durocher et al. 2014):

$$F_{{\mathrm{CH}}_4}\left(T\right)=F_{{\mathrm{CH}}_4-20°\mathrm{C}}{e}^{\left(\frac{E_A}{k_b}\left(\frac{1}{293.15}-\frac{1}{T}\right)\right)}$$ (1)

where $F_{{\mathrm{CH}}_4}\left(T\right)$ is the methane flux at temperature T in K, $F_{{\mathrm{CH}}_4-20°\mathrm{C}}$ is the methane flux at 20°C, E _A (eV) is the activation energy, and k _b is the Boltzmann constant (8.62 × 10⁻⁵ eV K⁻¹). The activation energy was estimated by fitting Equation (1) to the daily‐averaged CH₄ flux measurements from both years combined (Figure 5a), yielding E _A  = 1.89 ± 0.01 eV (R ² = 0.9989). The sensitivity to water temperature was approximately 70% higher during the wet year (E _A  = 2.60 ± 0.02 eV) than during the dry year (E _A  = 1.53 ± 0.02 eV). When separating ebullition and diffusion, the fitted activation energy for the ebullitive fluxes was E _A  = 2.02 ± 0.02 eV (R ² = 0.9979), twice the activation energy for the diffusive fluxes (E _A  = 1.00 ± 0.01 eV, R ² = 0.9980; Figure S14). The seasonal pattern exhibited hysteresis in the relationship between water temperature and methane emissions, with higher emissions at a given temperature during the second half of the year (Figure 5a). This hysteretic behavior is explained by the second most important driver, the temporal evolution of water column depth (Figure 4b): as water depth—and thus hydrostatic pressure—decreases over the course of the summer season, the release of methane as bubbles is facilitated (Figure 5a). Chlorophyll‐a (Chl‐a) was the third most important seasonal predictor in the RF analysis (Figure 4b).

FIGURE 5.

Hysteretic relationship between water temperature and methane fluxes, and diurnal decomposed ebullitive and diffusive signals. (a) Bottom water temperature vs. methane flux. Red and blue dots represent the original daily‐averaged methane fluxes plotted against the measured daily‐averaged water temperature at 0.5 m above the lake sediment. The filled scattered circles indicate the relationship between the decomposed seasonal bottom‐water temperature T _bot and methane flux signals, with colors denoting the corresponding seasonally decomposed water‐column depth at the EC platform location. For visualization purposes, the seasonal relationships for the wet year are shown in a separate inset. Black arrows indicate the counterclockwise direction of the hysteresis over the annual cycle. The black dashed line shows the best‐fit Arrhenius relationship based on the two‐year dataset, with the corresponding activation energy (E _A ) indicated. The gray shaded area around the Arrhenius fit represents the 95% confidence interval. (b) Diurnal decomposed ebullitive signal and the ΔP _tot/Δt signal, and (c) diurnal decomposed diffusive signal and wind speed for a selected summer period during the dry year. Vertical red lines in (b,c) indicate the time of flux peaks. Time is in UTC.

The magnitude of the diurnal component of the ebullitive fluxes varied with ΔP _tot /Δt (Figure 5b; Figure S15b), with higher ebullition rates occurring during rapid decreases in total pressure in the afternoon. The diurnal cycle commonly exhibited a bimodal pattern, with a secondary peak in the early morning, corresponding to a secondary drop in total pressure. In contrast, the decomposed diurnal pattern of diffusive methane fluxes followed the diurnal cycle of wind speed over the reservoir, with peak values occurring in the afternoon (Figure 5c; Figure S15a). The linear mixed‐effects model applied to year‐round original hourly methane fluxes, using daytime/nighttime as a fixed effect, indicated that methane fluxes were significantly higher during the daytime. On average, daytime fluxes were 50% and 20% higher than nighttime fluxes during the dry and wet years, respectively (Table 1; Figure S16).

TABLE 1.

Linear mixed‐effects model results for day‐night differences in CH₄ and CO₂ fluxes.

Year	Gas	Effect	Estimate a (μmol m−2 s−1)	95% CI (μmol m−2 s−1)	p	% change a
2022	CH4	Day vs. Night	0.0759	[0.065, 0.0867]	2.84e‐38 c	49
2022	CO2	Day vs. Night	0.0610	[0.0243, 0.0977]	0.001 c	14
2024	CH4	Day vs. Night	0.0240	[0.0165, 0.0314]	5.30e‐10 c	20
2024	CO2	Day vs. Night	−0.0388	[−0.077, −0.0006]	0.046 b	−8

^a The estimate indicates the difference between daytime and nighttime fluxes. The percentage change measures the differences between the mean daytime and nighttime fluxes divided by the mean nighttime fluxes.

^b Significant at 95% level.

^c Significant at 99% level.

Unlike CH₄ fluxes, CO₂ fluxes showed a weak dependence on temperature, with low activation energies of ≈0.03 eV (Figure 6a,b). During the dry year, the decomposed seasonal CO₂ flux signal decoupled from the seasonal surface water temperature signal (Figure 6c), and a clear hysteretic pattern emerged in the relationship between water temperature and CO₂ fluxes (Figure 6b), with lower emissions at a given temperature during the second half of the year as the reservoir water level decreased. As with diffusive methane fluxes, the decomposed diurnal pattern of CO₂ fluxes showed a clear synchronous relationship with wind speed (Figure 6d). The linear mixed‐effects model applied to year‐round hourly CO₂ fluxes, with daytime/nighttime included as a fixed effect, indicated contrasting patterns between years. During the dry year, CO₂ fluxes were significantly higher during the daytime, by an average of 14%. In contrast, during the wet year, CO₂ fluxes were significantly lower during the daytime, by an average of 8% (Table 1; Figure S16).

FIGURE 6.

Sensitivity of CO₂ fluxes to temperature and diurnal emissions. (a,b) Surface water temperature versus CO₂ flux during the (a) dry and (b) wet years. Red and blue dots represent the original daily‐averaged CO₂ fluxes plotted against the measured diurnal‐averaged water temperature at 0.5 m below the lake surface. Filled scatter symbols show the relationship between the seasonally decomposed surface water temperature (T _surf) and CO₂ flux signals, with colors indicating the corresponding seasonally decomposed water‐column depth at the EC platform location. Black arrows in (b) indicate the clockwise direction of hysteresis over the annual cycle. The black dashed lines in (a, b) show the best‐fit Arrhenius relationships for each year, with the corresponding activation energies (E _A ) indicated. The gray shaded areas around the Arrhenius fits represent the 95% confidence intervals. (c) Seasonally decomposed T _surf and CO₂ flux signals. (d) Diurnal decomposed CO₂ flux signal and wind speed for a selected summer period during the wet year. Vertical blue lines indicate the timing of flux peaks. Time is in UTC.

4. Discussion

Average methane emissions from Cubillas fall within the range reported for tropical reservoirs (e.g., Prairie et al. 2021) and lie within the upper decile of emission rates reported for reservoirs globally (e.g., Prairie et al. 2021; Deemer et al. 2016). Strong seasonal patterns in methane fluxes as found in Cubillas have been observed in shallow temperate, boreal, and subarctic lakes and reservoirs (e.g., Waldo et al. 2021; Jansen et al. 2020; Podgrajsek et al. 2016), as well as in wetlands (e.g., Knox et al. 2019; Sturtevant et al. 2016). Seasonal dependencies between methane fluxes and air, water, or sediment temperature have also been documented in reservoirs (e.g., Kim et al. 2025), lakes (e.g., Jansen et al. 2020; Jammet et al. 2017; Wik et al. 2014), and wetlands (e.g., Knox et al. 2021) worldwide. These patterns ultimately reflect the temperature dependence of methanogens' metabolic activity during the decomposition of organic matter in anoxic sediments (e.g., Yvon‐Durocher et al. 2014). The relationship between CH₄ emissions and temperature is commonly exponential and described using an Arrhenius‐type equation (e.g., Iwata et al. 2020; Jansen et al. 2020; Natchimuthu et al. 2016). The exponential temperature dependence observed at Cubillas is consistent with previous year‐round measurements in the reservoir using a flux chamber coupled to a gas analyzer (Rodríguez‐Velasco et al. 2024), and the estimated activation energies for ebullition (E _A  ≈2 eV) and diffusion (E _A  ≈1 eV) fall within the range of values reported in earlier studies (Praetzel et al. 2021; Jansen et al. 2020; Aben et al. 2017; DelSontro et al. 2016).

Here, we further show that CH₄ emissions exhibit a pronounced hysteretic pattern, with higher emissions at a given temperature during the second half of the year (Figure 5a). Such counterclockwise hysteretic behavior is commonly observed in wetlands (Chang et al. 2021), but has been rarely reported in lakes or reservoirs (Jansen et al. 2020). This hysteretic pattern has been attributed to seasonal differences in microbial abundance and/or substrate availability and characteristics in the sediment (e.g., Chang et al. 2020; Updegraff et al. 1998), both of which influence methane production. In Cubillas, chlorophyll‐a (Chl‐a) arose as the third most important predictor of the seasonal component of CH₄ fluxes in the dry year. Chl‐a concentrations increase as the summer season progresses, potentially enhancing CH₄ production as this readily degradable material is deposited in the sediments (e.g., Martínez‐García et al. 2024; Grasset et al. 2018; West et al. 2015, 2012), thereby reinforcing the hysteretic effect. However, water column depth—and thus hydrostatic pressure—arises in our RF analysis as the major driver controlling CH₄ flux hysteresis in Cubillas. Together, near‐sediment temperature and water column depth explain more than 99% of the seasonal variability in CH₄ fluxes. Reduced hydrostatic pressure lowers the threshold for bubble formation and release, increasing the contribution of ebullition to the overall fluxes, which also bypasses water‐column oxidation. Consequently, methane fluxes are higher at a given temperature later in the season under shallower water depths, producing the observed hysteresis in the methane flux‐temperature relationship. By showing a linear increase in CH₄ fluxes with decreasing hydrostatic pressure, other authors (e.g., Iwata et al. 2020; Deshmukh et al. 2014) have suggested that CH₄ bubbling sites are seasonally controlled by hydrostatic pressure. In Cubillas, CH₄ fluxes are primarily seasonally controlled by temperature; however, during periods of declining water level, a linear relationship is still observed between CH₄ fluxes and hydrostatic pressure (Figure S17). Similarly, temperature and depth continue to modulate CH₄ emissions at the multiday scale, consistent with observations in other systems (Hounshell et al. 2023; Knox et al. 2021).

The diurnal variability of CH₄ fluxes in Cubillas is better predicted by wind speed, near‐sediment temperature, and the temporal gradient of total pressure (ΔP _tot/Δt). Wind‐driven turbulence enhances gas transfer velocity at the air–water interface, controlling diffusive CH₄ fluxes (e.g., Lorke and Peeters 2006; Crusius and Wanninkhof 2003). Accordingly, CH₄ diffusive fluxes in Cubillas vary synchronously with wind forcing at the diurnal scale (Figure 5c; León‐Palmero et al. 2025). Surface cooling can generate convective turbulence that also affects gas exchange (e.g., Tedford et al. 2014; Read et al. 2012; Soloviev et al. 2007), although RF analysis indicates that, on an annual basis, buoyancy‐driven convection plays a secondary role overall in Cubillas, in contrast to other systems (e.g., Podgrajsek et al. 2014). Wind can also induce bottom currents that enhance near‐bed shear stress, thereby favoring the release of gas bubbles from the sediment (Joyce and Jewell 2003). Temporal decreases in total pressure are well‐known to control ebullition (e.g., Varadharajan and Hemond 2012; Mattson and Likens 1990), which in Cubillas exhibits a clear diurnal bimodal pattern consistent with semidiurnal atmospheric pressure oscillations (Figure 5b), and similar to that reported by Deshmukh et al. (2014) in a tropical hydroelectric reservoir. Near‐sediment temperatures, pressure drops, and wind speed tend to peak during daytime hours, contributing to higher daytime CH₄ fluxes (by ~ 50% or ~ 20%, depending on the year; Table 1; Figure S16), consistent with daytime‐elevated methane emissions reported for four lakes by Sieczko et al. (2020), where higher daytime wind speeds were identified as the dominant driver.

Average annual CO₂ emissions in Cubillas fall within the wide range reported for temperate (10¹–10³ mgC m⁻² d⁻¹ in absolute values, e.g., Prairie et al. 2021) and Mediterranean (León‐Palmero, Morales‐Baquero, and Reche 2020) reservoirs, with diurnal variability dominating total flux variance. CO₂ fluxes in Cubillas fall within Cluster A identified by de Eyto et al. (2025) from 41 diel campaigns across 21 lakes, characterized by afternoon maxima and dawn minima, and primarily driven by wind forcing. The predominance of wind forcing as a driver for CO₂ fluxes is also consistent with previous flux‐chamber measurements in the study reservoir (León‐Palmero et al. 2025; Rodríguez‐Velasco et al. 2024) and other year‐long EC studies in lakes and reservoirs (e.g., Golub et al. 2023). Multiday CO₂ flux variability was primarily associated with water depth, temperature, chlorophyll‐a concentration, and GPP, highlighting the combined influence of hydrological conditions and ecosystem metabolism on CO₂ dynamics. Air and water temperatures are common predictors of CO₂ fluxes across multiple temporal scales in lakes and reservoirs (Rodríguez‐Velasco et al. 2024; Golub et al. 2023; Reed et al. 2018; Ouyang et al. 2017). Temperature likely controls short‐term respiration and organic carbon mineralization rates, while depth affects how CO₂ and dissolved inorganic carbon are distributed and buffered within the water column. Chlorophyll‐a, as a proxy for phytoplankton biomass, influences CO₂ fluxes through photosynthetic uptake, respiration, and the supply of labile organic carbon for microbial processing. The influence of chlorophyll‐a concentration at multiday timescales is consistent with previous studies; for example, Ouyang et al. (2017) and Shao et al. (2015) reported negative correlations between monthly CO₂ fluxes and chlorophyll‐a concentrations in a western bay of Lake Erie. The emergence of GPP as the second major predictor of multiday variability in the wet year aligns with previous findings in aquatic systems, where air–water CO₂ fluxes track short‐term variations in primary production rather than standing biomass (e.g., Wang et al. 2026).

Inflow discharge was not a significant predictor of seasonal CO₂ flux variability, suggesting that short‐term external carbon inputs are not the dominant control on seasonal emissions. Instead, water depth—the major predictor of the seasonal component of the CO₂ flux—likely integrates the cumulative effects of hydrology and carbon storage through pelagic respiration, while surface water temperature, the second most important predictor, regulates internal metabolic processing and CO₂ availability. During the wet year, the reservoir rewetted areas that had been dry for more than 1 year, as maximum water levels declined from 2021 to 2023 (Figure 1a). For example, 5% of the in‐reservoir 90% EC footprint during the wet year corresponded to areas that were rewetted above the maximum water level reached in the previous year. Rewetting of previously dry areas may have promoted higher dissolved CO₂ concentrations within the EC footprint, either through the presence of rewetted areas within the footprint itself or through enhanced advection of CO₂‐rich water from other rewetted reservoir locations. These processes likely contributed to the early‐season peak in CO₂ fluxes observed during the wet year by increasing substrate availability and microbial activation in already warm (T _surf > 16°C) rewetted areas (the “Birch effect”, Amaral et al. 2022; Jarvis et al. 2007; Birch 1958). For example, rain events have been shown to cause at least a 5‐fold increase in CO₂ fluxes compared to pre‐rainfall values in boreal forests (Makhnykina et al. 2024). Hysteresis between CO₂ fluxes and surface water temperature during the wet year may have been further reinforced by the rapid summer decline in water level (4.5 m from June to September), leading to the flushing of water, nutrients, and carbon. Consequently, CO₂ fluxes exhibited hysteresis with respect to surface water temperature, with higher emissions at a given temperature during the rising (rewetting) phase than during the falling (drawdown) phase, reflecting temporal changes in carbon availability rather than temperature alone. Additional measurements are needed to confirm this mechanism.

Although RF analysis does not explicitly account for dependencies among drivers or assign causation, we are confident that our conclusions regarding the key role of hydrological and thermal variability controlling GHG emissions from shallow eutrophic Mediterranean reservoirs are robust and consistent with our mechanistic understanding of GHG dynamics. Interannual differences in CH₄ emissions were primarily driven by changes in water level and hydrostatic pressure, which regulated the timing and intensity of the high‐emission season via effects on ebullition and, in the case of water level, thermal stratification and sediment temperature. Lower water column depth and an earlier drawdown during the dry year led to substantially higher cumulative methane emissions (44% higher), highlighting how projected increases in drought frequency and altered hydrological regimes may enhance reservoir CH₄ emissions. Water level dynamics are also influenced by reservoir operation, so management strategies may partly modulate these responses, although reduced summer withdrawals during drought may be unfeasible in water‐limited Mediterranean regions. Drought‐induced changes in reservoir surface area must also be considered when scaling GHG emissions at the reservoir level. In contrast to methane, CO₂ fluxes were less sensitive to interannual hydrological variability and were dominated by short‐term physical drivers, particularly wind speed, although seasonal water‐level changes still influenced CO₂ emissions, emphasizing how water‐column depth affects carbon storage and subsequent CO₂ release. Finally, chlorophyll‐a–linked seasonal and multiday variability in CH₄ and CO₂ fluxes indicates that eutrophication enhances GHG emissions by boosting primary production and labile sedimentary carbon supply, reinforcing methane hysteresis and ebullitive temperature sensitivity.

Mediterranean‐climate regions are among the most densely populated and water‐stressed globally, relying on reservoirs to buffer seasonal and interannual water scarcity. The Mediterranean region is also a climate change hotspot, warming at rates that exceed global averages (e.g., Hassoun et al. 2025; Pausas and Millán 2019), but similar highly variable hydrological conditions are increasingly observed in other regions, with substantial portions of the global land surface transitioning toward drier aridity classes (Crapart et al. 2026). Our findings show that the mechanisms driving the observed interannual variability in CH₄ emissions are broadly applicable to managed freshwater systems and relevant to regions experiencing increasing aridity. Importantly, we observed a clear decoupling in GHG dynamics, whereby drought conditions disproportionately enhanced emissions of the high‐radiative‐forcing gas CH₄ while having little effect on CO₂ fluxes, thereby substantially amplifying the warming impact per unit of stored water. By relying on averaged or static hydrology, global upscaling approaches may miss interannual variability, adding uncertainty to estimates of freshwater–climate feedbacks. Although the limited range of hydrological conditions examined here precludes establishing robust quantitative relationships, our findings indicate that future changes in reservoir GHG emissions will depend on the combined effects of climate‐driven and anthropogenic‐driven hydrological changes alongside increasing nutrient enrichment. This underscores the need for continued long‐term, high‐frequency measurements across a broader range of hydrological conditions to improve emission estimates and projections of climate feedbacks from managed freshwater systems.

Author Contributions

Cintia L. Ramón: conceptualization, data curation, formal analysis, funding acquisition, investigation, methodology, project administration, resources, software, validation, visualization, writing – original draft. Isabel Reche: conceptualization, funding acquisition, investigation, project administration, resources, supervision, writing – original draft, writing – review and editing. Rodrigo J. Gonçalves: data curation, investigation, software, writing – review and editing. Andrés Martínez‐García: data curation, investigation, writing – review and editing. Miriam García‐Alguacil: data curation, investigation, writing – review and editing. Sergio López‐Padilla: data curation, investigation, writing – review and editing. Javier Herrero: data curation, investigation, writing – review and editing. Rafael Morales‐Baquero: investigation, writing – review and editing. Enrique P. Sánchez‐Cañete: methodology, software. Alicia Cortés: investigation, writing – review and editing. Sara Valiente: data curation, investigation, writing – review and editing. Francisco J. Rueda: conceptualization, funding acquisition, investigation, project administration, resources, writing – review and editing.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Annual footprints of the eddy covariance fluxes. Areas contributing to different percentages of the EC footprint centered on the EC tower location (colored lines), the bathymetry of the Cubillas Reservoir expressed as isolevels of water elevation (m.a.s.l.; gray color scale), and wind roses for (a) 2022 (dry year) and (b) 2024 (wet year). Thick and thin black contour lines indicate the minimum and maximum water levels reached in each year, respectively. Map lines delineate study areas and do not necessarily depict accepted national boundaries.

Figure S2: Rotated vs non‐rotated CH₄ and CO₂ fluxes. Comparison of 30‐min fluxes calculated with and without accounting for pitch and roll rotation of the floating platform. The black solid line indicates the 1:1 relationship. Root mean square errors are shown at the top of each panel. Differences between rotated and non‐rotated fluxes were tested using a linear mixed‐effects model, with rotation included as a fixed effect and day as a random intercept. No significant differences were detected between the two signals (p > 0.05).

Figure S3: Gap‐filled hourly time series of total CH₄ fluxes and their decomposed seasonal, multiday, and diurnal components.

Figure S4: Gap‐filled hourly time series of ebullitive CH₄ fluxes and their decomposed seasonal, multiday, and diurnal components.

Figure S5: Gap‐filled hourly time series of diffusive CH₄ fluxes and their decomposed seasonal, multiday, and diurnal components.

Figure S6: Gap‐filled hourly time series of CO₂ fluxes and their decomposed seasonal, multiday, and diurnal components.

Figure S7: Variables used as predictors in the random forest analysis. Variables: T _air = Air temperature (°C). T _surf = Surface temperature (°C), 0.5 m below the surface. T _bot = Bottom temperature (°C), 0.5 m above the sediment. ΔT = Stratification strength (°C) (=T _surf−T _bot). P _atm = Atmospheric pressure (hPa). Depth = Water column depth (m). ΔP _tot/Δt = Gradient of total pressure (atmospheric + hydrostatic) (Pa h⁻¹). Wind = Wind speed (m s⁻¹). O _2,bot  = Oxygen concentration 0.5 m above the sediment (mg L⁻¹). Chl‐a = Chlorophyll‐a (μg L⁻¹). NEP = Net ecosystem production (gO₂ m⁻³ d⁻¹). GPP = Gross primary production (gO₂ m⁻³ d⁻¹). Inflow = Inflow discharge (m³ s⁻¹). B _0,net = Net surface buoyancy flux (W kg⁻¹).

Figure S8: Spearman's correlation coefficients for predictor variables for Year 2022 (dry year). Statistical significance: *p < 0.05; **p < 0.01; ***p < 0.001. Variables: T _air = Air temperature. T _surf = Surface temperature (°C), 0.5 m below the surface. T _bot = Bottom temperature (°C), 0.5 m above the sediment. ΔT = Stratification strength (°C) (= T _surf−T _bot). P _atm = Atmospheric pressure (Pa). Depth = Water column depth (m). ΔP _tot/Δt = Gradient of total pressure (atmospheric + hydrostatic) (Pa h⁻¹). Wind = Wind speed (m s⁻¹). O _2,bot = Oxygen concentration 0.5 m above the sediment (mg L⁻¹). Chl‐a = Chlorophyll‐a (μg L⁻¹). NEP = Net ecosystem production (gO₂ m⁻³ d⁻¹). GPP = Gross primary production (gO₂ m⁻³ d⁻¹). Inflow = Inflow discharge (m³ s⁻¹). B _0,net  = Net surface buoyancy flux (W kg⁻¹). SWR = shortwave radiation (W m⁻²).

Figure S9: Reservoir‐wide total emissions. (a, b) Reservoir‐wide cumulative (a) CH₄ and (b) CO₂ emissions, expressed in moles and assuming that the EC footprint is representative of emissions from the entire reservoir, and (c) reservoir surface area during the wet and dry years.

Figure S10: Random forest predictions for the original and decomposed CH₄ flux signals during the dry year. Panels show 1:1 relationships (a, c, e, g) and corresponding time series (b, d, f, h) for the observed signals and the Random Forest (RF)–predicted signals. Black lines in panels (a, c, e, g) indicate the 1:1 relationship. For each predicted signal, the number of predictors (N), the selected predictors, and the coefficient of determination (R ²) are shown. The five most important predictors were initially selected based on their permutation‐based out‐of‐bag (OOB) importance, and further predictor reduction was applied when R ² exceeded 0.95 using a smaller subset of predictors.

Figure S11: Random Forest predictions for the original and decomposed CH₄ flux signals during the wet year. Panels show 1:1 relationships (a, c, e, g) and corresponding time series (b, d, f, h) for the observed signals and the Random Forest (RF)–predicted signals. Black lines in panels (a, c, e, g) indicate the 1:1 relationship. For each predicted signal, the number of predictors (N), the selected predictors, and the coefficient of determination (R ²) are shown. The five most important predictors were initially selected based on their permutation‐based out‐of‐bag (OOB) importance, and further predictor reduction was applied when R ² exceeded 0.95 using a smaller subset of predictors.

Figure S12: Random Forest predictions for the original and decomposed CO₂ flux signals during the dry year. Panels show 1:1 relationships (a, c, e, g) and corresponding time series (b, d, f, h) for the observed signals and the Random Forest (RF)–predicted signals. Black lines in panels (a, c, e, g) indicate the 1:1 relationship. For each predicted signal, the number of predictors (N), the selected predictors, and the coefficient of determination (R ²) are shown. The five most important predictors were initially selected based on their permutation‐based out‐of‐bag (OOB) importance, and further predictor reduction was applied when R ² exceeded 0.95 using a smaller subset of predictors.

Figure S13: Random forest predictions for the original and decomposed CO₂ flux signals during the wet year. Panels show 1:1 relationships (a, c, e, g) and corresponding time series (b, d, f, h) for the observed signals and the Random Forest (RF)–predicted signals. Black lines in panels (a, c, e, g) indicate the 1:1 relationship. For each predicted signal, the number of predictors (N), the selected predictors, and the coefficient of determination (R ²) are shown. The five most important predictors were initially selected based on their permutation‐based out‐of‐bag (OOB) importance, and further predictor reduction was applied when R ² exceeded 0.95 using a smaller subset of predictors.

Figure S14: Bottom water temperature vs. methane ebullitive and diffusive fluxes. The black dashed line shows the best 2‐year Arrhenius fit [Equation (1)], with the corresponding activation energy (E _A ) indicated. The gray shaded area around each Arrhenius fit represents the 95% confidence interval.

Figure S15: Random forest predictor importance for the decomposed diurnal diffusive (a) and ebullitive (b) CH₄ fluxes. Diffusive and ebullitive components were extracted from the total eddy covariance CH₄ fluxes using the wavelet analysis approach proposed by Iwata et al. (2018). The decomposed temporal flux components were obtained using the maximal overlap discrete wavelet transform (MODWT) and multiresolution analysis. Each Random Forest analysis was repeated 50 times using different random seeds to assess model stochasticity. Bars represent mean predictor importance across runs, and black horizontal lines indicate ±1 standard deviation.

Figure S16: Daytime versus nighttime fluxes. Daily‐averaged gap‐filled CH₄ (a, b) and CO₂ (c, d) fluxes during daytime (PAR ≥ 10 μmol photons m⁻² s⁻¹) and nighttime (PAR < 10 μmol photons m⁻² s⁻¹) conditions for the dry year (a, c) and the wet year (b, d).

Figure S17: Hydrostatic pressure versus methane fluxes. Relationship between daily‐averaged hydrostatic pressure at the lake sediment surface and methane fluxes. Solid gray lines show the seasonally decomposed relationship between the two variables, and arrows indicate the direction of progression over the annual cycle. Linear regressions are shown for periods during which CH₄ fluxes increase linearly with decreasing hydrostatic pressure (i.e., water depth). The gray shaded area around each regression represents the 95% confidence interval.

Table S1: Coefficient of determination (R ²) for random forest predictions as the number of variables (within brackets) increases from 1 to 5, or until a combination reaching R ² > 0.95 is achieved. Variables 1–5 were selected based on their predicted importance (the most important first).

Data Availability Statement

The data supporting the findings of this study are openly available in https://doi.org/10.5281/zenodo.20800380.