Vegetation influence and environmental controls on greenhouse gas fluxes from a drained thermokarst lake in the western Canadian Arctic

Abstract

Abstract. Thermokarst features are widespread in ice-rich regions of the circumpolar Arctic. The rate of thermokarst lake formation and drainage is anticipated to accelerate as the climate warms. However, it is uncertain how these dynamic features impact the terrestrial Arctic carbon cycle. Methane (CH4) and carbon dioxide (CO2) fluxes were measured during peak growing season using eddy covariance and chambers at Illisarvik, a 0.16 km2 thermokarst lake basin that was experimentally drained in 1978 on Richards Island, Northwest Territories, Canada. Vegetation in the basin differs markedly from the surrounding dwarf-shrub tundra and included patches of tall shrubs, grasses, and sedges with some bare ground and a small pond in the centre. During the peak growing season, temperature and wind conditions were highly variable, and soil water content decreased steadily. Basin-scaled net ecosystem CO2 exchange (NEE) measured by eddy covariance was −1.5 [CI95 %±0.2] g C−CO2 m-2d-1; NEE followed a marked diurnal pattern with no day-to-day trend during the study period. Variations in half-hourly NEE were primarily controlled by photosynthetic photon flux density and influenced by vapour pressure deficit, volumetric water content, and the presence of shrubs within the flux tower footprint, which varied with wind direction. Net methane exchange (NME) was low (8.7 [CI95 %±0.4] mgCH4m-2d-1) and had little impact on the growing season carbon balance of the basin. NME displayed high spatial variability, and sedge areas in the basin were the strongest source of CH4 while upland areas outside the basin were a net sink. Soil moisture and temperature were the main environmental factors influencing NME. Presently, Illisarvik is a carbon sink during the peak growing season. However, these results suggest that rates of growing season CO2 and CH4 exchange rates may change as the basin's vegetation community continues to evolve.

Controls on carbon dioxide and methane fluxes from a low-center polygonal peatland in the Mackenzie River Delta, Northwest Territories

“…The NN analysis identified PPFD as the primary control over GPP, and highlighted VPD as the main limiting factor of GPP, which was expected (Aubinet et al, 2012). This relationship has been identified by NN analysis in other studies (Mofat et al, 2010;Skeeter et al, 2020) and has been noted at other wet tundra sites across the Arctic (Kwon et al, 2006;Fox et al, 2008). The model selected polygon center temperatures at 5cm as the dominant driver, which makes sense given that polygon centers cover the majority of the land area (66%) at the site.…”

“…Neural Networks (NN) were used to: identify relevant environmental controls over NEE and NME, gap fill the time series, and partition NEE into its component fluxes ER and GPP. Various studies have applied NNs to flux data to identify and analyze flux drivers (Moffat et al, 2010;Briegel et al, 2020;Skeeter et al, 2020), investigate the influence of spatial heterogeneity (Morin et al, 2015;Skeeter et al, 2020), and to gap-fill time series of NEE and NME Valentini, 2003, Dengel et al, 2013;Knox et al, 2015;Skeeter et al, 2020). The steps for identifying relevant controls, gap filling, and flux partitioning are discussed in section 2.5.…”

“…A NN trained on CO 2 fluxes will just estimate NEE rather than the component fluxes ER and GPP. To compensate for this, we estimated daytime ER with the pruned NN by using an artificial "dark" input set (Skeeter et al, 2020). The inputs Daytime and PPFD were both set to zero and any R n > 0 were also set to zero.…”

“…Jarvis and Stauch, 2003), we have no way of knowing what it is actually doing. Plotting the model output response to specific drivers is beneficial (Moffat et al, 2010;Skeeter et al, 2020), but when working in multidimensional feature space there is no way to assess all possible combinations of drivers. To our knowledge, this is the first study to apply the weights method (Gevrey et al, 2003) to a NN analysis of EC data.…”

“…Instead, we used the Daytime variable to give the model the ability to differentiate between night and day without splitting our training data. Skeeter et al (2020) approximated ER using a NN by setting radiative inputs to "nighttime" values, but they did not use a Daytime variable, and their NN underperformed relative to a Q 10 curve. We found adding Daytime drastically increased the model's ability to resolve ER which we support with comparisons to the Q 10 and logistic temperature response curves.…”