Drained agricultural peatlands are characterized by some of the highest greenhouse gas emissions from terrestrial ecosystems due to soils that are often rich in organic carbon (C) and nitrogen (N) content (Leifeld and Menichetti 2018; Hemes et al. 2019; Anthony and Silver 2020; Anthony et al. 2023). The oxidation of soil organic C to carbon dioxide (CO2) during drainage drives the majority of the global warming potential (GWP) from agricultural peatland soils (Freeman et al. 2022). However, nitrous oxide (N2O) and methane (CH4) can represent up to 35% of CO2-equivalent (CO2e) agricultural peatland emissions (Anthony and Silver 2021). Combined agricultural peatland CO2, N2O and CH4 emissions are estimated to contribute to 32% of global cropland emissions while accounting for only 3% of global land area and providing only 1% of crop kilocalories worldwide (Carlson et al. 2017; Freeman et al. 2022).

Soil N2O and CH4 emissions are often characterized by hot moments, where the majority of annual emissions occur on a timescale of days to weeks (Molodovskaya et al. 2012; Savage et al. 2014; Krichels and Yang 2019; Anthony and Silver 2021; Anthony et al. 2023). Hot moments of N2O and CH4 high emission events are often poorly constrained, but can be triggered by weather events (rainfall, freeze/thaw) or management induced changes (fertilization, irrigation, flooding, drainage) to the soil environment (Conrad 1989; McNicol and Silver 2014; Wagner-Riddle et al. 2017; Krichels et al. 2019a). The drivers of spatial variability in N2O and CH4 fluxes are even less well understood. Low lying areas, legacies of prior land use, and areas of particularly high resource availability may be important locations for high emission potential, termed hot spots, on the landscape. There are several ways in which hot spots of high fluxes may manifest in agricultural ecosystems (sensu McClain et al. 2003, Wagner-Riddle et al. 2020). Hot spots could be characterized by consistently high greenhouse gas emissions from a single locale or set of locales such as a drainage ditch or furrow which have near constant low or dynamic redox and high substrate availability. Alternatively, locations with periodic high flux events (i.e., hot moments) may be limited to a single or few locations relative to a larger ecosystem matrix, thus linking hot spots with hot moments. Locations with a higher probability for hot moments of high flux may be characterized by microtopography (e.g., low lying areas), properties of vegetation (e.g., high litterfall inputs), or land use legacies (e.g., N fixing crops) (Krichels and Yang 2019; Hall et al. 2023). Finally, hot spots could occur inconsistently as locales with a higher proportion of hot moments than the surrounding matrix for a defined period (month, season, year) that vary over space and time. This final category of greenhouse gas emission hot spot would be particularly difficult to characterize but could be linked to dynamic soil heterogeneity (Totsche et al. 2010), land management activities such as tillage, planting, or fertilizer application that may reduce spatial variability in soil conditions (Adewale et al. 2016), or result from random, or chaotic processes associated with ecosystem complexity (Baveye et al. 2018; Sokol et al. 2022).

The datasets needed to constrain hot spots and hot moments of ecosystem processes are inherently large and complex. Denitrification and methanogenesis are often regulated by non-linear and interactive effects which are difficult to model using traditional approaches. Machine learning models can help capture trends and identify important drivers and predictors of ecosystem processes; relationships are determined from available data, allowing for unbiased projections while providing insights into functional relationships (Saha et al. 2021). Machine learning models can also be used to make projections at larger spatial and temporal scales (Alber et al. 2019). Supervised machine learning models, such as random forest modelling, utilize an algorithm for classification and regression based on the principle of recursive partitioning (Breiman 2001), and independent of the assumption of functional relationships between the response and predictor variables (Saha et al. 2021).

An important limitation of machine learning models is that they require large, high-quality training datasets to appropriately train the model without using a significant proportion of the observed data (Qiu et al. 2016). Accuracy likely improves with increased spatial and temporal data inputs and is therefore inherently limited by the scope of a training dataset. This makes it difficult to predict the occurrence and magnitude of hot spots and hot moments without adequate representation of large flux events within training datasets. Unexplained variability in previously developed models is likely a function of consistently underfitting hot moments or under sampling key driving variables (soil O2, moisture, and temperature) across space or time. Continuous long-term (multi-month or year) measurements of N2O and CH4 fluxes and their drivers across scales are rare, limiting our ability to scale hot spots and hot moments across time and space.

Continuous automated measurements can improve the temporal resolution of greenhouse gas emissions data (Holst et al. 2008; Barton et al. 2015). However, these approaches are still typically limited to 6–12 sampling locations and study duration is often less than one full year. Similar challenges exist for the determination of controlling variables occurring within soil environments (spatial scales from aggregate to pedon and temporal scales from seconds to days). Patterns in substrate availability, gas diffusivity, and microbial activity throughout a soil profile are likely to be important drivers of observed spatiotemporal variability in agricultural peatland greenhouse gas emissions. Soil moisture content, redox conditions, and C and nutrient availability often vary with depth in soils, leading to differential patterns in C and N cycling (Jobbagy and Jackson 2000; Krichels et al. 2019b; Thorup-Kristensen et al. 2020). Incorporating physical and biogeochemical conditions both vertically and horizontally in soils can significantly improve predictions of greenhouse gas fluxes (Potter 1997; Chatskikh et al. 2005; Xing et al. 2011; Feng et al. 2021). Thus, challenges persist for the prediction of greenhouse gas emissions at an ecosystem scale and corresponding calibration data is likely to be critical to better predict spatiotemporal variability.

In agricultural peatlands, soil nitrate (NO3) availability, water filled pore space (WFPS), and its effects on O2 availability can be significant drivers of hot moment N2O flux events (Pärn et al. 2018; Anthony and Silver 2021; Anthony et al. 2023). Soil NO3 can accumulate under oxic, well-drained soil conditions via N mineralization and nitrification following fertilization or crop harvest (Kirk et al. 2015; Anthony and Silver 2021) and may be supplemented by Fe-coupled anaerobic ammonium oxidation in Fe- and C-rich peatlands soils (Martikainen et al. 1993; Golovchenko et al. 2007; Yang et al. 2012; Yang and Liptzin 2015). Due to a relatively unlimited organic C and N supply (in comparison to most other soils), many temperate agricultural peatland soils experience high rates of N mineralization and nitrification (Danevčič et al. 2010; Jerman et al. 2017; Oktarita et al. 2017) suggesting a sustained and active N cycling microbial community that likely facilitates the production and emission of N2O.

Net CH4 fluxes from agricultural peatlands are assumed to be minimal under drained conditions (Maljanen et al. 2010; Günther et al. 2019) as this is thought to favor aerobic respiration pathways that outcompete methanogens for available substrate and/or stimulate methanotrophic activity (Conrad 2007). But irrigation or intense rainfall events can create periods of anaerobic conditions ideal for CH4 production (Hemes et al. 2019; Anthony and Silver 2021). At the ecosystem scale, the influence of soil temperature, water table fluctuations, and plant activity drive CH4 fluxes in both drained and restored wetlands, often as scale emergent controls that can only be determined with high density eddy covariance measurements (Sturtevant et al. 2016; Oikawa et al. 2017; Chamberlain et al. 2018). However, spatial prediction of CH4 fluxes often suffers from the same pitfalls as N2O flux predictions, with methods lacking agreement between observed and predicted fluxes and poor performance when extrapolated to other locations (Ehrhardt et al. 2018; Gaillard et al. 2018).

Cover type and management are likely to be important determinants of the patterns and magnitude of greenhouse gas emissions from drained agricultural peatlands. Agriculture practices in drained peatlands are generally determined by soil organic matter content and nutrient availability, and can range from grazed pastures to annual or perennial agricultural systems (Hatala et al. 2012; Knox et al. 2015), all of which may vary in rates and pathways of greenhouse gas fluxes. In some drained temperate peatlands including California’s Sacramento-San Joaquin Delta, soils with elevated organic matter content, and thus high rates of organic matter mineralization, are frequently chosen for corn (Zea mays) agriculture given their inherent productivity. As annual cropping systems, these ecosystems are often prioritized for winter flooding of fallow fields for weed management and migrating waterfowl habitat that also coincide with hot moments of CH4 and N2O flux (Pellerin et al. 2014; Anthony and Silver 2021). In contrast, more degraded peat soils with higher mineral content are often prioritized for alfalfa (Medicago sativa) agriculture as this perennial is an N-fixing crop and can be productive over many years with a single planting and only periodic summer irrigation (Anthony et al. 2023). The most degraded soils are often managed as ditch-irrigated grazed pasture as they are not economically suitable for crop agriculture (Teh et al. 2011; Knox et al. 2015). Drained and irrigated peatlands generally experience predictable periods of inundation followed by well-aerated conditions; this combination fosters the biogeochemical environment needed to drive high greenhouse gas fluxes (Hemes et al. 2019; Anthony and Silver 2021; Anthony et al. 2023). Dynamic redox conditions over time can support high substrate availability for both denitrifers (e.g., higher redox) and methanogens (e.g., lower redox) (Conrad 1989). Methane and N2O fluxes can also be supported by microtopographic variability that leads to temporary ponding in low-lying regions with gradients of reducing conditions and well-aerated environments.

To better constrain greenhouse gas budgets and their drivers from agricultural peatlands, we conducted continuous in-situ soil greenhouse gas flux measurements from three different agricultural peatland land uses in a small geographical area (~ 60 km2): an annual corn cropping system, a perennial alfalfa cropping system, and a grazed irrigated pasture. These systems provided a unique template to explore the variability in greenhouse gas fluxes across land uses while controlling for climate and general soil conditions (e.g., drained peatland). We tested the hypothesis that drained peatland agricultural creates persistent hot spots of CH4 and N2O emissions due to low or fluctuating redox conditions, respectively, and that the majority of hot moments would be limited to a relatively few hot spot chambers. We also hypothesized that this method of quantifying hot spots and hot moments would identify relatively less CO2 hot spots and hot moments as soil respiration would be more uniformly distributed across these agricultural peatland ecosystems with variability driven by seasonality in temperature and precipitation. We predicted that the annual corn system would have fewer emission hot spots due more uniformity in inputs and surface soil disturbance from intensive management, while the perennial alfalfa system and pasture would have relatively greater hot spot variability due to the development of greater heterogeneity over time in the absence of intensive management. We used continuous cavity ringdown spectroscopy and automated chambers to measure CO2, CH4, and N2O fluxes. Soil sensing data of O2, moisture, and temperature across three soil depths (10, 30, and 50 cm) from the corn and alfalfa sites allowed us to explore potential drivers. Our goal was to quantify the spatiotemporal distribution of chamber-scale hot spots and hot moments in agricultural peatlands, the relationships between hot spots and hot moments of greenhouse gas flux, and explore machine learning methods to predict these hot spots and hot moments across different agricultural peatland land uses.


Site description

We used data from three agricultural peatland sites in the Sacramento-San Joaquin Delta region of California. The sites experienced a similar climate over a relatively small regional scale (~ 60 km2) but had contrasting land uses and soil conditions (Table 1). The datasets included over 98,000 chamber flux measurements from an organic-rich annual maize agroecosystem (38.1091° N, 121.5351° W; hereafter referred to as corn). Preliminary data on temporal flux patterns from years 1 to 3 were reported in Anthony and Silver (2021); here we added an additional year of observations and explored spatial dynamics not previously considered. We also included data from a continuous perennial alfalfa agroecosystem (chambers: 38. 1076° N, 121.5021° W, tower: 38.0992° N, 121.4993° W; hereafter referred to as alfalfa) with 103,000 flux measurements; as with the corn site previous work explored the temporal dynamics of greenhouse gas emissions from this site, but did not report spatial patterning (Anthony et al. 2023). Finally, we included data from a continuously grazed irrigated pasture 38.0402° N, − 121.7272° W; hereafter referred to as pasture) with 33,000 flux measurements. These are the three dominant land uses for drained agricultural peatlands in the region, accounting for 53% of crop acreage (The Delta Protection Commission 2020). Importantly, grazed pastures often occur on degraded peatland soils where other agricultural activity is less likely to be profitable (Hatala et al. 2012; Buschmann et al. 2020). Corn was the only system to receive N fertilizer inputs, which were applied at a rate of 118 kg N ha−1 year−1 (Anthony and Silver 2021). Here we used the three datasets, where we include both fluxes to the atmosphere from soil and fluxes from the atmosphere to the soil, to explore spatial dynamics both within and across agricultural peatland ecosystems and relationships to temporal patterns which previous studies did not address.

Table 1 Descriptions of soil C (% ± SE), soil N (% ± SE), and soil pH (± SE) in the top 0–15 cm and relevant management information

Greenhouse gas flux measurements

All datasets collected soil fluxes of CO2, CH4, and N2O continuously using an automated chamber system for 1 to 4 years. Each system consisted of nine opaque, automated gas flux chambers (eosAC, Eosense, Nova Scotia, Canada) connected to a multiplexer (eosMX, Eosense, Nova Scotia, Canada). The multiplexer allowed for dynamically signaled chamber deployment and routed gases to a cavity ring-down spectrometer (Picarro G2508, Santa Clara, CA, USA). Chambers were measured sequentially over a 10-min sampling period with a 1.5-min flushing period before and after each measurement. Chambers were deployed in a 10 × 10 m grid design, with each chamber approximately 5 m from other chambers. Vegetation was included in all pasture chambers and randomly assigned to bare soil (n = 4) or plants (n = 5) in alfalfa. In corn, chambers were randomly assigned to beds (n = 3) or furrows (n = 6) without vegetation included due to the height of the stalks. Data filtering and detection limits are described in detail in Anthony & Silver (2021) and Anthony et al. (2023).

Site-level net ecosystem CO2 exchange (NEE) and calculated site-level global warming potential (GWP) utilizing annual flux measurements from paired Ameriflux towers in corn (US-Bi2) and pasture (US-Snf) sites (Kasak et al. 2020; Rey-Sanchez et al. 2021) was employed to contextualize the importance of hot moments on net ecosystem CO2e budgets. For the alfalfa site, we utilized annual net ecosystem exchange (NEE) estimates from a nearby (< 1 km) Ameriflux tower (US-Bi1) in an alfalfa agroecosystem with identical management practices and soil type (Anthony and Silver 2020; Rey-Sanchez et al. 2022b). The eddy covariance technique captures continuous, long-term exchange of CO2, CH4, water, and energy fluxes between the landscape and the atmosphere, along with measurements of environmental drivers (Eichelmann et al. 2018). Fluxes were measured at a frequency of 20 Hz using open-path infrared gas analyzers (LI-7500 for CO2, LI-7700 for CH4, LiCOR Inc., Lincoln, NE, USA). Sonic anemometers measured sonic temperature and three-dimensional wind speed at 20 Hz (WindMaster Pro 1590, Gill Instruments Ltd, Lymington, Hampshire, England). To convert N2O and CH4 flux measurements to CO2e, we used the IPCC AR5 100-year GWP values of 28 CO2e for CH4 and 298 CO2e for N2O (Myhre et al. 2013).

Environmental sensors

In both the corn and alfalfa systems, two sets of soil sensors per site were installed at depths of 10 cm, 30 cm, and 50 cm and included O2 and soil temperature sensors (SO-110, Apogee Instruments, Logan, UT) and moisture sensors (CS616, Campbell Scientific, Logan, UT) connected to a datalogger (CR1000, Campbell Scientific, Logan, UT) from September 2018-February 2021 in alfalfa and October 2018-July 2021 in corn. Sensors were installed as close to corresponding chamber grid locations as possible while minimizing soil disturbance and limiting interference with agricultural activities. Except for periodic agricultural events, sensors remained installed throughout the year. Data filtering and data gaps are described in detail in Anthony & Silver (2021) and Anthony et al. (2023).

Statistical analyses and machine learning

Following data filtering, annual fluxes and the quantity and magnitude of hot moments and hot spots were determined. Annual fluxes were determined by taking the mean annual flux of each chamber as well as the mean annual flux for each field location. We defined hot moments as flux measurements with values greater than four standard deviations from the mean (Anthony and Silver 2021; Anthony et al. 2023), as statistically 99.9% of the population should fall within four standard deviations of the mean given a normally-distributed dataset. All individual flux datasets (CO2, CH4, and N2O) from each field site were approximately log-normally distributed. Calculating standard deviations on non-Gaussian datasets generally overestimates the true standard deviation (Carter 2013), but is still appropriate in describing the spread of the variable of interest. Therefore, quantifying hot moments in this manner is generally conservative, as using an inflated standard deviation quantifies individual flux observations that are significantly outside the range of expected observations. Determining the variation across datasets can provide a metric to compare how moments vary across fluxes (CO2, CH4, and N2O) and across ecosystem. We defined a hot spot as any individual chamber with greater than 11% of hot moment observations within an annual flux dataset. This metric was chosen because if hot moments were evenly distributed across chambers, each chamber would represent 11% of hot moment observations. Therefore, any hot spot chamber with greater than 11% of hot moment observations would thus disproportionately contribute to the total ecosystem flux. Spatiotemporal differences in hot spots were tested with two-way analysis of variance (ANOVA) of annual mean greenhouse across chambers and site-years for each location.

The greenhouse gas flux and soil sensor datasets (corn and alfalfa only) were collected with sufficient spatiotemporal frequency to utilize machine learning approaches to determine drivers of N2O and CH4 fluxes. We used a random forest machine learning model (Hamrani et al. 2020; Saha et al. 2021) for both datasets separately to determine if increased sampling density can predict N2O and CH4 fluxes. Input variables included individual observations of N2O and CH4 fluxes and volumetric water content (VWC), O2, and temperature across a soil profile (10, 30, and 50 cm) as predictor variables. As sensors were not deployed with each individual chamber, some spatial resolution could not be accounted for with sensor observations. The relative importance of each input variable in the machine learning models were determined by utilizing a node impurity measure, which is related to the corresponding response variance in the variable of interest (Segal and Xiao 2011).


Spatiotemporal patterns in emissions across ecosystems

The three land uses differed in patterns of greenhouse gas emissions and net ecosystem exchange (Table 2). The alfalfa and pasture were net CO2e sinks over the study periods. The corn had mean annual emissions of 46.7 ± 0.8 Mg CO2e ha−1 year−1 with CO2 accounting for 58.0% of this, followed by 34.9% as N2O and 7.1% as CH4. Together, the average annual GWP of N2O plus CH4 was 19.6 ± 0.8 Mg CO2e ha−1 year−1 (Table 2). The alfalfa site was on average a net CO2e sink of -23.3 ± 0.5 Mg CO2e ha−1 year−1. Soil N2O emissions decreased the net CO2e sink by 8.0% and the average GWP of N2O plus CH4 was 1.84 ± 0.01 Mg CO2e ha−1 year−1. The pasture site had the lowest annual fluxes with -9.66 ± 0.11 Mg CO2e ha−1 year−1, with N2O decreasing the net CO2e sink by 12.8% and an average GWP of N2O plus CH4 of 1.54 ± 0.11 Mg CO2e ha−1 year−1.

Table 2 Annual mean [ecosystem respiration (Reco)] ecosystem exchange (NEE) and GWP from N2O, CH4, and total net ecosystem carbon dioxide equivalent (CO2e) budgets by site

The frequency of hot moments varied by site. In the corn, hot moments of N2O and CH4 emissions represented 1.1 ± 0.2 and 1.3 ± 0.2% of measurements, respectively, but contributed to 45 ± 1% of annual N2O fluxes and to 140 ± 9% of annual CH4 fluxes. In the alfalfa system hot moments of N2O emissions were only 0.1 ± 0.05% of annual measurements but were 31.6 ± 0.2%of the annual flux (Anthony et al. 2023). The alfalfa ecosystem did not experience significant changes in annual CH4 budgets from either hot spots or hot moments and overall was a small net CH4 sink (Fig. 4, middle). In the grazed pasture, hot moments of N2O flux accounted for 22% of total N2O emissions but were again less than 1% of measurements. Hot moments of CH4 flux were only 0.3% of fluxes but decreased the net CH4 sink by 53% (from − 24.1 to − 11.3 g m−2 year−1). Hot moments of CO2 emissions consistently contributed less to overall fluxes than hot moments of N2O or CH4. Hot moments of CO2 emissions were of the greatest importance to the total ecosystem GWP flux in the pasture ecosystem relative to the corn or alfalfa. In the pasture only 1.0% of CO2 flux measurements represented 5.4% of net CO2 emissions (Figure S6, top). In the corn system, CO2 hot moments were only 0.4% of observations but represented 4.3% of net CO2 emissions (Figure S4, top). Hot moments of CO2 fluxes were the lowest percent of total emissions in the alfalfa system, with hot moments representing 0.5% of measurements but 2% of net CO2 emissions (Figure S5, top).

Hot spots of greenhouse gas emissions across ecosystems

We used the individual chamber measurements to explore the spatial and temporal distribution of hot spots of greenhouse gas emissions across time. Across all three field sites, no individual chamber was consistently a hot spot of CO2, CH4, or N2O emissions on annual timescales (Figs. 1, 2, 3). We found significant temporal variability in the magnitude of greenhouse gas fluxes both within and cross chamber locations (p < 0.001, Figs. 1, 2, 3). For CO2, the corn site had four chambers with over 11% of hot moments, with the highest at 19.8% followed by 19.1%, 16.9%, and 14.3%. On an annual basis, the location of hot spot chambers varied across years. There were also four chambers were over 11% of the CO2 flux in alfalfa, with one hot spot representing 27.5% of all hot moment observations, followed by 21.1%, 16.4%, and 12.6% in the other chambers; hot spot chambers also varied over time on an annual time scale. The pasture site only had two chambers over 11% of the CO2 flux, with one chamber representing 65.6% of all hot moments, followed by 11.4% of observations in the other chamber. Importantly, all chambers exhibited individual hot moments of CO2 flux in both the corn and alfalfa site, while two of the nine chambers in the pasture site had no hot moments observations. However, hot moments of CO2 flux were not significant proportions of total CO2 soil flux budgets, representing a maximum of only 4% of annual soil CO2 fluxes.

Fig. 1
figure 1

Annual mean (± standard error) (left) CO2 (g CO2 m−2 day−1), (middle) CH4 (mg CH4 m−2 day−1), and (right) N2O (mg N2O m−2 day−1) fluxes by chamber from an annual corn agricultural peatland from June 2017 through June 2021. White, orange, blue, and gray bars are annual mean flux measurements with black error bars representing annual standard error

Fig. 2
figure 2

Annual mean (± standard error) (left) CO2 (g CO2 m−2 day−1), (middle) CH4 (mg CH4 m−2 day−1), and (right) N2O (mg N2O m−2 day−1) fluxes by chamber from a perennial alfalfa agricultural peatland January 2017 through January 2021. White, orange, blue, and gray bars are annual mean flux measurements with black error bars representing annual standard error

Fig. 3
figure 3

Annual mean (± standard error) (left) CO2 (g CO2 m−2 day−1), (middle) CH4 (mg CH4 m−2 day−1), and (right) N2O (mg N2O m−2 day−1) fluxes by chamber from a grazed pasture agricultural peatland April 2019 through April 2020. Red bars are annual mean flux measurements with black error bars representing annual standard error

The relative contribution of hot spots to N2O hot moment fluxes was similar to the distribution for CO2 hot spots in all three sites, with the largest individual hot moment observations for both CO2 and N2O observed in the same chambers. Hot moments of N2O emissions were relatively common, with the majority of chambers exhibiting hot moments of N2O fluxes in all three agroecosystems (Fig. 5). There were individual locations with higher annual N2O fluxes than all other chambers, but hot moments of N2O flux were observed across multiple chamber locations (Fig. 1, 2, 3). In the corn site the highest contribution was 19.1% of N2O hot moments from one chamber, followed by 16.6%, 14.0%, 12.9%, and 11.6% from four other chambers with over 11% of the hot moments. There were only three chambers with greater than 11% of N2O hot moment observations in the alfalfa site, with a single hot spot representing 32.8% of observations (followed by 22.9% and 16.4%, respectively). At the pasture site, a single hot spot represented 68.5% of N2O hot moments from one chamber that was also the largest contribution to CO2 hot spots.

Fig. 4
figure 4

Annual mean (± standard error) CH4 (mg CH4 m−2 day−1) fluxes with and without hot moments by chamber from an (top) annual corn agricultural peatland from June 2017 through June 2021, (middle) a perennial alfalfa agricultural peatland from January 2017 through February 2021, and (bottom) a grazed pasture agricultural peatland from April 2019 through June 2020

Most CH4 hot moments were generally focused in one location. In the corn site, 57.3% of all CH4 hot moment fluxes occurred in only one chamber, with all other chambers accounted for less than 11% of total hot moment observations. Alfalfa and pasture sites had similar distributions of CH4 hot spots. In alfalfa, 33.3% of hot moments were observed from one chamber, followed by 15.3% and 13.6% in two other chambers while in the pasture site 24% of hot moments from one chamber with only two other chambers over 11% of observations, at 21.3% and 16%, respectively. In the corn site hot moments of CH4 flux were predominantly driven by fluxes from a single location that contributed disproportionately to a net ecosystem CH4 flux (i.e. a hot spot, Fig. 4). In contrast, in the perennial alfalfa and grazed pasture systems, CH4 hot spots only had a small impact on the net CH4 sink (Fig. 4).

Machine learning predictions of nitrous oxide and methane fluxes

A machine learning model using the corn flux and sensor dataset was able to account for 81% of the variance observed in N2O fluxes with soil VWC, O2, and temperature sensor data across depths (10, 30, and 50 cm) (Fig. 6). Node impurity measurements suggested both VWC and temperature across depths had the highest relative importance in determining N2O fluxes. With the alfalfa agroecosystem dataset, a machine learning model was only able to account for 16% of the variance observed in N2O fluxes (Fig. 6) and soil O2, VWC, and temperature all had similar values of relative importance. Both the alfalfa and corn machine learning models systematically underfit the largest N2O fluxes (Fig. 6). For CH4 fluxes, machine learning performed significantly worse. The models were only able to account for 24% of the variance in the annual corn cropping system and 6% of the variance in the perennial alfalfa cropping system (Fig. 7).

Fig. 5
figure 5

Annual mean (± standard error) N2O (mg N2O m−2 day−1) fluxes with and without hot moments by chamber from an (top) annual corn agricultural peatland from June 2017 through June 2021, (middle) a perennial alfalfa agricultural peatland from January 2017 through February 2021, and (bottom) a grazed pasture agricultural peatland from April 2019 through June 2020

Fig. 6
figure 6

Comparison of observed N2O fluxes (nmol m−2 s−1) to machine learning predicted N2O fluxes (nmol m−2 s−1) from agricultural peatland corn (a) and (b) alfalfa agroecosystems. Predictor variables were identical for each ecosystem (10, 30, and 50 cm depths with oxygen, moisture, and temperature sensors)

Fig. 7
figure 7

Comparison of observed CH4 fluxes (nmol m−2 s−1) to machine learning predicted CH4 fluxes (nmol m−2 s−1) from agricultural peatland corn (a) and (b) alfalfa agroecosystems. Predictor variables were identical for each ecosystem (10, 30, and 50 cm depths with oxygen, moisture, and temperature sensors)


The relationship between hot spots and hot moments of emissions

Hot spots of greenhouse gas emissions can be characterized in several ways: as sites with consistently high emissions, sites with a higher proportion of hot moments than the surrounding matrix, or locations with high emissions for a defined period (month, season, year) that vary over space and time. It is well known that the rates and reactions of biogeochemical processes can vary greatly in space and time, and it is often postulated that similar mechanisms drive the occurrence of both hot spots and hot moments (McClain et al. 2003; Bernhardt et al. 2017). This often leads to the assumption that hot spot and hot moment phenomena occur at similar rates for these biogeochemical processes, thus exhibiting similar influences on elemental cycles and ecosystem greenhouse gas budgets. Here we found that the occurrence and magnitude of hot spots and hot moments of CO2, CH4, and N2O fluxes were not consistent across time or space, were not consistently distributed across ecosystems, and differed by gas. Hot moments of N2O fluxes were common across chambers in both the corn and alfalfa sites, and hot spots of emissions differed across the years in the two longer term datasets. In the pasture, there was a single hot spot of N2O emissions that accounted for 69% of the hot moments over the 1-year study period. Interestingly, the most productive hot spots of N2O emissions occurred primarily in the same chambers as high CO2 emission at all sites. This may indicate that these hot spots have high labile C availability and organic matter mineralization rates, characteristics that can support both nitrifier denitrification and canonical denitrification (Firestone and Davidson 1989; Wallenstein et al. 2006). Methane hotspots were also tied to hot moments of emissions and were largely limited to a single chamber location in the corn site. The corn site had a single chamber with a disproportionate impact on total ecosystem fluxes. Patterns for CH4 in alfalfa and pastures ecosystems were more variable over time.

Predicting hot spots and hot moments requires intensive sampling across scales

Historically, predictive N2O and CH4 models utilized controlling variables together with assumptions regarding rate constants and kinetics (Müller et al. 2007; Rütting and Müller 2007; Riley et al. 2011). Machine learning approaches are becoming more popular for predicting N2O fluxes, but their wider application has been limited by the lack of sufficient within site replication and the high sampling frequencies of relevant variables across scales needed to effectively train models. Due to methodological limitations in how we measure greenhouse gas fluxes, we are currently only able to measure these phenomena at limited scales in the field (either chamber- or eddy covariance-scales). Here we used datasets with 33,000 to 103,000 flux measurements at the chamber-scale. In the corn ecosystem, VWC and temperature across depths had the highest relative importance in determining N2O fluxes, similar to previous observations using wavelet coherence analyses (Anthony and Silver 2021). The lower predictability of N2O fluxes observed in the alfalfa ecosystem and unexplained variance in the corn system may be due to low spatial resolution of the sensor measurements and underfitting of hot moments of emissions. Both systems only had two sets of sensors per field and were installed at 10, 30, and 50 cm depths, which may or may not have corresponded to locations where hot moments of N2O production occurred, and the spatiotemporal controls on them may be more variable across a field. As both machine learning models systematically underfit the largest N2O hot moments, this is likely representative of the unaccounted spatial and temporal variance driving these extreme hot moments in both models. This underfitting may be occurring if the input soil variables are not capturing the spatiotemporal variability, and/or our limited spatial resolution during hot moment observations.

In both the pasture and alfalfa sites a low number of significant CH4 hot moments, such as those observed in wetlands or thawing permafrost (Elder et al. 2021; Rey-Sanchez et al. 2022a), is not surprising as these land uses are predominantly drained throughout the year. There are several potential explanations for the poor CH4 model performance in the machine learning models used here. First, some other unmeasured variable, such as C substrate availability or the presence of reducible Fe or sulfate species, could be a strong driver of soil CH4 flux (Chamberlain et al. 2018, 2019). Another potential explanation may be that there was insufficient spatial coverage of soil profile measurements of O2, temperature, and VWC. This would suggest that increasing spatial measurements of the variables controlling CH4 hot spots, potentially alongside each individual chamber, is needed to appropriate quantify the variables controlling the spatial variability in CH4 fluxes at the chamber scale observed.

Combining chamber-based measurements with larger scale measurements (such as eddy covariance) may help to better constrain the spatiotemporal drivers of ecosystem-scale variability. High density eddy covariance-based measurements have significantly increased the amount and availability of ecosystem-scale CH4 flux datasets (Knox et al. 2019; Delwiche et al. 2021). These eddy covariance datasets have also highlighted the increased ability to determine time-scale emergent drivers at ecosystem scales (Sturtevant et al. 2016; Chamberlain et al. 2018). However, they are generally not able to measure the dynamics and controls of CH4 fluxes at the spatial scales needed to accurately determine the distribution and drivers of hot spots and hot moments. Additionally, eddy covariance measurements of CH4 have detection limits that are too high to accurately determine CH4 consumption. Eddy covariance measurements alongside chamber measurements can provide insights to spatiotemporal variability of CH4 fluxes and help determine their importance at the ecosystem-scale. Another approach that may improve the ability to model hot spots and hot moments utilizes eddy covariance flux measurements of CH4 fluxes and footprint-weighted hot spot detection (Rey-Sanchez et al. 2022a) alongside automated chamber and soil sensor measurements at smaller scales to constrain the spatiotemporal distribution of CH4 hot spots and hot moments. Unfortunately, current eddy covariance methods for N2O measurements are often prohibitively expensive and have higher detection limits than chamber-based measurements (Tallec et al. 2019; O’Connell et al. 2022).

Drivers of nitrous oxide and methane hot spots

Drivers of soil greenhouse gas fluxes often vary across scales, from seconds (e.g., respiration) to years (e.g., succession) and from micrometers (e.g., soil aggregate) to kilometers (e.g., whole ecosystems) (Anthony et al. 2023). Interactions between soil O2 availability, temperature, moisture, and substrate concentrations are thought to drive hot spots and hot moments of soil N2O and CH4 flux (Butterbach-Bahl et al. 2013; Nazaries et al. 2013; Pärn et al. 2018). The magnitude of soil N2O and CH4 fluxes also likely depends on the previous biogeochemical conditions as these will influence the composition of the microbial community and its response to events such as soil inundation or changes in substrate availability (Wallenstein et al. 2006). High rates of denitrification require both NO3 and labile C (Dietz and Clausen 2006). In contrast, methanogenesis, as a low-energy yielding oxidation–reduction reaction, is assumed to only occur in the absence of other electron acceptors under very low redox conditions (Conrad 1989). Thus, controls on the spatiotemporal availability of these substrates likely facilitate increased variability in hot spots and hot moments.

In agricultural peatlands, these variables are often controlled by both the rate and extent of soil drainage. The amount of time that peat soils are drained, the rate of soil inundation, and soil organic matter content are directly related to land use practices. The severity of peatland degradation determines economically viable land uses, but these land uses often lead to further peatland degradation (Leifeld et al. 2019, 2020). Corresponding land use and the extent of peatland degradation may also play a role in the magnitude of hot moments. Soil NO3 concentrations accumulate under oxic, well-drained soil conditions following soil organic N mineralization (Kirk et al. 2015; Anthony and Silver 2021), and the rate of accumulation is likely dependent on both the concentration of mineralizable organic matter and the amount of time and extent of soil drainage that provides optimal conditions for nitrification. In the corn agroecosystem, the elevated organic matter content and high organic matter mineralization rates coupled with extended periods of soil inundation were likely drivers of the magnitude of hot spots and hot moments (Anthony and Silver 2021). High C and NO3 substrate availability, along with spatiotemporal variability in redox conditions driven by variations in moisture, O2, and temperature, likely contributed to the distribution of hot spots and hot moments observed. The distribution of hot spots and hot moments in the alfalfa agroecosystem likely suggests that elevated N availability from alfalfa N-fixation and short-term changes in O2 availability stimulated N2O fluxes but did not produce redox conditions suitable for CH4 production. In the degraded pasture, the ditched subsurface irrigation likely maintained anaerobic conditions throughout the soil profile, limiting hot moments of N2O flux from incomplete denitrification. The patterns observed in the pasture agroecosystem were likely to be affected by the lower C substrate availability, which may help explain the spatiotemporal variation in CH4 production and limited CH4 hot moments. Combined, these results suggest variation in substrate availability, soil moisture content, flooding duration, and associated effects on O2 availability and soil redox conditions likely contribute to the variation in hot spots and hot moments observed across these agricultural peatland ecosystems. These differences also contribute to the variation in ecosystem greenhouse gas budgets and the effects of hot spots and hot moments on these budgets observed across the three land uses in this study.

The magnitude and distribution of hot spots and hot moments in agricultural peatlands vary across time and space

The findings reported here highlight the importance of hot moments in overall N2O and CH4 budgets in agricultural peatland soils. We found significant variability in hot spots of greenhouse gas fluxes over time and space. Hot spots (here defined as a location with a disproportionate number of hot moment observations) occurred in all three ecosystems but were generally not characterized by sustained high emissions from a single location, except in the pasture site with a shorter observation period. Hot spots were more commonly characterized by high proportion of hot moments in single changes over time, or high emissions from different chambers that varied within and across years. Variability in the controls on greenhouse gas fluxes, such as soil moisture, soil O2, and substrate availability are likely driving these differences. However, poor spatiotemporal resolution of these variables or other potential unmeasured drivers are likely limiting our ability to predict the occurrence of hot spots and hot moments. Given the elevated substrate availability and predictable management-induced changes in soil redox conditions in many agricultural peatland soils, they are an ideal place to study the biogeochemical processes controlling greenhouse gas fluxes. Quantifying the drivers of this variability in agricultural peatland soils will greatly improve our understanding of the processes, controls, and distribution of hot spots and hot moments of greenhouse gas fluxes and improve land management decisions that to help mitigate climate change. Given the significant variability in hot spots and hot moments across agricultural peatland ecosystems, spatiotemporally intensive measurements of CO2, CH4, and N2O fluxes and their associated drivers are needed to appropriately constrain their respective greenhouse gas budgets. Emerging technologies are dramatically increasing soil sensor capabilities while simultaneously decreasing per sensor cost (e.g., Baumbauer et al. 2022). This can help us improve both sampling density and sampling frequency, allowing us to capture the drivers of greenhouse gas fluxes across spatial and temporal scales. Overall, these findings contribute to a better understanding of the biogeochemical patterns and pathways of soil greenhouse gas fluxes in agricultural peatland soils, highlight the considerable variability in hot spot and hot moment emissions across space and time, and the need for datasets with increased spatiotemporal resolution of both N2O, and CH4 fluxes and their controls to properly model ecosystem-scale fluxes.