Spatiotemporal variability of soil respiration in a seasonal tropical forest

Abstract We monitored soil CO 2 effluxes for over 3 years in a seasonally wet tropical forest in Central Panama using automated and manual measurements from 2013 to 2016. The measurements displayed a high degree of spatial and temporal variability. Temporal variability could be largely explained by surface soil water dynamics over a broad range of temporal scales. Soil moisture was responsible for seasonal cycles, diurnal cycles, intraseasonal variability such as rain‐induced pulses following dry spells, as well as suppression during near saturated conditions, and ultimately, interannual variability. Spatial variability, which remains largely unexplained, revealed an emergent role of forest structure in conjunction with physical drivers such as soil temperature and topography. Mean annual soil CO 2 effluxes (±SE) amounted to 1,613 (±59) gC m−2 year−1 with an increasing trend in phase with an El Niño/Southern Oscillation (ENSO) cycle which culminated with the strong 2015–2016 event. We attribute this trend to a relatively mild wet season during which soil saturated conditions were less persistent.


year
Lowland terra firme tropical rainforest 2,649 Edisol eddy covariance system IRGA Sotta (1998)  Climatic factors such as precipitation and radiation largely drive temporal variability in soil respiration, influencing soil moisture, temperature, and many biotic processes such as root, soil microbes, and litterfall dynamics. Wet and moist tropical climates are characterized by intense and frequent rainfall events, with or without a seasonal cycle. Soil moisture is probably the most important abiotic factor influencing soil respiration within tropical forests (e.g., Li, Xu, & Zou, 2006;Sotta et al., 2006). Soil CO 2 efflux can be suppressed in both low and high soil water content (Davidson, Belk, & Boone, 1998;Linn & Doran, 1984). High water content creates a barrier to gas diffusion at the soil-atmosphere interface, limiting the escape of CO 2 and supply of oxygen (Liptzin, Silver, & Detto, 2010), thereby reducing both, production and diffusion of CO 2 (Davidson, Samanta, Caramori, & Savage, 2012;Fang & Moncrieff, 1999). At low soil moisture conditions, decomposition is limited by soluble carbon availability (Davidson et al., 2012;Linn & Doran, 1984).
Several nonlinear relationships have been proposed to link soil respiration rate and soil water content (Cook & Orchard, 2008;Davidson, Verchot, Cattanio, Ackerman, & Carvalho, 2000), indicating optimal conditions for microbial decomposition and root respiration at intermediate moisture conditions. However, these relationships remain empirical, and it is unknown how they vary with soil, climate, and forest type. In addition, these relationships can be altered during rain-induced pulses, which can be caused by large amounts of watersoluble carbon leaching from the litter or dead microbes, accumulated during dry periods, known as the "Birch effect." Although these pulses are often observed in dry ecosystems (Ma, Baldocchi, Hatala, Detto, & Yuste, 2012), are less documented in tropical forests (Cleveland & Townsend, 2006). Because this nonmonotonic and nonlinear response, the effect of changing in rainfall variability is difficult to predict in a particular forest without available observations.
In contrast, soil temperature fluctuations in the tropics are small, especially in areas covered by dense vegetation, where little radiation reaches forest floors. Although temperature is a direct factor affecting root and microbes metabolic rates (Kuzyakov & Gavrichkova, 2010;Lükewille & Wright, 1997) and is responsible for the temporal variation in soil respiration, primarily in temperate (e.g., Hanson, Wullschleger, Bohlman, & Todd, 1993;Vargas, Detto, Baldocchi, & Allen, 2010) and boreal ecosystems (Shibistova et al., 2002), it might play a secondary role in tropical forests (Davidson et al., 2000). For example, results from a rain exclusion plots suggest that the positive effect of temperature on soil CO 2 efflux is still constrained by soil moisture availability (Wood, Detto, & Silver, 2013).
Relative to temporal variability, spatial variability is less understood, and it can be driven by heterogeneity in below ground physical, chemical, and biological soil properties, landforms, and vegetation cover (Hanson et al., 1993;Xu & Qi, 2001). Spatial variability is also known to be very large at small scales (Epron, Bosc, Bonal, & Freycon, 2006;Kursar, 1989). By shielding intercepting rainfall and determining root distribution, above ground forest structure can create microheterogeneity in the physical and biotic drivers of soil respiration (Raich & Tufekcioglu, 2000). Species composition may also play an important role, as plant species differ in the production and quality of detritus (Raich & Tufekcioglu, 2000), root system, and associations with microorganisms such as fungi and bacterial communities (Barberán et al., 2015). Plant-soil feedbacks have an important role in the ecosystem nutrient cycling and soil carbon exchange through productivity and carbon input into the soil (Balogh et al., 2011;Bardgett, Freeman, & Ostle, 2008;Sayer et al., 2011). For example, diurnal fluctuations in soil CO 2 effluxes may also be caused by translocation of photosynthates from leaves to roots Kuzyakov & Gavrichkova, 2010).
At larger scales, topographic features (slopes, plateau, and valley) influence hydrological processes and determine heterogeneity in water availability, soil texture, and nutrients (Silver, Scatena, Johnson, Siccama, & Sanchez, 1994;Weintraub et al., 2015). However, the effects on soil respiration are still unclear and literature reports mixed results. Sotta et al. (2006) and Hanson et al. (1993) found no differences between landforms in the Eastern Amazonia and South East US, respectively, while other studies found strong relationship between soil respiration and topographic position, decreasing from hills to bottomlands (Chambers et al., 2004;Epron et al., 2006) or increasing from ridge to lower slopes (Takahashi et al., 2011).
The spatiotemporal variability complicates ground-based monitoring of soil respiration because of the inherent trade-off between temporal and spatial sampling resolution. Recent advances in automated systems have greatly improved our ability to monitor temporal variability up to a half hourly resolution, which is comparable to the scale of variation of many climatic and hydrological drivers. However, systems such as dynamic chambers rely on a centralized gas sampling design, which limits their applications in spatially heterogeneous environments. Manual measurements are more adaptable to spatially stratified sampling designs, but sampling frequency is often insufficient to resolve all scales of variation, and there are other logistic problems limiting sampling at night or during, and immediately after rain events. A combination of manual and automated measurements could provide sufficient information to accurately quantify the soil CO 2 effluxes. Unfortunately, tropical studies that integrate this approach on a sufficiently large temporal horizon are scant (see Table 1).
The objectives of this study were to quantify soil respiration in a lowland seasonally wet tropical forest and analyze their spatial and temporal variability. We used both, manual and automated systems, collecting more than three years of measurements, which include a strong El Niño event. In particular, we investigated the effects of soil moisture, soil temperature, topography, and forest structure. Finally, in order to compute integrated seasonal and annual budgets, we assimilated the measurements in a statistical model at daily scale using Artificial Neural Network.

| Site description
The study site is located on Barro Colorado Island (BCI), Panama (9°9′N, 79°50′ W), a 15 km 2 island in the middle of the Lake Gatun.
The forest is tropical moist with a distinct dry season between January and April. Mean annual temperature is 27°C, with minimal seasonal variation; mean annual rainfall is 2642 (±566) mm.

| Automated dynamic chamber system
An automated CO 2 chamber system (LI8100, Licor Bioscience) was installed from June 2013 until January 2015 and from April 2016 to end of August 2016. The system included four automated Dynamic and automatic closed chamber, a multiplexer and a close-path infrared gas analyzer. Each chamber operated every 20 min and CO 2 concentration measurements were taken every second for 2 min after the chamber was closed. A 30-sec prepurge and 45-s postpurge were introduced to allow flushing the system between each measurement.
The soil collars had an outside diameter of 11.4 cm and were installed 5 cm into the soil 1 month before measurements commenced.
The collars were located at the vertices of a 20 × 20 m square centered at the microclimatic tower, where power grid was available.
Data collection was interrupted few times due to power losses (about 33 days total). On May 2014, the pump failed and the sensor was sent for repair and factory recalibration; it was reinstalled on July 2014 until January 2015, when it was designated to another experiment. A new, factory calibrated, identical system was installed in April 2016. The automated system had been operative for 617 days, during which it was visited regularly every 1-2 weeks to check the operative status and clean the chambers from litter or other debris that could prevent perfect closure.

| Portable static chamber system
In

| Soil temperature and soil moisture
Continuous soil temperatures were taken in proximity of the auto-

| Forest structure
We recorded the diameter of all the trees at breast height, within 5 m distance from each collar to calculate basal area. We calculated the gap fraction (indirect site fraction) of the canopy above each ring taking hemispherical photographs with a digital camera (Canon EOS 6D, Canon Inc. Japan) provided with a fisheye lens (Sigma 8 mm f/3.5 EX DG Circular Fisheye Lens, Sigma Corporation of America).

| Statistical analysis
Correlation analysis was used to determinate the relationship between measurements taken by automated and manual chambers, the spatial relationship between temperature and soil CO 2 effluxes and the relationship between the magnitude of rain-induced pulses and soil moisture fluctuations. From the automated time series, ten pulses were selected during the dry seasons and their magnitudes estimated as the percentage increase between the values of the effluxes immediately before the rain event and at the attained maximum.
A quadratic relationship between spatially averaged soil CO 2 effluxes and soil moisture was fitted using a robust nonlinear leastsquares method implemented in the function fit (MATLAB 2014a). The relationships were fitted independently for automated and manual samples.
Coefficient of variation (CV), semivariogram (γ), and autocorrelation functions (acf) were used to quantify the temporal and spatial variability of the effluxes during the entire period and during the dry and wet periods separately (for CV and γ only). The autocorrelation as function of time interval τ was defined as: where E[ ⋅ ] denotes expectation and ⟨f⟩ t spatial average for all mea- Diurnal pattern of soil effluxes was computed by averaging continuous measurements as function of time of the day during the 2016 dry season. To compare soil temperature, soil moisture, and soil CO 2 efflux, the diurnal patterns were normalized between 0 and 1.

| Artificial Neural Networks (ANN)
Because manual and automated measurements were acquired at different temporal resolution and irregular frequency, averaging across space and time is problematic. ANN is a statistical technique commonly used to gap fill biological fluxes (gross primary productivity, ecosystem respiration, and evapotranspiration) obtained from eddy covariance measurements (Papale & Valentini, 2003) in order to produce meaningful time integrated budgets. Here, we used ANN with the same scope.
All data were assimilated at daily time step into the ANN which uses meteorological inputs (soil moisture, solar radiation, temperature, and pressure) to predict soil CO 2 effluxes at each location. Once all the series have been put on the same time frame, they were easily averaged.
To train the network, the dataset was randomly divided into a training set (70%), a validation set (15%) and a test set (15%). A twolayer feed-forward network with 10 sigmoid hidden neurons and linear output neurons was trained using the Levenberg-Marquardt algorithm until the mean square error (MSE) of the validation set stop improving (Hagan & Menhaj, 1994).Performance, in term of MSE, was evaluated using the test set at the end of the training. This procedure was repeated 100 times for each location to produce 100 estimates of daily soil CO 2 effluxes. Training multiple times generates different results due to different initial conditions and random sampling of the three sets. Ensemble at any locations i and day j was obtained as weighted average from the 100 ANN predictions using the inverse of MSE of the test set as weights according to: The ANN was implemented using the Neural Network Toolbox in MATLAB 2014a.

| RESULTS
Time series of soil respiration measurements (Figure 1a,b) showed a clear seasonal pattern in phase with soil moisture (Figure 1c). There was a large spatial variation in the weakly manual measurements and in the four automated measurements. Automated measurements also displayed large temporal variability at finer time scales. Soil CO 2 effluxes ranged from as low as 0.06 to a maximum of 14.07 μmol CO 2 m −2 s −1 and 0.25-14.57 μmol CO 2 m −2 s −1 , for the manual and automated system, respectively.
In addition to seasonal cycle, surface soil moisture exhibited highfrequency fluctuations in correspondence to rain events (Figure 1c).
Lower values of soil CO 2 efflux were found when the soil was either dry or completely wet, immediately following heavy rain events in the wet seasons. In contrast, even moderate rain events after long dry spells generated CO 2 pulses of variable magnitude.
Soil temperature exhibited less variation, both diurnal and seasonal, ranging from 22 to 30 degrees Celsius across the entire record.
Highest values were reached during the 2016 dry season in correspondence to a strong El Niño event (Figure 1d).
Automated and manual measurements were in good agreement when compared on a daily scale. Figure 2 shows a scatterplot representing the average of the 32 manual measurements for each census against the average of the four chambers during the time of the day correspondent to the duration of the field campaign. Although there was a good correlation (R 2 = 0.78, p < 1 × 10 −10 ), manual measurements were higher for low soil CO 2 effluxes and lower for high soil CO 2 effluxes.
Temporal and spatial variabilities of soil respiration were both high with coefficient of variations (CVs) ranging from 0.27 to 0.46, and 0.17 to 0.47 for temporal and spatial CV, respectively (Figure 3a,b). During the dry season, CVs were higher than during the wet season, indicating strong relative variability during low flux periods. Temporal autocorrelation function showed a long-term correlation (>15 months) with annual periodicity (Figure 4c). In contrast no spatial structure was detected in the semivariogram (Figure 3d), suggesting that from a 20 m distance, measurements can be considered statistically independent.
The automated and manual measurements showed a consistent quadratic relationship between soil moisture and soil CO 2 effluxes ( Figure 4). For both automated and manual measurements, the peak of soil CO 2 effluxes was at ~0.45 cm 3 /cm 3 . (1) Pulses of soil respiration during dry season were strongly correlated with the magnitude of the soil moisture fluctuations, with the soil CO 2 efflux doubling for a 20% increase in soil moisture ( Figure 5).
Automated measurements allowed us to study diurnal cycles in soil respiration, which were consistently detected during the dry seasons with an average amplitude of ~0.6 μmol m −2 s −1 . This diurnal variation was in phase with soil water content, as illustrated in Figure 6, while temperature was lagging for about 5 hr.
Spatial soil temperature variations were minimal, spanning approximately one degree Celsius in each of the four manual censuses ( Figure 7). Not surprisingly, these small spatial temperature differences did not explain much variation in the soil CO 2 effluxes. The correlation between manual CO 2 effluxes and soil temperature was significant in one census only (R 2 = 0.27, p = .002, n = 32), marginally significant in one census (R 2 = 0.10, p = .081, n = 32), and not significant in the other two censuses (R 2 = 0.05, p = .219, n = 32; R 2 = 0, p = .861, n = 32). When the significant temperature census was included in the multiway ANOVA, the overall effect of temperature was significant, but mostly in conjunction with gap fraction (Table 2).
Although within topographic features spatial variation was large (coefficient of variation was 20% and 22% along slope and plateau, respectively), the average soil CO 2 efflux along the slope was slightly bigger than on the plateau (4.38 ± 0.07 and 4.22 ± 0.05 μmol m −2 s −1 , respectively), and the difference was significant (p < 1 × 10 −6 ).
However, when topography was included in a multivariate analysis with other variables, only the interaction terms were significant, in particular with basal area and gap fraction ( Table 2).
The multiway ANOVA (Table 2) also revealed significant effects of basal area, gap fraction, temperature and several interaction terms that explain spatial variation in the manual measurements after removing

| DISCUSSION
Our long-term observations displayed large spatial and temporal variability, the latter over a broad range of scales, with similar degree of variation among space and time. Despite lower emissions, the dry season exhibited higher relative variability because dry soil responded dynamically to rainfall events as commonly observed in water-limited ecosystems (Ma et al., 2012), where they can be responsible for more than 10% of the carbon losses over a year (Jarvis et al., 2007). These temporal processes reflected also on the spatial variability, which was relatively higher during dry season most likely because each location responded differently to these rainy events. Automated measurements showed that soil CO 2 effluxes were greatly reduced immediately following an intense rain event. This effect can last several hours, with a complete recovery of prestorm conditions after approximately 2 days. This means that during the high of the wet season soil respiration is always limited because storm frequency averages one every 2-3 days.
Previous studies highlighted the control of soil moisture on soil respiration (e.g., Davidson et al., 2000;Schwendenmann, Veldkamp, Brenes, O'Brien, & Mackensen, 2003;Sotta et al., 2006;Xu & Qi, 2001). In our study, soil moisture confirmed to be the primary driver of temporal variability in this tropical forest, with temperature playing a secondary role. However, although the effect of soil moisture was F I G U R E 4 Quadratic relationships between soil moisture and soil CO 2 effluxes obtained from automated and manual measurements. Blue dots represent daily average flux measurements with daily averaged soil water content between 0 and 15 cm. Fitted equation, R 2 and root mean square error are also reported F I G U R E 3 Spatiotemporal variation in the manual measurements shows a large degree of variability in both, temporal and spatial axes, strong temporal autocorrelation and lack of spatial structure. Boxplot of coefficient of temporal variation (CV) of soil CO 2 effluxes among locations during all periods, wet (swc > 0.35) and dry (swc < 0.35) conditions (a). Boxplot of coefficient of spatial variation of temporally averaged soil CO 2 effluxes during all period, wet and dry conditions (b). Autocorrelation function (c). Semivariogram during wet and dry conditions (d) well described by a quadratic relationship, large scatter still remains at intermediate moisture conditions, evident from the automated measurements which have higher temporal resolution and a better representation of the rain pulses. The soil CO 2 effluxes peak, at 0.45 cm 3 / cm 3 , was consistent with other studies that have used a parabolic function to describe the relationship between soil moisture and soil respiration (e.g., Chambers et al., 2004;Sotta et al., 2004). For example, Wood et al. (2013) found higher soil CO 2 efflux at 0.375 m 3 /m 3 , and Schwendenmann et al. (2003) found that the peak ranged from 0.35 to 0.50 m 3 /m 3 depending upon the type of soil.
In tropical forests, where temperature variations, both temporal and spatial, are generally small, soil temperature does not significantly influence soil respiration. Actually, the time series displayed an apparent negative correlation (Figure 1). This confounding effect, analogous to what was found by Davidson et al. (1998), is caused by soil moisture to mask the effect of temperature, because the covariation between cold/hot and wet/dry conditions.
In fact, the largest fluctuations in soil temperature were observed during the dry season. In this period, clear sky conditions and canopy opening left by the dry deciduous species, allow larger amounts of solar radiation to reach the forest floor, generating deep diurnal cycles in soil heat. It is in concomitance with these fluctuations that flux measurements showed a detectable diurnal cycle (Figure 7). An indepth analysis of the lags between temperature and soil CO 2 effluxes demonstrated that these cycles were not consistent with a temperature dependent process, because effluxes peaked at 8:00 a.m. while temperature peaked at 13:30 p.m. Considering that gas soil diffusivity is relatively high in dry conditions, it seems unlikely that temperature produced a lagged response of more than 18 hr.
Again, this diurnal pattern is almost in phase with soil moisture, which in the dry season also exhibited detectable diurnal cycles generated by root water uptake. This suggests two possible explanations: (1) Sugars are synthetized during the day and translocated to roots. These sugars are respired by roots or released as exudates in the rhizosphere, and used by soil microorganisms such as mycorrhizae (Kuzyakov & Gavrichkova, 2010). The time of this coupling, which is determined by phloem transport dynamics (Mencuccini & Hölttä, 2010), generates a lagged response between photosynthesis (or transpiration) and soil respiration.
(2) Part of the carbon respired by roots can be transported F I G U R E 7 Relationship between manual CO 2 effluxes and soil temperature measured in the proximity of the collars during four filed campaigns. Date of the census, R 2 and p-value are indicated in the legend. Linear regressions are indicated by dashed lines for reference F I G U R E 6 Mean diurnal variation of soil CO 2 effluxes is out of phase with soil temperature and soil moisture during 2016 dry season. All variables are normalized between 0 and 1 for comparison F I G U R E 5 Rain-induced pulses of soil CO 2 effluxes (F) explained by relative changes in soil moisture (swc). Each point represents a pulse with magnitude expressed as the relative difference of the flux measured just before the rain event (denoted as 0), and the maximum flux of the pulse (denoted as max). Least-squares line, R 2 and p-value are shown for reference aboveground through the xylem rather than being diffused through the soil (Aubrey & Teskey, 2009), reducing measured soil CO 2 effluxes during periods of high transpiration. Although this pattern was extremely interesting, the magnitude of the fluctuations was relatively small, and the phenomenon was present only during dry periods, with marginal contribution for annual budgets. For this reason, it was neglected for upscaling manual measurements.
We did not find spatial correlation in our sampling design, which might indicate a lack of large-scale structure, confirmed by the weak effect of topography. This simplified the analyses because the samples could be considered statistically independent. However, it does not shed light on the spatial scales of variation and the possible mechanisms that generated such variation, suggesting a presence of an unresolved fine scale variability, which might be much smaller that our sparse (20 m × 20 m) sampling design, as previously observed by Kursar (1989).
Although spatial variability was less characterized than temporal variability, our study suggests that forest structure has important direct and indirect influences. For example, the effect of temperature was small but significant, as indicated by the multiway ANOVA, but it was the interactions between temperature and gap fraction that explained most of the variance. This suggests that spatial temperature effects were mediated by canopy structure, which may prove important for upscaling soil CO 2 effluxes using remote sensing products such as LiDAR. Similarly, higher soil CO 2 effluxes were observed from slope compared to plateau, a topographic effect generally attributed to hydrological processes related to water transport and drainage (Epron et al., 2006;Ngo et al., 2013;Zhou et al., 2016).
However, the multiway ANOVA showed that the topographic effect was mostly significant when interacting with basal area. This might be consistent with the observation that within these forests, mean canopy height, a predictor of above ground biomass, is strongly correlated with hydrological terrain attributes, with taller forests in the proximity of the drainage network (Detto, Muller-Landau, Mascaro, & Asner, 2013).
Other studies investigated the spatial correlation between the soil CO 2 efflux and tree proximity, but results were not consistent and the mechanisms unclear. For example, Sotta et al. (2004)  T A B L E 2 Multiway analysis of variance (ANOVA) for testing the effects of multiple factors and their interactions on the mean of the residual soil CO 2 effluxes (after removing the temporal dependence on soil moisture with a quadratic model) traits. Conversely, Shibistova et al. (2002) found that rates from relatively open areas were about half of those observed below or around trees, a result that was attributed mostly to root density. Vargas and Allen (2008) found that gaps in the canopy could explain fluctuations in soil volumetric water content, causing soils to become wetter or drier faster during or after a rain event, respectively. Strong correlation with forest stand structure at ecosystem scale was also found by Katayama et al. (2009) andShi, Gao, Cai, andJin (2016).
The study period was characterized by an ENSO cycle that culminated with a strong El Niño in 2015-2016. During El Niño, this region of the tropics experiences relatively dry and warm conditions.
Because the relationship between soil respiration and soil moisture is not monotonic, dry conditions can enhance or reduce annual carbon losses. Our study indicated an increase of annual CO 2 fluxes during El Niño, mainly driven by a reduction in the frequency of saturated water soil conditions during the wet season. Considering that the frequency and intensity of ENSO are predicted to increase, these results will be useful to understand the impact of climate change on tropical forest carbon cycle.
This study showed how an innovative combination of long-term automated and manual measurements could help to better quantify soil CO 2 effluxes in a temporally and spatially variable environment and provide an accurate estimation of landscape soil carbon losses at annual scale. Our soil CO 2 effluxes fall within the range of the tropical studies presented in Table 1. The annual budget is not statistically different from the mean of the distribution computed from the averaged effluxes reported in these studies (p-value .1608, t-test with 48 dof). However, the length of the study period and the method used for measuring the soil CO 2 effluxes varies among studies, making meaningful comparison difficult, highlighting the importance of long-term observations, and standardized method.
Statistical models as the ANN implemented in this study are useful tools to integrate the measurements in space and time, but mechanistic models will be necessary to project soil carbon losses from tropical forests under different climate change scenarios. For this reason, it is essential that future research aims to understand some of the unresolved variability of soil respiration, especially the spatial variability. Important mechanisms that should receive better attention in tropical forests are ecological drivers, as the carbon translocation from above ground to roots and the interactions with soil microorganisms.