Net ecosystem fluxes for a suite of hydrocarbons were measured on an hourly basis using the relaxed eddy accumulation (REA) method. REA is a micrometeorological flux measurement technique that permits in situ flux measurements for chemical species that cannot be measured at the high frequency required for eddy covariance techniques (Businger and Oncley, 1990). To date, no light alkene sensor meets the requirements for detection limit, accuracy, sensitivity and response time for eddy covariance measurements in natural ecosystems. REA systems have been successfully used for other biogenic volatile organic compounds, including isoprene (Bowling et al., 1998; Guenther et al., 1996; Haapanala et al., 2006) and OVOCs (Schade and Goldstein, 2001; Baker et al., 2001).

The REA technique is described in detail in Businger and Oncley (1990); therefore, only a brief description is provided here. Air samples are conditionally sampled into an updraft reservoir, a downdraft reservoir or a neutral bypass controlled by fast response valves that respond to high-frequency 3-D sonic anemometer measurements of the vertical wind velocity (w). Mean vertical wind velocity (w‾) is determined for a flux averaging period, and the instantaneous vertical wind velocity is calculated (w′=w(t)-w‾). The REA method is derived from the eddy accumulation method (Desjardins, 1977) but “relaxes” the requirement of sampling at flow rates proportional to the vertical wind speed. In both methods, a turbulent flux is derived from the differences between averaged concentrations in the updraft (c+‾) and downdraft (c-‾) reservoirs collected over some flux averaging period (typically 30–60 min). In the surface layer, the concentration differences are scaled by the standard deviation of w (σw) and the dimensionless Businger–Oncley parameter (b) to yield the vertical flux (Eq. 1): F=bσwc+‾-c-‾. In theoretical solutions, b was found to be a weak function of atmospheric stability (Businger and Oncley, 1990). Wyngaard and Moeng (1992) simulate b to be fairly constant (b∼0.627) assuming a Gaussian joint probability density function between w and c. Empirical approximations based on direct eddy covariance measurements show some variation in the b coefficient on a diurnal basis, and although it varies for different scalars, estimates usually fall in the range of 0.51<b<0.62 (Katul et al., 1996; Ruppert et al., 2006; Baker, 2000; Pattey et al., 1993; Baker et al., 1992). Consequently, a dynamic b value is often used, calculated for each REA averaging interval based on concurrent eddy covariance (EC) measurements of a proxy scalar under the assumption of scalar similarity (Pattey et al., 1993). In this case, c is replaced with the proxy scalar of temperature, measured by the sonic anemometer. The value of b can be calculated from the sonic temperature and by rearranging Eq. (1) as follows (Eq. 2): b=w′T′‾σwT+‾-T-‾, where (w′T′‾) is the covariance between instantaneous fluctuations of w and temperature, i.e., the heat flux, averaged over the chosen time interval and (T+‾,T-‾) values are the mean temperatures during updraft and downdraft sampling, respectively. Ruppert et al. (2006) investigated scalar similarity between water vapor, sonic temperature and carbon dioxide and found a diurnal pattern in scalar correlation coefficients leading to an error of FREA≤10 %.

To increase the accuracy of conditional sampling and maximize the signal-to-noise ratio in Δc‾, samples during very small w′ are discarded via a neutral bypass as part of a “deadband” (Baker, 2000). For each flux averaging interval, a symmetrical threshold (w0) around the mean wind velocity is applied, whereby the updraft reservoir is sampled when w′≥w0 and the downdraft is sampled when w′≤-w0. Oncley et al. (1993) analytically solved the ratio between an increase in the uncertainty of c‾ due to shorter sampling intervals with increasing w0 and an improvement in the signal-to-noise ratio; they report an optimum at w0=0.6σw, which was used in this study. For each flux averaging interval, the Businger–Oncley parameter is computed from Eq. (2) using the same deadband. The deadband-related increase in ΔT‾ consequently leads to smaller b values that are ∼ 0.4.

In REA measurements, both w‾ and σw need to be initialized in real time to determine what constitutes an updraft and downdraft within each flux averaging interval. Based on the analysis of Turnipseed et al. (2009), we chose to use w‾ and σw from the previous flux averaging interval.

The physical REA instrumentation consists of two subsystems: (1) an air sampling subsystem to segregate the sample flow into an up- and down-line (or neutral bypass line) according to the vertical wind velocity and (2) a reservoir system, for storage, transfer and evacuation of the sampled air (Fig. 2). The subsequent description follows the flow of air through the system.

The air sampling subsystem consisted of a sonic anemometer and segregator box, both mounted 25.1 m a.g.l. on the end of a 1.2 m boom (metal cross beam) extending outward from the top level of the walk-up chemistry tower. Vertical wind velocity was measured with an ultrasonic anemometer (model 81000; R. M. Young, Traverse City, MI, USA), which transmitted data at a 5 Hz frequency via RS-232 to a CR-1000 data logger (Campbell Scientific Inc., Logan, UT, USA). A 75 cm long 1/8′′ outer diameter by 1/16′′ inner diameter PTFE tube (EW-06605-27; Cole Parmer, Vernon Hills, IL, USA) was attached to the sonic anemometer (horizontal offset ≈ 0 cm and vertical offset = 10 cm with respect to the center of the anemometer's measurement path). Sample air was drawn into the segregator box (also mounted on the boom) via a micro-diaphragm pump (UNMP805; KNF Neuberger Inc., Trenton, NJ, USA), with airflow restricted by a stainless steel needle valve. The segregator split the airflow into an up-line, down-line and neutral line by two logger-controlled PTFE diaphragm solenoid valves (Vup and Vdn, Fig. 2; 100T3MP12-62M; Bio-Chem Fluidics Inc., Boonton, NJ, USA). The neutral line was activated when vertical wind velocities fell into the deadband (see Sect. 3.1 above). Neutral airflow was directed through an airflow sensor (AWM3300V; Honeywell International Inc., Morris Plains, NJ, USA) and finally vented out of the segregator.

For the sample set being measured, air from each bag was transferred sequentially (18 min each) through solenoid valves V4u or V4d (Fig. 2). Two sample lines (1/4′′ PTFE tubing wrapped in foam insulation) extended down to the laboratory trailer at the base of the tower, and air samples were drawn from the reservoir bags to the gas chromatograph (see next section).

To address the potential issue of different storage time in the bags, the order of sample analysis alternated between each hourly flux sampling interval (e.g., 13:00: up bag, down bag; 14:00: down bag, up bag). After the transfer, airflow to the GC was shut off and the remaining air in the up or down reservoir bag was evacuated for 15 min through solenoid valve V3u or V3d using a vacuum pump (UNMP805; KNF Neuberger Inc., Trenton, NJ, USA; Fig. 2), with less than 2 % carryover from one sample to the next; additional details are described in the Supplement.

Real-time measurements of vertical wind velocity (w) were collected on a data logger (CR1000; Campbell Scientific Inc., Logan, UT, USA), which also relayed the signal following the sampling lag time (see Supplement) to control the segregator sampling line valves, Vup and Vdn, accordingly. The high-frequency time series of sonic temperature (T) were stored in the data logger's memory for subsequent calculation of the covariance of w and T: (w′T′‾). Sonic temperature was also conditionally averaged into T+‾ and T-‾ for calculation of the b coefficient (Eq. 2). At the end of each flux averaging interval, w‾ and σw were calculated by the data logger and used to initialize the deadband for the following sampling hour and to compute the instantaneous fluctuations of vertical wind speeds (w′). In addition, the logger also triggered the bag selection valves (V1 and V2) when switching to the other pair of updraft and downdraft reservoirs (set A vs. set B bags; Fig. 2). For quality control, the volume of sampled air in each bag, the volume of expelled neutral air and the average sampling flow rate were saved on the data logger's memory. Quality control for each hourly REA flux measurement was checked against eight potential flags associated with the sample volumes, meteorological conditions or footprint analysis (Fig. S1, Supplement).

Flux detection limits (Fmin) were calculated by using Eq. (3): Fmin=bσw2σc_SD, where 2 σc_SD is the analytical precision based on 2 standard deviations of hourly repeated GC-FID runs of the calibration standard (see Sect. 3.6 below). The lowest flux detection limit (LDL) was determined for isoprene (Fmin=3.4 µg m-2 h-1), followed by ethene and butene (Fmin=4.1 µg m-2 h-1) and propene (Fmin=4.7 µg m-2 h-1). Flux observations that were negative or below Fmin were included in the overall statistical analyses (median and percentiles, means and standard deviations) but excluded for the curve fitting in response to temperature and PAR. The number of fluxes < LDL varied as follows: ethene (n=12), propene (n=33), butene (n=93), isoprene (n=105), acetylene (n=380) and benzene (n=158).

Understory flux measurements were performed on a single day, 2 September 2014 (about 1 month after the main experiment), to provide insight on the magnitude of fluxes that may be emanating from the surface instead of the tree canopy. These understory fluxes were measured by mounting the REA sampling system to a separate smaller scaffold, with the inlet line and sonic anemometer placed at 2 m a.g.l. Hourly fluxes were measured starting at 06:00 and ending at 17:00, with the up and down bag samples being transferred to electropolished stainless steel canisters for later analysis in the laboratory on the same gas chromatograph used during the field season.

The challenge with understory measurements is that they are prone to sampling artifacts due to flow distortion and low wind speeds. Furthermore, turbulence tends to be intermittent, and there is a lack of universal theories on sub-canopy turbulence characteristics, i.e., (co)spectral models (Launiainen et al., 2005).

In this study, the understory turbulence (defined here as the standard deviation of vertical wind) evolved over the course of the day from 0.04 m s-1 at night and in the early morning to over 0.1 m s-1 at 09:00 MST to a maximum of ∼ 0.4 m s-1 (Fig. S2). In previous sub-canopy flux studies, a σw mixing criterion was empirically determined at 0.1 m s-1 (Launiainen et al., 2005). Thus, measured fluxes in periods with insufficient mixing (small σw) do not represent the real surface–atmosphere exchange. Our observations support the use of a similar criterion: sensible heat fluxes were highly variable under low turbulence conditions but showed weak dependence on σw with increasing σw. A site-specific σw threshold was determined at 0.4 m s-1.

Flux measurement time series are often fragmented due to questionable turbulence statistics, unfavorable wind directions or sensor failure. Hence diurnally or seasonally averaged fluxes can be biased if time series are not gap filled. Gap filling the REA-derived fluxes was performed here using an artificial neural network (ANN) approach (Moffat et al., 2007; Papale et al., 2006). ANN is increasingly used in eddy covariance studies because of its ability to resolve nonlinear relationships and complex interactions between flux drivers (Dengel et al., 2013; Papale and Valentini, 2003). Input variables included air temperature, photosynthetically active radiation, water vapor flux and standard deviation of the vertical wind speed. Prior to gap filling, input variables were normalized on a scale of -1 (for minimum value) to +1 (for maximum value). Inputs variables (n=1223 each) were then divided into k=20 clusters via the k-means method, a cluster analysis tool that partitions n observations into k≤n clusters by minimizing the inner-cluster variance. From those clusters, explanatory data were proportionally sampled into train, test and validation subsets. This procedure aims at avoiding a bias in network training towards data subsets with better data coverage. In total, 20 extractions out of these subsets were performed and run for 5 network architectures with increasing complexity. The best architecture for each of the 20 extractions was chosen according to the lowest root mean square error (through comparison with the validation subset, which is not used for training the networks) and the lowest complexity and then used to compute a predicted flux. Gap filling was finally performed using the median of the 20 resulting predictions.

Goodness of prediction was quantified by using coefficients of determination (r2) between median prediction and measured data, as well as by root mean square error (RMSE). For ethene r2=0.70 and RMSE = 32.1 µg m-2 h-1, for propene r2=0.71 and RMSE = 27.7 µg m-2 h-1, for butene r2= 0.80 and RMSE = 8.6 µg m-2 h-1 and for isoprene r2=0.3 and RMSE = 38.9 µg m-2 h-1. The lower performance for isoprene was due to the difficulty in predicting intermittent large negative fluxes.

Hydrocarbons (C2-C5 alkenes including isoprene, C2-C6 alkanes, acetylene and some aromatics) were measured with a gas chromatograph with a flame ionization detector (GC-FID; Fig. S3). The automated GC-FID was originally developed for aircraft operation, with 45 hydrocarbons resolved on the capillary column with a detection limit of 2 to 5 ppt for a 350 cm3 STP sample (Goldan et al., 2000; Kuster et al., 2004). The system was modified here to optimize light hydrocarbon measurements using 20 min run times, and calibration standards were analyzed between sample runs to produce daily calibration curves, from which concentrations were derived (Supplement). This study focused on ethene, propene, isoprene, acetylene, benzene and the three butene isomers (trans-2-butene, 1-butene and cis-2-butene), which were all well resolved by the chromatography. However, the trio of butene isomers had retention times that were clustered together, and these were all present in equal amounts in the calibration standards. Only one of the butene isomers showed consistently significant signals in this study, and this compound was identified tentatively as cis-2-butene based on its retention time. This compound is reported in this study as “butene” to account for its molar mass and chemical makeup while allowing for the uncertainty of the specific isomer being measured (Supplement).

Between 2009 and 2014, turbulent fluxes of CO2, water, heat and energy were measured at MEFO (Ortega et al., 2014) using the eddy covariance (EC) method (Baldocchi et al., 1988). An ultrasonic anemometer (CSAT3; Campbell Scientific, Logan, UT, USA) was mounted at 25.1 m of measurement height, along with a weather transmitter (WXT520; Vaisala, Vantaa, Finland) to measure absolute temperature and relative humidity. Air was drawn from the tower through a Teflon inlet line into the trailer and measured for CO2 and water vapor measurements using a closed-path IRGA (Li-7000; Licor Biosciences, Lincoln, NE, USA). In this study, fluxes were averaged for 30 min intervals and underwent a quality control scheme including a test on stationarity and on the integral turbulence statistics (Foken and Wichura, 1996). Fluxes from periods failing both tests were removed from the data set (13 %); data failing only one test were flagged (53 %).

Analysis of the tower's suitability for micrometeorological measurements was performed previously during the BEACHON campaigns (Kaser et al., 2013a). Flux source regions (i.e., the flux footprint) for this campaign were computed using an analytical model (Hsieh et al., 2000), and the median 90 % flux footprint recovery during unstable (blue) and stable (green) atmospheric conditions was spatially mapped (Fig. 3); 90 % flux recovery stretched up to 1400 m (median 670 m) upwind from the tower for unstable atmospheric conditions and 5000 m (median 2200 m) for stable atmospheric conditions. Data from easterly winds were flagged for suspicious footprints due to the presence of a lightly traveled paved highway approximately 500 m away. Further data with 90 % flux recovery exceeding 1.9 km were flagged due to possible source–sink inhomogeneity.

Ambient alkene concentrations, calculated as the average of the up and down bag reservoirs for the same hour-long period and reported as the end time, showed large fluctuations over the course of the field campaign (Fig. 4). Median and mean daily concentrations were the highest for ethene (318 and 303 ppt, respectively), followed by propene (176 and 182 ppt), isoprene (115 and 148 ppt), acetylene (79 and 86 ppt), butene (52 and 51 ppt) and benzene (43 and 44 ppt; Tables 1 and S2).

Ethene, propene, butene and isoprene concentrations exhibited clear diurnal cycles; the lowest concentrations were observed at nighttime, with a minimum typically occurring between 04:00 and 07:00 MST (Fig. 5, red points). From 07:00 MST onwards, concentrations sharply increased and reached maxima at 13:00 MST for ethene and propene. Butene and isoprene were also elevated during midday, although concentration peaks were not as pronounced. During the afternoon, all of these compounds showed a slow decrease towards the nighttime minima. In contrast, benzene showed only a minor enhancement in concentration during the daytime, and acetylene concentrations showed no measurable diurnal cycle (Figs. 4 and 5).

Gaps in the measurement period complicate the picture for larger-timescale fluctuations in concentrations. The highest concentrations for ethene, propene and butene occurred between days 198 and 206 during midday. Ethene and propene also had high concentrations in the early measurement period between days 176 and 181 when butene concentrations were not monitored. The highest daytime isoprene concentrations occurred between days 200 and 208, also during midday. Acetylene had two periods of higher concentrations, between days 197 and 201 and days 220 and 223, with the highest concentrations occurring either in the daytime or at night. Benzene showed no obvious temporal trends.

Approximately 450 net fluxes (Fig. 6) were quantified over the course of the summer, of which 19 % were critically flagged and omitted from further analysis (Supplement). Ethene had the largest overall median and mean flux (46 and 71 µg m-2 h-1, respectively), followed by propene (36 and 59 µg m-2 h-1), butene (12 and 23 µg m-2 h-1) and isoprene (0.6 and 14 µg m-2 h-1; Tables 1 and S2).

The time series of alkene fluxes show distinct diurnal patterns of emissions that are similar for ethene, propene and butene (Fig. 5, blue points). Median and mean daytime emissions were large for ethene (123 and 123 µg m-2 h-1, respectively), followed by propene (95 and 104 µg m-2 h-1), butene (39 and 44 µg m-2 h-1) and isoprene (17 and 32 µg m-2 h-1), but these elevated fluxes were concentrated between 10:00 and 17:00 MST. In general, light alkene fluxes were low (but generally positive) at nighttime, with a rapid rise during the morning and a rapid drop in the evening. Isoprene fluxes on average showed a similar pattern but decreased earlier in the afternoon (15:00 MST) and had roughly zero flux at nighttime. In contrast, acetylene and benzene showed no diurnal flux patterns and scatter around zero: 1±13 µg m-2 h-1 for acetylene and -2±17 µg m-2 h-1 for benzene (Fig. 5, Table 1).

In addition to the diurnal patterns, multiday (∼ 5 day) fluctuations were visible in the measured peak daytime fluxes for the alkenes (Fig. 6). Daytime maximum emissions rose and fell 50 % between days 198 and 205 and again between days 215 and 220. The pattern resembles the broad temporal trends in temperature, radiation and water flux (Fig. 6).

Gap filling REA fluxes (Fig. 6) using artificial neural networks (i.e., modeled results) removes the temporal bias in averaging the quality-controlled observations. The ANN-derived gap filling of missing hourly data yields 20 % higher median (Table 1) and 7–8 % higher mean (Table S2) emission rates for the light alkenes. However, these differences between groups of modeled and observed fluxes were nonsignificant (ANOVA, α=0.05), suggesting that the selectivity of quality-controlled measurements might lead to only a minor underprediction of diurnal averages. When negative alkene fluxes were measured, they usually failed quality control owing to stable nocturnal atmospheric conditions; however, a limited number (small proportion) of quality-ensured fluxes suggest apparent uptake at night, with n=12 (3.3 %) for ethene (-48.3 µg m-2 h-1), n=24 (8 %) for propene (-28 µg m-2 h-1), n=12 (3.1 %) for butene (-8.7 µg m-2 h-1) and n=124 (34 %) for isoprene (-20.9 µg m-2 h-1) being larger than flux detection limits. Negative fluxes were too infrequent and small to be captured in ANN model predictions for the light alkenes (Table 1).

Over the sampling period (24 June–9 August 2014), Manitou Forest acted as a net CO2 source of 2.6 g m-2 d-1 on average (Fig. 6). Characteristic diurnal flux patterns show nighttime to morning respiration (2–8 µmol m-2 s-1) and net CO2 uptake (up to -8.6 µmol m-2 s-1) between 09:00 and 18:00 MST. A simple one-level storage term evaluation was performed (Rannik et al., 2009). The venting of stored CO2 was on the order of magnitude of measured EC fluxes in the morning (06:00–08:00 MST), leading to apparent emission during the onset of turbulence. Storage occurred at night (19:00–24:00 MST), leading to an underrepresentation in measured nighttime respiration on the order of ∼ 25 %. Over the course of a day, the positive and negative storage terms cancel each other out.

The diurnal CO2 flux cycle increased in amplitude following the onset of significant seasonal rainfall. In the first half of the measurement period, 24 June through 11 July (DOY 175 through 192), daily maximum and minimum CO2 fluxes were relatively small, averaging 4.4±1.4 and -3.1±1.7 µmol m-2 s-1, respectively. Following a strong rain event on 12 July (DOY 193, between 15:00 and 17:00 MST), these averaged 7.5±2.4 and -5.8±1.6, respectively, through the end of the campaign on 9 August (DOY 193 to 221). During this latter time period, numerous significant rainfall events also occurred (Fig. 6).

H2O fluxes have a characteristic diurnal pattern, with negligible fluxes during nighttime, a sharp increase during sunrise (07:00 MST), maxima at 12:00 MST and a steady decrease during afternoon. On overcast days, peak emissions were on the order of 1.2 mmol m-2 s-1, whereas on sunny days fluxes reached up to 7.8 mmol m-2 s-1. H2O storage was found to be negligible. As with CO2, the amplitude of water vapor fluxes increased from 12 July (DOY 193) onwards. Average daily maximum water vapor fluxes were 2.9±0.2 and 5.4±0.7 mmol m-2 s-1 for the measurement periods before and after 12 July, respectively.

Sensible heat fluxes (HS) ranged from -100 to 500 W m-2. Typical diurnal patterns indicated nighttime inversions from 20:00–07:00 MST and peak emissions at 12:00 MST. Computing the Bowen ratio (B = sensible heat divided by latent heat fluxes) gives insight into the ecosystem's response to water availability. In the dry period prior to day 193, B was strictly > 1 (median = 2), which is typical for semiarid water-limited ecosystems. During this time, evaporation was restricted, favoring elevated sensible heat flux. After rainfall events, B dropped below 1 (median = 0.4) due to higher latent heat fluxes and hence less sensible heat flux.

For the following analysis, correlations are quantified between two independent variables using the Pearson correlation coefficient (ρ). Above-canopy concentrations of ethene, propene and butene were highly correlated (0.73<ρ≤0.88), whereas correlations including isoprene were slightly weaker (ρ= 0.4–0.5; red and black dots, Fig. 7). Concentrations of these light alkenes and isoprene were poorly correlated with those of acetylene or benzene (ρ<0.5); however, benzene and acetylene showed a strong correlation with each other (ρ=0.78). For the correlated pairs, median molar concentration ratios were propene / ethene (0.55), butene / ethene (0.18), butene / propene (0.31) and benzene / acetylene (0.51).

Similar to the concentrations, the net fluxes of ethene, propene and butene showed high correlation coefficients with each other 0.52<ρ≤0.93, whereas correlations with isoprene, acetylene and benzene were weak ρ<0.2. Unlike their concentrations, benzene and acetylene fluxes were not correlated (ρ=0.24). The strong correlation between ethene and propene fluxes was particularly notable (ρ=0.93). The median of mass flux ratios (excluding those < LDL) were propene / ethene (0.87), butene / ethene (0.31) and butene / propene (0.35).

The understory flux measurements on 2 September 2014 can help partition the above-canopy fluxes between surface and canopy sources. Of the 10 REA flux samples collected that day, 8 flux samples exceeded the σw threshold of 0.4 m s-1; the 2 samples that fell beneath the threshold occurred during the early morning hours (Fig. S2). For the light alkenes, the understory fluxes greatly contrasted with the above-canopy fluxes. The understory REA measurements showed detectable consumption overall for ethene, propene and butene as opposed to the large emissions observed from the above-canopy fluxes (Table 1, Fig. S3).

In contrast, the isoprene, acetylene and benzene fluxes were in similar ranges to the above-canopy fluxes. Isoprene showed relatively large emissions during the day at the surface, which are in the upper range of observed daytime emissions from the above-canopy measurements. Acetylene and benzene showed small fluxes that scattered around zero, similar to the above-canopy measurements.

While isoprene is well known to be a biogenic volatile organic compound, the biogenic sources for the light alkenes are not as well determined. In this study, ethene, propene and butene appear to originate from local sources that are also biogenic in origin, in particular from the forest canopy.

The large diurnal fluctuations of both ambient concentrations and net fluxes of the alkenes follow sunlight and temperature cycles, which is typical for biogenic VOCs. For example, prior studies at Manitou Forest showed that summertime VOCs with diurnal cycles were predominantly biogenic, with the highest contributions from 2-methyl-3-buten-2-ol (232-MBO or MBO), methanol, ethanol, acetone, isoprene and, to a lesser extent, monoterpenes (mostly α-pinene, β-pinene and Δ-3-carene; Kim et al., 2010; Greenberg et al., 2012). Diurnal patterns of alkene concentrations agree with observations of the sum of MBO + isoprene. Monoterpene emissions are biogenic but occur throughout the day and night; their diurnal concentration pattern is inverted, with a buildup in the shallower boundary layer over nighttime and depletion during daytime, the latter due to a combination of dilution in the growing boundary layer and reactivity with O3 and OH (Kaser et al., 2013b).

In contrast, no such diurnal patterns in concentration are observed for the primarily anthropogenic compounds (acetylene and benzene), and their fluxes are near zero (Table 1). Consequently, correlations between the light alkenes and either acetylene or benzene are poor (concentrations) or nonsignificant (fluxes). Acetylene is considered to be a tracer of combustion originating from biomass burning or urban areas (Xiao et al., 2007). The two general periods of elevated ambient acetylene concentrations, between days 197 and 201 and days 220 and 223, did not correspond to the highest concentrations of the light alkenes. Also, elevated acetylene concentrations typically occurred at nighttime, not at midday like the biogenic VOCs. Benzene appeared to have a slight diurnal fluctuation, but this compound may also have a minor biogenic source in addition to its anthropogenic sources (Misztal et al., 2015). In prior studies at Manitou Forest, it was shown that on days with long-range transport from the Front Range cities (Colorado Springs, Denver), anthropogenic VOCs were present, although typically at low concentrations, and no significant local anthropogenic emissions were detected in the area around the site (Ortega et al., 2014).

The REA method requires a measurable concentration difference based on vertical winds. Thus, the observation of alkene emissions points to a local source, and the flux footprint during the daytime is predominantly ponderosa pine forest. The vertical concentration gradient of any source outside of the flux footprint would be erased because of mixing by the time it reached the tower, perhaps generating elevated concentrations but no measurable flux. The benzene and acetylene measurements support this; elevated concentrations in ambient air were occasionally observed for these compounds, presumably from distant anthropogenic sources, but they were not associated with emissive fluxes at the site.

The understory measurements demonstrate that these light alkenes are emitted from the forest canopy, not from the surface litter or soils (Table 2). In fact, light alkenes showed a small downward flux to the surface, suggesting potential consumption. Very small emission rates of light alkenes from a boreal forest floor in Finland (< 1.8 µg m-2 h-1 for ethene, < 0.5 µg m-2 h-1 for propene and < 0.05 µg m-2 h-1 for cis-2-butene; Hellén et al., 2006) may also be consistent with the present study, given that the light alkene emissions appear to be from the canopy, not from the forest floor.

In contrast to the light alkenes, surface isoprene emissions were relatively large and comparable in magnitude to the above-canopy emissions during the growing season. The understory included grasses and herbaceous flower plants (forbs), which were not predicted to be significant sources of isoprene. Leaf and needle litter emissions of BVOCs were measured from ponderosa pine (the dominant tree species) at Manitou Forest previously, and a compound with the ion m/z=69 (such as isoprene) was measured using PTR-MS. This compound was tentatively identified as pentanal because of the lack of known isoprene-emitting vegetation at the site (Greenberg et al., 2012), but our measurements suggest that a small local isoprene surface source exists. The relatively small fluxes of isoprene are consistent with BEACHON campaign measurements, which showed that isoprene amounted to ∼ 10–20 % of MBO concentrations at Manitou Forest (Karl et al., 2014). Benzene and acetylene show negligible fluxes in the understory, similar to above-canopy fluxes. Taken together, these observations suggest that the canopy is the source for the light alkenes and the understory is a source for isoprene.

A direct comparison between tower-based and understory fluxes cannot be made because only one REA system was available. However, the light (1300–1700 µmol m-2 s-1) and temperature (20–26 ∘C) conditions during the understory measurements on 2 September 2014 can be inserted into the temperature and PAR parameterizations from the tower-based measurements to calculate expected fluxes (Sect. 5.4). Doing this yields a predicted isoprene emission of 91±57 µg m-2 h-1, which is within 20 % of the averaged measured understory flux (Tables 1 and S2) and supports the hypothesis that the understory is the dominant source for isoprene.

We can assess the relative importance of light alkenes in the overall emission of reactive VOCs from this ponderosa pine ecosystem by comparing the light alkene emissions measured in this study with the other BVOCs measured during the BEACHON campaigns. In order to do this, it is important to place 2014 in the context of prior years using ecological parameters measured across all of these years. Eddy covariance flux measurements of CO2 and heat allow for this type of comparison: CO2 fluxes, PAR and net radiation flux observed from June–August 2014 (Fig. 6) were similar to observations made during the 2008–2013 BEACHON campaigns both in magnitude and seasonal pattern (Ortega et al., 2014). For example, the summer net ecosystem exchange (NEE) is usually positive, while the spring NEE is negative. Also, the increase in CO2 emissions following the onset of precipitation has been observed at this site in previous years. This has been attributed to the “Birch effect” found in semiarid, Mediterranean and African ecosystems, whereby precipitation triggers a burst of organic matter decomposition with subsequent CO2 emissions, significantly reducing or inverting NEE in forest ecosystems (Jarvis et al., 2007).

The overall seasonally averaged sum of ethene, propene and butene flux measurements is ∼ 150 µg m-2 h-1, and this amount is substantial even in comparison to the other BVOCs previously measured at the site. For example, the daytime average (10:00–18:00 MST) flux of combined light alkenes was ∼ 270 µg m-2 h-1. This is approximately 15 % of the combined MBO + isoprene flux of 1.84 mg m-2 h-1 (combined because the PTR-MS measurements were not able to fully discriminate between these compounds), and it is two-thirds of the methanol emissions (0.42 mg m-2 h-1; Kaser et al., 2013a). Thus, the light alkenes contribute a significant amount of reactive carbon to the atmosphere at this coniferous forest ecosystem and may even play a bigger role in ecosystems that do not emit MBO.

To assess the relative importance of the light alkenes and isoprene to the total OH reactivity of the BVOCs, we utilized the daytime fluxes from this study compared with the MBO, methanol, monoterpene, acetic acid, glycolaldehyde, acetaldehyde, ethanol, acetone, propanal and formic acid fluxes reported previously for this site (DiGangi et al., 2011; Kaser et al., 2013a). Multiplying the mixing ratios of these compounds by their OH rate constants provides a measure of relative OH reactivities (Ryerson et al., 2003; Fantechi et al., 1998; Ravishankara and Davis, 1978; Atkinson et al., 1986, 1997; Baulch et al., 1994; Huang et al., 2009; Picquet et al., 1998). We utilized fluxes instead of concentrations to provide a measure of OH reactivity that is independent of elevated concentrations associated with pollution events and more representative of site-specific sources. Accordingly, the dominant BVOC for OH reactivity is MBO, accounting for 65 %, followed by monoterpenes at 11 % and isoprene at 5 % (Fig. 8). Ethene, propene and butene accounted for 3, 5 and 4 % of the OH reactivity, respectively. Combined, the light alkenes accounted for 11.6 % of the total OH reactivity, which is comparable to the monoterpenes and second only to MBO. Thus, the light alkenes are an important component of the atmospheric chemistry of ponderosa pine forests. It is possible that unmeasured or underestimated emissions of the light alkenes can contribute to the problem of missing OH reactivity observed in other forests, as the reactive source for the missing OH has the temperature response characteristics of a BVOC (Di Carlo et al., 2004; Mogensen et al., 2011; Nölscher et al., 2013).

Net ecosystem fluxes of light alkenes have been reported for one other forested site: a temperate deciduous forest in Massachusetts (Harvard Forest; 42∘ N, 72∘ W; Goldstein et al., 1996). Using a flux gradient method, average emission fluxes were derived for ethene, propene and butene (1-butene) of 44.1, 28.4 and 13.8 µg m-2 h-1, calculated as the integrated mean diurnal fluxes between 1 June and 31 October 1993. In the present study, observed Manitou Forest emissions were larger by factors of 1.6 to 2.1 (71.3, 59.0 and 22.8 µg m-2 h-1, respectively). However, this study focused on the summer months of July–August 2014, and the much higher fluxes are partly a consequence of averaging fluxes over a period of higher temperature and PAR. A simple extrapolation for the whole season at Manitou Forest, assuming linear increases and decreases from/to zero during the shoulder months, still yields 30–70 % larger seasonal fluxes, suggesting that the coniferous Manitou Forest indeed emits more per unit area than the deciduous Harvard Forest. A more detailed model extrapolation for the shoulder season is applied in Sect. 5.5.

In both studies, the fluxes of these alkenes were correlated with each other, although with slightly different ratios. Goldstein et al. (1996) report molar ratios of emissions of ethene and butene versus propene of 1.8±0.22 (SD error) and 0.41±0.06, respectively, whereas this study yielded ratios of 1.1±0.17 and 0.52±0.14 (SD error), respectively. While the butene / propene ratio appears to be similar, a key difference is that the butene isomer identified by Goldstein et al. (1996) was 1-butene, whereas in this study the butene isomer is tentatively identified as cis-2-butene.

Strong diurnal cycles of ethene, propene and butene fluxes were observed in both forests, but MEFO fluxes more closely tracked temperature than incident light, whereas Harvard Forest exhibited the reverse. This was illustrated both by the temporal synchronicity and the stronger correlation between the alkene fluxes and ambient temperature (for MEFO) or PAR (for Harvard Forest; see Sect. 5.4). At MEFO, ambient temperature usually peaked 1–2 h after PAR starts declining, similar to the alkene fluxes.

A brief comparison can be made with other observed biogenic emissions of light alkenes. Ethene emission rates from plant shoots compiled by Sawada and Totsuka (1986) averaged 1.5 ng of ethene per gram fresh weight (gfrw) per hour, with a range of 0.6–3.2 ng (gfrw)-1 h-1. Emission rates were combined with biomass and surface area estimates of biomes to derive a net areal flux from coniferous forests for the growing season of 29.8 µg m-2 h-1 from plant shoots and leaves. This is roughly 40 % of the average (71 µg m-2 h-1) and 65 % of the median (46 µg m-2 h-1) ethene flux measured here. Given the fact that the prior study was based largely on a very limited number of laboratory incubations of non-arboreal species, it is remarkable that the emission rates are within a factor of 3 of each other. On the other hand, the emission rates from coniferous forests during the warmest part of the summer appear to exceed the previously assumed upper range of emissions.

There is a striking similarity in the multiday patterns observed in both the biogeochemical fluxes and environmental parameters at MEFO. The mesoscale temporal patterns in the fluxes are illustrated by a rise and fall of peak midday values (Fig. 6), such as the one occurring between DOY 198 and 212 followed by another between DOY 202 and 223. A similar pattern is evident in the peak midday H2O flux, the maximum daily air temperature and the net radiation and/or PAR. These trends were measured independently with separate instruments using different methods. The relationship between the fluxes and environmental parameters suggests that sunlight and temperature control the variability in the alkene fluxes and evapotranspiration rates.

To describe temperature and light responses, alkene fluxes have been averaged into bins of (a) 200 µmol m-2 s-1 PAR and (b) 2 ∘C temperature classes (Fig. 9). The light response flux, F(PAR), was parameterized according to Eq. (4) (Harley et al., 1998): F(PAR)=αCL1PAR1+α2PAR2×F1000, where α and CL1 are empirical coefficients (Table 2), PAR is the photosynthetically active radiation (µmol m-2 s-1) and F1000 is the observed flux at PAR = 1000 µmol m-2 s-1. This relationship was originally developed for emissions of isoprene, which has light-dependent production.

The temperature response flux was divided into light-independent and light-dependent fractions. The light-independent fraction (LIDF) of the temperature emission response refers to volatilization processes that do not depend on light but are still temperature dependent, such as the volatilization of pools of organics stored within plant tissues. The flux of the light-independent fraction of temperature responses, F(TLIDF), was parameterized according to Eq. (5) (Schade and Goldstein, 2001): FTLIDF=Fref×exp⁡βT-Tref, where β is an empirical coefficient (Table 3), T is the ambient temperature (∘C) and Fref is the observed flux at reference temperature Tref=30 ∘C. For the purposes of comparison, fluxes have been normalized to equal 1 at a temperature of 30 ∘C prior to response curve fitting (see Table 3). Temperature responses in the 0 to 30 ∘C range follow an exponential function, are fairly similar between individual alkenes and agree well with several other BVOCs, such as methanol, ethanol, acetone, acetaldehyde, monoterpenes and α-pinene (Table 3 and Fig. 9).

The light-dependent fraction (LDF) of the temperature emission response refers to the emission of compounds that have recently been produced and emitted without being stored. The flux of the light-dependent fraction of temperature responses, F(TLDF), was parameterized according to Eq. (6) (Schade and Goldstein, 2001; Guenther et al., 2012): FTLDF=EoptCT2eCT1xCT2-CT11-eCT2x,x=Topt-1-T-1R, where CT1 and CT2 are empirical coefficients, Eopt is the maximum emission capacity at temperature Topt , which was set to 312 K, and R is the universal gas constant (Table 3). This relationship is similar to the ones governing MBO emissions, which are considered to have a light-dependent temperature response curve (Fig. S5).

Emissions show a strong relationship to PAR (Fig. 9), although both temperature response curves showed higher correlation coefficients than the light response curves (Table 2 vs. 3). The curve fits of the temperature emission response using the light-dependent equation (Eq. 6) are slightly better than the fits using the light-independent fraction equation (Eq. 5), suggesting that the alkene emissions have a high light-dependent fraction (LDF). However, the range of temperatures in this study is within the range in which both temperature response curves are similar, thus limiting the assessment of which equation performs better at high temperatures (Figs. 9 and S5).

The similar response curves to other BVOCs further suggest that these alkenes are biogenic in origin and emitted from the canopy during photosynthetically active periods. The MBO flux profile measurements show that MBO emissions are light dependent and increase with height up to 12 m (Karl et al., 2014; Ortega et al., 2014). Ethene, propene and butene flux responses show an almost linear increase at PAR < 1000 and asymptotic behavior at PAR ≈ 2000 µmol m-2 s-1. The isoprene light response, on the other hand, showed less of an asymptote at high PAR. It should be noted that the PAR measurements employed to compute the light response curves were measured above the canopy, while the observed source of isoprene appears to be in the vegetated understory, which experiences more diffuse light. In fact, PAR intensity measured near ground level (2 m a.g.l.) was on average 50±30 % (standard deviation) of the measured PAR above the forest canopy. Hence, the sub-canopy isoprene source(s) may experience an optimum quantum yield at much larger incident PAR (measured above the canopy) than the other alkene source(s) within the ponderosa pine canopy, explaining the different light response curves.

The light alkenes (ethene and propene) are included in the Model of Emissions of Gases and Aerosols from Nature version 2.1 (MEGAN 2.1), which is used to determine the BVOC input into the atmosphere from terrestrial and oceanic ecosystems. Perhaps the best-characterized BVOC in MEGAN 2.1 is isoprene, and it is noteworthy that the modeled parameters for isoprene flux in this study are in excellent agreement with MEGAN 2.1, with nearly identical parameterizations (CL1 =1.80 and α=0.0007 in this study; CL1 = 1.74 and α=0.0007 in MEGAN 2.1).

The choice of which temperature-dependent flux response equation to apply varies among different compounds and different studies, as illustrated in Table 3. In our study, both the light-dependent fraction (LDF) and the light-independent fraction (LIDF) equations for temperature response performed better than the PAR response curve. In addition, the PAR response curve goes to zero as PAR goes to zero, although it appears that emissions of ethene, propene and butene still occurred at nighttime when PAR equaled zero. We therefore utilized a combination of the temperature-based equations, scaled by the LDF reported in the MEGAN 2.1 model (last column in Table 3), to extrapolate flux results to the remainder of the season for which flux measurements were not determined. Between 1 May and 31 October 2014, the extrapolated seasonal flux yielded an average of 61.5, 51.7, 24.3 and 18.0 µg m-2 h-1 for ethene, propene, butene and isoprene, respectively. For the light alkenes, this represents a 40–80 % higher emission rate than that observed over the same season length at Harvard Forest (Goldstein et al., 1996). This is slightly larger than the simple linear extrapolation described in Sect. 5.3 above.

In MEGAN 2.1, ethene is classified as a “stress VOC” owing to its known biochemical production during times of abiotic and biological stress (Abeles et al., 2012), while propene and butene are classified as “other VOCs”. In this study, propene and butene fluxes highly correlate with ethene fluxes and show a very similar light and temperature response. Hence, our results suggest that propene and butene can be categorized together with ethene, and their temperature-dependent emissions should have similar LDF values. In MEGAN 2.1, global butene emissions are only 30 % of ethene and 50 % of propene, which is similar to the ratios found here (30 and 40 %, respectively). Modifying the light and temperature parameterizations for light alkenes in the vegetation emissions model will lead to a corresponding increase in estimated global emissions for these compounds. This would generally support the conclusion of Goldstein et al. (1996) that “terrestrial biogenic emissions could provide a significant global source for two important reactive olefins, propene and 1-butene”, with the caveats that the specific butene isomer remains in question and that other terrestrial ecosystems need to be surveyed.