The study site is a cropland field in Oensingen, canton of Solothurn, Switzerland. The site has a mean annual temperature of 10.1 °C and receives an average of 1151 mm of precipitation per year (2004–2023). CO₂ and water vapor (H₂O) fluxes have been continuously measured at the site by an EC station (CH-Oe2, part of FLUXNET and Swiss FluxNet; 47°1^′11.1^′′ N, 7°44^′01.5^′′ E; 452 m a.s.l.) since December 2003 (Emmel et al., 2018). The field where the station is positioned has an area of 1.55 ha, and the soil is classified as Eutric-Stagnic Cambisol with a silty clay texture. In the top 0.25 m, the soil consists of 43 % clay, 47 % silt, and 10 % sand, with 2.8 % organic matter (Alaoui and Goetz, 2008). In the top 0.1 m depth, soil pH is 6.4 (Maier et al., 2025), and bulk density is 1.27 g cm⁻³. Carbon and nitrogen stocks in the 0–0.7 m soil profile, derived from Emmel et al. (2018), are 18.6 kg C m⁻² and 1.9 kg N m⁻², respectively. The field is managed according to an integrated farming label (IP Suisse) and follows the Proof of Ecological Performance (PEP; Swiss Federal Council, 2025). PEP sets baseline requirements for receiving direct payments, including rules for crop rotation, farm-based fertilizer use, and measures to protect soil and biodiversity. Over the past 20 years, the crop rotation included various annual crops, with winter wheat as primary crop. A temporary grassland (grass-clover mixture; STEFFEN 3003M, Samen STEFFEN AG) was established in 2020 and managed until its final cut in late September 2022, followed by herbicide application (Glyphosate) and direct-seeding of winter wheat. This study was conducted over one year, from September 2022 to October 2023, covering the last month of the temporary grassland, the full winter wheat growing season, and the subsequent cover crop season, a mixture of legumes, grasses, and forbs (Terra-FIT Quattro, Samen STEFFEN AG). Winter wheat (Montalbano, bred by Agroscope and DSP) was sown on 6 October 2022, and harvested on 15 July 2023, with 5.03 t ha⁻¹ grain yield (dry matter). The winter wheat received three applications of mineral and organic (dairy slurry) fertilizers between late tillering and early stem elongation, summing up to 138 kg N ha⁻¹ (Table 1). After the harvest, the straw was removed from the field, and the soil was cultivated for the subsequent sowing of the summer cover crop. The cover crop was mulched in late September 2023, before the soil was cultivated again to prepare for the sowing of winter barley in early October 2023.

Cumulative N₂O, CH₄, and CO₂ fluxes were calculated separately for the winter wheat cropping season and the subsequent cover crop period using gap-filled flux datasets. The winter wheat cropping season was defined as the period from sowing to harvest, while the cover crop period was defined as the period from cover crop sowing to mulching. To express fluxes in terms of CO₂-equivalents, global warming potential (GWP₁₀₀) values of 273 for N₂O and 27 for CH₄ were applied, in accordance with current estimates (Forster et al., 2021). The sum of N₂O, CH₄ and CO₂, i.e., the cumulative fluxes expressed in CO₂-equivalents (CO₂-eq), is hereafter referred to as net GHG exchange (NGHGE). This term is used to distinguish the EC-based atmospheric exchange from the net GHG budget, which additionally accounts for lateral C imports and exports.

The net GHG budget was derived from the C budget, which combines NEE with lateral C fluxes from management (imports and exports). C exports comprised harvested grain and straw, and C imports comprised organic fertilizer additions. Exports were treated as positive fluxes (C removal from the ecosystem) and imports as negative fluxes (C addition). The C budget was converted to CO₂-eq using the molecular mass ratio of CO₂ to C, then N₂O and CH₄ fluxes (in g CO₂-eq) were added to obtain the net GHG budget, following Ammann et al. (2020). Negative values indicate a net GHG sink, positive values a net GHG source.

To assess the sensitivity of NGHGE to friction velocity (u*) filtering (Table A3), cumulative N₂O, CH₄ and CO₂ fluxes and the resulting NGHGE were calculated under alternative u* thresholds, corresponding to the 16th and 84th percentiles of the bootstrapped u* threshold distribution (CUT16 and CUT84, respectively), in addition to the default threshold (CUT50; median), following Pastorello et al. (2020).

The N₂O emission factor (EF) for the winter wheat season was calculated by dividing the cumulative N₂O-N emissions from sowing to harvest by the total amount of N from mineral and organic fertilizers. The resulting EF is expressed as the percentage of the applied N that is emitted as N₂O-N. Because no unfertilized control treatment was available, non-fertilization-induced emissions could not be subtracted. The resulting value, therefore, represents an apparent EF rather than an IPCC Tier 1 EF. To account for N₂O emissions beyond the winter wheat season, we additionally calculated an annual apparent EF for the one-year period starting at winter wheat sowing, including the bare-soil period after harvest and the subsequent cover crop period. This metric was calculated as cumulative annual N₂O-N emissions divided by the fertilizer N applied during the same period.

The meteorological variables air temperature and relative humidity (CS215, Campbell Scientific Ltd., Logan UT, USA), atmospheric pressure (PAA-33X, Keller AG, Switzerland), short- and longwave radiation (CNR4, Kipp & Zonen, Delft, Netherlands), photosynthetic photon flux density PPFD (BF5, Delta-T Devices, United Kingdom), and precipitation (heated tipping bucket rain gauge, Lambrecht meteo GmbH, Germany) were measured continuously at the EC station. Soil water content and soil temperature were also measured continuously close to the EC station along one soil profile (at 0.05, 0.15, 0.3, and 0.5 m soil depths) using 5TM sensors (Decagon Devices, Inc., Pullman WA, USA). All meteorological measurements were recorded at one-minute time resolution and, after screening for outliers, averaged to 30 min values (precipitation was summed). Water-filled pore space (WFPS; Eq. 1) was calculated for each soil depth as: 1WFPS=SWC1-BDPD⋅100 where SWC is the volumetric soil water content (in m³ m⁻³), BD is the bulk density (in g cm⁻³), with values of 1.16, 1.40, and 1.33 g cm⁻³ for the 0–0.05, 0.15–0.3, and 0.3–0.5 m depth intervals, respectively, obtained from Emmel et al. (2018), and PD is the particle density, assumed to be 2.65 g cm⁻³ (Danielson and Sutherland, 1986). Because SWC was measured at a single soil profile, absolute WFPS estimates may not fully represent spatial variability across the EC footprint. However, similar SWC dynamics observed at an independent sensor profile located a few meters away indicated that temporal patterns were representative. The WFPS conversion further relied on bulk density values from Emmel et al. (2018), which may not fully reflect topsoil conditions during this experiment. Therefore, uncertainty is expected mainly in absolute WFPS values rather than in temporal WFPS dynamics.

Five opaque static PVC (polyvinyl chloride) chambers, each with a volume of 17.7 L and a surface area of 0.07 m², were deployed at five randomly selected locations within a 12 m radius of the EC station (Fig. 1). The chambers were first deployed on 24 February 2023, before the first fertilization event, and positioned to enclose plants from two adjacent wheat rows. To ensure that regular field management could continue unhindered, the chambers were removed immediately after sampling when management operations were scheduled before the next sampling date and subsequently reinstalled at the same positions for the following measurements. In total, six sampling campaigns were conducted between 24 February and 17 March 2023, encompassing the period around the first fertilizations (27 and 28 February) and two campaigns later in the season (31 May and 19 July). The sampling was performed around midday (typically between 11:00 and 13:00 local time) to minimize diurnal variability in fluxes. For each chamber and time point, two headspace samples were collected: one immediately after chamber closure (t0), and one after a 1-hour enclosure period (t60). Air samples were drawn from the chamber headspace with a 60 mL syringe. For subsequent N₂O concentration analysis, 20 mL of the air sample was injected into pre-evacuated 12 mL vials. For stable isotope analyses, 180 mL of the air sample was injected into pre-evacuated 110 mL serum crimp vials. During each sampling campaign, a soil sample was collected for each chamber by combining two soil cores (0–0.1 m depth) taken approximately 0.3 m from the chamber edge. The pooled samples were placed into Schott tubes, which were immediately sealed with Parafilm, enclosed in plastic bags, and stored in a cooling box before being transferred to a freezer at -18 °C. Soil water was later extracted from the samples using cryogenic vacuum distillation (Ehleringer and Osmond, 2000). The stable oxygen isotope ratio (δ18O) of the extracted water was then measured using an isotope ratio mass spectrometer (IRMS; DeltaplusXP, Finnigan MAT, Bremen, Germany) and expressed in the delta notation referenced to the Vienna Standard Mean Ocean Water (V-SMOW) in ‰ (Eq. 2) as: 2δ18O=RsampleRreference-1 where R denotes the ratio of ¹⁸O / ¹⁶O.

N₂O gas samples were analyzed for their stable N and oxygen isotope ratios (δ15N^bulk, δ18O, respectively) as well as the site preference (SP-N₂O) using a trace gas preparation unit (Elementar, Manchester, UK) coupled to an IsoPrime100 IRMS (Elementar, Manchester, UK). Calibration and data correction followed the procedure described in Verhoeven et al. (2019) and in Gallarotti et al. (2021), including the use of multiple working standards and correction for instrumental drift, span, and linearity. In addition, instrument performance was regularly assessed using standards treated as samples (Werner and Brand, 2001), ensuring data quality and interlaboratory comparability (Mohn et al., 2014). Site preference (SP) represents the difference between the N isotope ratios at the central (α) and terminal (β) N positions of the N₂O molecule. Unlike δ15N^bulk, SP-N₂O is independent of the initial substrate, making it a more direct indicator of the N₂O production pathway (Toyoda et al., 2002). The isotopic signature of emitted N₂O was calculated using the following two-point mixing equation (Eq. 3): 3δZX=N2Oe⋅δZXe-N2Os⋅δZXsN2Oe-N2Os where [N₂O]_s and [N₂O]_e are the N₂O mixing ratios [ppb] at t0 and t60, respectively, measured using a gas chromatograph equipped with an electron capture detector (456-GC, Scion Instruments, Livingston, WLO, UK), while δZX_s and δZX_e are the corresponding isotope ratios (‰). A threshold of 30 ppb in N₂O concentration increase (Δ[N₂O] = [N₂O]_e - [N₂O]_s) was used to exclude flux estimates for which the change in N₂O concentration during the chamber closure was too small to yield reliable isotopic signatures (Harris et al., 2020b). N₂O fluxes on four of the eight sampling dates, 1, 3, 6, and 17 March, exceeded this threshold and were subsequently included in the analyses. Thus, isotope-based process identification was restricted to the three weeks following the first fertilization events at the end of February 2023 (Table A1).

To determine the main production pathways of N₂O, a dual isotope mapping approach was used, relating (i) δ18O-N₂O to SP-N₂O, and (ii) δ15N^bulk-N₂O to SP-N₂O. The mixing endmembers for nitrification (Ni), heterotrophic bacterial denitrification (bD), nitrifier denitrification (nD), and fungal denitrification (fD) were chosen according to Yu et al. (2020). The δ18O-N₂O endmembers of bD, nD, and fD were adjusted by the mean δ18O of soil water to account for the oxygen isotopic signature of the substrate (Lewicka-Szczebak et al., 2020). Since no measurements of the δ15N of soil N precursors were available, a δ15N value of 0 ‰ was assumed for precursor nitrate and ammonium in the δ15N^bulk-N₂O vs. SP-N₂O map. The contributions of different N₂O production pathways and the residual fraction of unreduced N₂O (r) were estimated with the FRAME software, implementing a Markov Chain Monte Carlo model (Lewicki et al., 2022). The FRAME model was configured using process-specific mean values and fractionation factors from Yu et al. (2020), and the δ18O-N₂O mean values were corrected for δ18O-H₂O. For each sample, the uncertainty of the isotopic signature of the emitted N₂O was obtained by Gaussian error propagation through Eq. (3), from the uncertainties of the four measured input quantities (δZX and [N₂O] at t0 and t60). As no replicate samples were available, conservative analytical precisions of 1 ‰ (1 SD) for the isotopic measurements and 2 % (1 SD, relative) for the N₂O concentrations were assumed. This assumed uncertainty for the isotopic measurements is larger than the analytical uncertainties typically reported for IRMS-based N₂O isotopocule analysis (Mohn et al., 2014). The resulting sample-specific uncertainties were then used as the measurement uncertainties in FRAME and are therefore reflected in the posterior distributions. FRAME was run using the mixing-reduction scenario, in which N₂O from different production pathways is mixed before partial reduction to N₂. Instead of the residual unreduced N₂O fraction (r), we report the reduced fraction (1-r), as it is more intuitive to interpret. To test sensitivity to the assumed order of mixing and reduction, the alternative reduction-mixing scenario was also tested, in which denitrification-derived N₂O is reduced before mixing with N₂O from the other pathways.

Leaf area index (LAI), vegetation height, and phenological development stages were monitored throughout the study. From March to July 2023, measurements were conducted at 2- to 3-week intervals to capture the growth dynamics during the most active vegetative period of winter wheat, and at key developmental stages before and after this period. LAI and canopy height were measured at six locations spread 5 m apart along four transects originating from the EC station (Fig. 1). LAI was measured using a LAI-2000 Plant Canopy Analyzer (LI-COR Biosciences, USA) by taking one above-canopy and six below-canopy readings for each transect. Under clear sky conditions, measurements were conducted towards the north to avoid direct sunlight on the lens. Vegetation height was determined by measuring the height of an expanded polystyrene plate of 0.25 m² dropped on top of the vegetation at each spot (Ammann et al., 2007). Phenological development stages of winter wheat were identified according to the BBCH-scale (Meier, 2018) for five plants randomly selected in the field.

Aboveground biomass (AGB) was sampled monthly, using a harvest frame with a defined area of 0.1 m². Three samples were randomly collected at each of the four defined field sections (Fig. 1), and entire wheat plants were cut 2 cm above the ground. At harvest, grain and straw were sampled separately to estimate C and N exports from the field. The dry weight was determined after drying all plant samples at 60 °C for 48 h. The dried samples were further processed by first cutting the biomass into smaller pieces, then a subsample was ground to a fine powder with a ball mill (MM200, Retsch, Germany). Finally, about 4 mg was weighed into small tin capsules, and carbon (C) and N concentrations (in %) were determined with a Flash EA 1112 elemental analyzer (Thermo Italy, former CE Instruments, Rhodano, Italy).

Soil mineral N was measured during the wheat cropping season to quantify the instantaneous extractable inorganic N pools. Soil samples were collected five times using a machine-assisted soil core sampler within the same field sections used for AGB sampling, with ten to fifteen soil cores taken at each section and bulked for each soil depth (0–0.3, 0.3–0.6, and 0.6–0.9 m). Ammonium (NH4+) and nitrate (NO3-) were extracted by an external laboratory using a 0.01 M CaCl₂ solution at a 1:4 soil-to-solution ratio, following the Swiss reference method (Buerge, 2020) established by the Federal Agriculture Research Center, Agroscope. Concentrations were photometrically determined by the same laboratory and reported in mg N per 100 g of dry soil. Soil mineral N (N_min) was then calculated in kg N ha⁻¹ using the same reference method.

Plant Root Simulator probes (PRS™, Western Ag, Saskatoon, Canada) were used as a complementary measure of plant-available N supply over time. In contrast to N_min measurements, which provide a snapshot of extractable inorganic N at the time of sampling, PRS probes integrate ion supply to exchange membranes over the burial period and therefore provide a time-integrated estimate of available N supply. Measurements of plant-available soil N complemented soil sampling from March to September 2023. Three PRS sets, each consisting of one anion and one cation exchange resin probe (17.5 cm² surface area each), were buried at 0–0.1 m soil depth. This depth was chosen to target the soil layer most strongly affected by management and where root density and nutrient uptake are highest. The probes were positioned along the wheat inter-row at the same five locations used for chamber measurements (Fig. 1). The probes were replaced every two to three weeks, resulting in seven burial periods. No probes were buried in June due to technical issues, nor in the second half of July due to tillage and sowing operations. Prior to shipment to Western Ag for analyses, the PRS probes were rinsed with deionized water and stored at 5 °C. Extraction of NH4+ and NO3- with 0.5 mol L⁻¹ HCl was conducted by pooling all three PRS sets from each location, yielding one value per location and sampling period (in µg N per 10 cm² per time of burial). Since all NH4+ measurements were below the detection limit (except for one sample), we excluded NH4+-N from further analyses.

Winter wheat was sown on 6 October 2022, following the final cut of a temporary grass-clover ley on 20 September 2022 (Table 1). The period after the final cut was wet, with 103 mm of precipitation in 15 d, which increased WFPS (Fig. 2a, b). During the winter wheat season, cumulative precipitation amounted to 698 mm, with rainfall concentrated in November–December 2022 (33 %) and March–April 2023 (29 %). A dry period occurred from late December to early March, although topsoil WFPS remained above 60 % (Fig. 2b), while a second dry period from mid-May to late June reduced WFPS to 20 %–30 % in the upper 0.3 m of the soil. Topsoil temperature ranged from 0.4 to 24.6 °C, with minimum values in winter and maximum values in July. Winter wheat development was slow during winter and early spring, with LAI remaining low until March (Fig. 2d). LAI increased rapidly from late March and peaked at the end of May, shortly before anthesis (Table A4). The first two fertilizer applications were conducted on 27 and 28 February during late tillering, and the final application was on 6 April during early stem elongation. During the short summer cover crop period following wheat harvest and soil tillage, soil temperature remained variable before decreasing toward the end of the period, while soil moisture increased, particularly in deeper soil layers (Fig. 2b, c), and cover crop development remained limited (Fig. 2d).

During the winter wheat cropping season (282 d), cumulative N₂O emissions were 5.5 kg N₂O-N ha⁻¹ (Table 2), with an average gap-filled flux of 0.81 nmol m⁻² s⁻¹. When scaled to crop yield, this corresponded to 1.1 kg of N₂O-N emitted per ton of wheat grain (dry matter). Based on the total N input from fertilizers of 138 kg N ha⁻¹ (Table 1), the resulting apparent emission factor (EF) was 4.0 %. Over the one-year period starting at winter wheat sowing, cumulative N₂O emissions amounted to 7.0 kg N₂O-N ha⁻¹, corresponding to an annual apparent EF of 5.1 %. Because no unfertilized control treatment was available, both values also reflect N₂O emissions from sources other than fertilizer application, such as those associated with mineralization of residues from the preceding grass-clover ley. During the subsequent 60 d cover crop season, cumulative N₂O emissions reached 1.1 kg N₂O-N ha⁻¹, and the average daily sum of N₂O emissions (17.6 g N₂O-N ha⁻¹ d⁻¹) was only slightly lower than that observed during the winter wheat season (19.6 g N₂O-N ha⁻¹ d⁻¹).

Throughout the winter wheat cropping season, the ecosystem acted as a net CO₂sink, with a cumulative NEE of -89 g CO₂-C m⁻² (Table 2). When expressed in CO₂-equivalents, concurrent N₂O emissions offset approximately 70 % of this seasonal net CO₂ uptake, while CH₄ emissions remained negligible. The resulting net GHG exchange (NGHGE) was therefore slightly negative, representing a small GHG sink with -76 g CO₂-eq m⁻². In contrast, the short summer cover crop period represented a net CO₂ source, emitting 108 g CO₂-C m⁻². N₂O emissions during this period accounted for approximately 10 % of NGHGE, while the net CH₄ exchange was again negligible. The NGHGE during the cover crop period showed large net GHG emissions (+441 g CO₂-eq m⁻²). The difference between the two crops was even more pronounced when comparing their daily NGHGE, with winter wheat averaging -0.27 g CO₂-eq m⁻² d⁻¹ and the cover crop averaging +7.4 g CO₂-eq m⁻² d⁻¹, resulting in a daily difference of approximately 8 g CO₂-eq m⁻² d⁻¹ (Table 2). However, when lateral C fluxes were included to calculate the net GHG budget, winter wheat was a substantial net GHG source (+1349 g CO₂-eq m⁻²), due to its large C export with harvested grain and straw, while the cover crop was a smaller net GHG source (+441 g CO₂-eq m⁻²; note: numbers are identical to NGHGE because no lateral C fluxes occurred). Sensitivity analysis across alternative u* thresholds (CUT16, CUT84) showed a NGHGE spread of 82 g CO₂-eq m⁻² for winter wheat versus 26 g CO₂-eq m⁻² for the cover crop, driven mainly by CO₂ fluxes (Table A3).

We used stable oxygen and nitrogen isotopes to identify the microbial processes underlying the measured N₂O fluxes during the three weeks following the first fertilization events on 27 and 28 February 2023. The δ18O and site preference (SP-N₂O) of emitted N₂O ranged between 17.3 ‰ and 39.8 ‰ and from -7.2 ‰ to +18.2 ‰, respectively (Fig. 7a). The dual isotope plots suggested that heterotrophic bacterial denitrification and/or nitrifier denitrification were the dominant N₂O production pathways (Figs. 7a, A5). Moreover, the observed trend of increasing enrichment of δ18O and SP-N₂O over time (Fig. 7a) suggested isotopic fractionation associated with the microbial reduction of N₂O to N₂ during denitrification. Results from the FRAME model supported the qualitative source attribution derived from the isotope mapping approach and enabled the quantification of both the relative contributions of individual N₂O-producing processes and the extent of N₂O reduction to N₂. However, following Buchen et al. (2017), we pooled heterotrophic bacterial denitrification with nitrifier denitrification, and nitrification with fungal denitrification, due to the large uncertainty in separating these processes given their similar endmembers. All four processes are nonetheless shown separately in Fig. A6. N₂O fluxes decreased during the sampling period (Fig. 7b), whereas the source contributions showed no clear temporal trend (Fig. 7c). Heterotrophic bacterial denitrification and/or nitrifier denitrification dominated N₂O production throughout the sampling period, contributing 74 %–86 % (±9 %–21 %) across the four sampling dates (Fig. 7c; based on the mixing-reduction scenario), while nitrification and/or fungal denitrification accounted for the remaining 14 %–26 %. The alternative reduction-mixing scenario produced very similar source-fraction distributions, with only slightly higher reduced N₂O fractions, while the overall interpretation remained unchanged (Fig. A6). The reduced N₂O fraction, reported as (1-r), remained between 40 % and 47 % on the first three sampling dates and increased markedly to 78 ± 13 % on 17 March, indicating a larger fraction of N₂O reduced to N₂ at the end of the sampling period, coinciding with decreasing chamber-based N₂O fluxes (Fig. 7b).

Two central objectives of this study were to quantify N₂O fluxes and determine when and how environmental and management drivers affected these fluxes during a winter wheat cropping season and the subsequent cover crop season. To address these objectives, we used an explainable ML framework (i.e., XGBoost-SHAP) as an interpretative tool to extract, from high-frequency EC observations, the temporal development of driver contributions to N₂O fluxes. We preferred tree ensembles over recurrent architectures such as long short-term memory networks, which can capture lagged driver-response relationships directly but require more data than a single site-year provides (Hamrani et al., 2020). Temporal dependencies were therefore represented through explicit feature engineering in our study. Such a framework offers several advantages over traditional statistical models. First, N₂O fluxes are highly episodic and respond non-linearly to interacting drivers, conditions under which ML models are expected to outperform linear models (Gnisia et al., 2025; Hamrani et al., 2020). Second, SHAP values allowed us to quantify not only the average importance of each driver, but also how its contribution evolved over time. Third, XGBoost captures interactions among drivers intrinsically through the hierarchical structure of the tree splits, without interaction terms needing to be specified a priori. Fourth, this framework enabled us to incorporate management history alongside continuous biophysical drivers, enabling interpretation of management effects not possible with classical statistical analyses.

As with any data-driven approach, these advantages come with trade-offs. SHAP quantifies how each driver contributes to the model's predictions rather than establishing process-based causal effects, and it can attribute variance only to predictors that are included. Two key drivers were not explicitly represented in our ML model: soil N mineralization and the availability of labile organic C. Their influence was therefore likely redistributed onto other predictors. Thus, effects of the mineralization of residues from the preceding grass-clover ley likely extended beyond the 30 d windows used here to represent soil disturbance, while the effect of labile organic C supplied with the slurry could not be separated from that of the fertilizer N and was therefore attributed to the engineered fertilization variables. Nevertheless, high-resolution environmental, management, and EC data over a full year provided a sound basis for the XGBoost-SHAP framework in this study, within clear limits. The model reproduced the timing of the main emission episodes but underestimated peak magnitudes and overestimated low fluxes, and it explained a moderate share of the variance in the independent hold-out period (R2=0.36; Fig. A2). Some temporal patterns of the modelled contributions might also reflect how predictors were constructed rather than changes in the underlying processes. For example, the fertilization and soil disturbance variables were engineered based on prior knowledge to represent a 30 d effect, so some driver contributions shifted abruptly once that window elapsed (Fig. 4b, e.g., early April and early May). SHAP attributions were nonetheless generally consistent with established process understanding and published field studies (see Sect. 4.2).

These trade-offs also delimit what an empirical ML framework can and cannot replace. Process-based biogeochemical models explicitly represent soil N and C pools and related transformations, allowing them to simulate hypothetical management scenarios and extrapolate beyond observed conditions, though their ability to reproduce N₂O fluxes and their short-term variation remains limited (Saha et al., 2021; Scheer et al., 2025). ML models cannot represent the underlying pools and instead describe the conditions under which N₂O fluxes occurred, but they require no site-specific parameterization and can exploit the full temporal resolution of EC data. Both approaches are therefore complementary rather than competing. Comparing SHAP-identified driver responses against the sensitivity of a process-based model to the same drivers can provide an independent check on whether the empirical attributions are mechanistically plausible. Furthermore, a limitation of our own setup points to a way forward: soil mineral N could not be used as a predictor because it was measured only on discrete dates (Sect. 2.7.1), whereas a calibrated process-based model could supply continuous simulated N_min and mineralization rates as ML predictors, replacing the engineered fertilization and soil disturbance proxies used here. Integrating ML with process-based models (Saha et al., 2021; Sharma et al., 2026) and adopting knowledge-guided ML approaches (Liu et al., 2022) are therefore promising next steps to address the complexity of N₂O emissions from managed ecosystems.

Two periods emerged as particularly relevant for N₂O emissions: crop establishment and early spring. Both were characterized by a mismatch between N supply and plant demand, with substantial mineral N supply from mineralization or fertilization coinciding with limited crop uptake, resulting in surplus soil mineral N, that under favorable moisture and temperature conditions, fuelled microbial N₂O production.

During crop establishment, elevated emissions were affected by soil disturbance and low crop activity (i.e., low GPP). Herbicide termination of the temporary grassland and direct seeding of winter wheat likely stimulated mineralization of soil organic matter and crop residues, as indicated by increased NEE (intensified soil respiration), while low GPP reflected limited plant N uptake (Fig. A1). Together, these processes raised soil N_min from 8.7 to 34.6 kg N ha⁻¹ between October and early December 2022 (0–0.3 m soil depth; Fig. 6d), consistent with observations after grassland termination (Beiküfner et al., 2025; Buchen et al., 2017; Helfrich et al., 2020). A comparable response occurred during cover-crop establishment after wheat harvest, when warm soil temperatures and soil tillage enhanced mineralization (Grandy and Robertson, 2006) and N_min rose from 12 to 27 kg N ha⁻¹ between late May and early August (Fig. 6d). Thus, management practices that leave soils disturbed and poorly vegetated favored N₂O emissions at our site, consistent with previous work suggesting that rapid and sustained soil cover, e.g. through optimized sowing dates, cover crops or cultivar choice, can reduce emissions during such periods (Maier et al., 2025; Quemada et al., 2020; Tallec et al., 2022).

In early spring, high N₂O emissions were predominantly driven by N fertilization, with WFPS contributing during peak events. The first fertilizer applications in late February combined mineral (NH₄NO₃) and organic (slurry) N, supplying a large N input (111 kg N ha⁻¹) at tillering, which coincided with high soil moisture, conditions typical for central European climates (Ruosteenoja et al., 2018). At this stage, low LAI (Fig. 2d) and low GPP (Fig. 4a) indicated limited plant N demand, so that soil N_min reached 74 kg N ha⁻¹ (Fig. 6d), which, together with precipitation events (Fig. 2a), created optimal conditions for microbial N₂O production. By contrast, the smaller final fertilization in early April (27 kg N ha⁻¹) was applied during stem elongation, when plant N demand was higher, as indicated by increased LAI (∼ 2 m² m⁻²) and GPP, which likely reduced substrate availability for N₂O production. Mitigation practices in such periods will depend on better synchronizing fertilizer N inputs with crop demand and soil N mineralization, e.g. through adjusted fertilizer timing and amounts (Hoben et al., 2011; Omonode et al., 2017), split applications, site-specific planning (Jordan-Meille et al., 2023), precision fertilization (Diacono et al., 2013) or enhanced-efficiency fertilizers (Akiyama et al., 2010).

Over the entire study period, N fertilization was the dominant driver increasing N₂O fluxes (Fig. A3), consistent with its role as the primary source of reactive N in croplands (Shcherbak et al., 2014). WFPS, GPP, soil temperature, and soil disturbance made substantial but temporally varying contributions, underlining the multifactorial and interactive nature of N₂O emissions. WFPS had both positive and negative effects but consistently emerged as a key driver during peak emission events. GPP, used here as a proxy for crop activity and N demand, was clearly negatively associated with N₂O fluxes (Fig. 5). We interpreted this result as an indirect, plant-mediated effect, reflecting microbial substrate limitation: as the crop grows, its N uptake depletes soil mineral N available for microbial N₂O production (Abalos et al., 2018; Feigenwinter et al., 2023; Maier et al., 2022; Timilsina et al., 2024). Because WFPS was included in the model, an alternative pathway, i.e., high transpiration rates lowering soil moisture and thereby suppressing N₂O production, was already accounted for. Although root C exudation might favor denitrification (Philippot et al., 2009), the negative relationship suggests substrate limitation dominated. Overall, our results are consistent with the “N synchrony framework” (Cassman et al., 2002; Cassman and Dobermann, 2022): the highest N₂O losses coincided with the largest mismatch between N supply and crop N demand, in line with previous work linking N surplus to N₂O emissions (Van Groenigen et al., 2010; Tallec et al., 2022).

Cumulative N₂O emissions during the winter wheat season were 5.5 kg N₂O-N ha⁻¹, substantially exceeding the ranges reported for winter wheat by Garnier et al. (2024) for northern France (0.1–2.7 kg N₂O-N ha⁻¹) and by Tallec et al. (2022) for southwestern France (0.95–2.91 kg N₂O-N ha⁻¹), both chamber-based studies, although our N input (138 kg N ha⁻¹) was well within their reported fertilizer ranges. Yield-scaled N₂O emissions in our study (1.1 kg N₂O-N Mg⁻¹grain) were approximately three times higher than global values reported for wheat by Yao et al. (2024; 0.301–0.346 kg N₂O-N Mg⁻¹grain). Part of this discrepancy may be methodological, since global syntheses rely largely on soil chambers with lower spatial and temporal coverage than continuous EC, which better captures short-lived peaks and integrates fluxes over larger areas (Butterbach-Bahl et al., 2026). In addition to methodological differences, the elevated N₂O emissions in our study likely reflected site- and management-specific conditions. At our site, substantial amounts of mineral N were probably supplied by mineralization of residues from the preceding three-year grass-clover ley, as indicated by the accumulation of soil mineral N during autumn in the absence of any fertilization (Sect. 4.2). In addition, the near-simultaneous application of mineral N and slurry at the end of February likely created conditions particularly conducive to N₂O production, owing to the input of readily degradable carbon (Dittert et al., 2005; Schils et al., 2008). Poor drainage in the silty clay Eutric-Stagnic soil likely amplified both effects, favoring the high soil moisture and limited aeration under which denitrification proceeds.

To our knowledge, this is the first study to report EC-based N₂O fluxes over a complete winter wheat season. Notably, our N₂O emissions only slightly exceeded those observed for maize in Switzerland (with EC: 4.8 kg N₂O-N ha⁻¹; Maier et al., 2022), despite the much shorter maize season (127 vs. 282 d). The two crops also clearly differed in their N use efficiency (NUE, calculated as the ratio between N output to N input, not accounting for soil N mineralisation): winter wheat took up 113 kg N ha⁻¹ from an input of 138 kg N ha⁻¹ (NUE of 82 %), whereas maize took up only 56 kg N ha⁻¹ from an input of 110 kg N ha⁻¹ (NUE of 51 %; derived from Table 1 in Maier et al., 2022). The higher wheat NUE is consistent with Yu et al. (2022) and may partly reflect crop-specific differences (e.g., C3 vs. C4 traits, soil N dynamics, weather) but should be interpreted with caution since N supplied through mineralization of the preceding 3-year grass-clover ley was not captured in the fertilizer-N input term. So, the apparent NUE given above likely overestimates the true NUE.

The N₂O losses substantially affected the EC-based net GHG exchange (NGHGE) of the winter wheat season, with N₂O emissions offsetting 70 % of the seasonal net CO₂ uptake. This offset is considerably higher than those reported by Maier et al. (2022) for pea and maize in Switzerland (12 % and 10 %, respectively) and for managed Central European permanent grasslands (Hörtnagl et al., 2018). This large offset partly reflected the small net CO₂ sink during the winter wheat season, likely driven by enhanced heterotrophic respiration from decomposition of the preceding grass-clover ley residues. When including lateral C fluxes, i.e., organic fertilizer inputs and harvest exports, to calculate the net GHG budget, winter wheat was a substantial net GHG source, consistent with net C losses and declining SOC stocks reported from long-term measurements at the same site (Emmel et al., 2018). The field acted as a small net CH₄ source, consistent with results from other agricultural soils (Cowan et al., 2021). CH₄ emissions occurred mainly during wet periods, while dry periods showed low uptake, indicative of aerobic oxidation (Le Mer and Roger, 2001). The net CH₄ source likely reflected the poorly drained silty clay soil at our site, which likely promoted anaerobic conditions favoring methanogenesis.

Despite the absence of fertilizer application, the average N₂O flux during the cover crop season was only slightly lower than that during the winter wheat season (0.72 vs. 0.81 nmol m⁻² s⁻¹, respectively). This most likely reflected low plant N uptake due to slow crop establishment and enhanced N mineralization due to warm temperatures and post-harvest soil disturbance, as indicated by increased soil N_min between late May to early August (Fig. 6d). Such conditions are consistent with recent studies showing that bare or sparsely vegetated soils promote N₂O emissions through reduced plant N uptake and enhanced microbial activity (Maier et al., 2025; Shang et al., 2024; Tallec et al., 2022; Timilsina et al., 2024). Moreover, daily mean NEE remained positive until late August (Fig. 3a), indicating that the soil was a net CO₂ source for approximately half of the 60 d cover crop season. As a result, the cover crop season was characterized by net CO₂ release, accounting for 90 % of the total GHG budget during this period, with N₂O contributing the remaining 10 %, while CH₄ fluxes were negligible (Table 2). However, this net CO₂ release reflected NEE during a short, disturbed post-harvest establishment phase rather than an intrinsic property of cover crops: a summer cover crop between two winter cereals is not standard practice at this site and was implemented for the first time. Indeed, cover cropping reduced post-harvest GHG emissions relative to bare soil: in 2019, when the field was left bare after winter wheat harvest, daily post-harvest CO₂-eq fluxes were roughly double those measured during this study (16 vs. 7.4 g CO₂-eq m⁻² d⁻¹; Maier et al., 2025), driven primarily by increased CO₂ emissions. Although management and climatic conditions differed between the two years, this pronounced difference highlights the benefit of maintaining a green cover crop between two winter crops, also consistent with modelling results at the same site (Emmel et al., 2018) and current Swiss policy promoting continuous soil cover (Art. 71c DZV, Swiss Federal Council, 2025).

Stable isotope sampling on selected dates in March 2023, following fertilizations on 27–28 February, coincided with the period of highest N₂O fluxes: cumulative EC-based emissions in March alone totalled 2.5 kg N₂O-N ha⁻¹, i.e. 45 % of the seasonal total for winter wheat. The FRAME model by Lewicki et al. (2022) allowed quantification of the relative contributions of different N₂O-producing processes over time. Denitrification, in particular heterotrophic bacterial denitrification, was the dominant source of N₂O (Figs. 7, A6), consistent with poor soil drainage and persistently high WFPS (> 65 % at 0.05 m), which promoted anaerobic conditions that favor denitrification (Bateman and Baggs, 2005; Davidson, 1991). Improving drainage in such settings may therefore help to mitigate N₂O emissions by reducing the extent and duration of waterlogging and the associated anaerobic conditions (Bouwman, 1996; Grossel et al., 2016). The low reduced N₂O fraction (∼ 40 %–50 %) in the week following fertilization may reflect elevated NO3- concentrations, leading to NO2- accumulation and inhibition of the final reduction step of N₂O to N₂ (Blackmer and Bremner, 1978; Senbayram et al., 2019). The fraction then increased markedly to ∼ 80 % on 17 March, coinciding with topsoil WFPS exceeding 80 %, thereby favoring complete denitrification. The near-simultaneous application of mineral N and slurry likely further promoted denitrification by providing both NO3- and easily degradable organic C, creating conditions favorable for heterotrophic denitrifier activity (Dittert et al., 2005). Because such high-risk combinations can occur in mixed crop-livestock systems, policy and management guidance should discourage closely timed applications (Schils et al., 2008).