Respiratory loss during late-growing season determines the net carbon dioxide sink in northern permafrost regions

Northern high latitude (NHL, >50°N) ecosystems have shown increased terrestrial net CO2 uptake and a larger atmospheric CO2 seasonal amplitude over recent decades, but the balance between photosynthetic uptake and respiratory losses remains uncertain due to their asynchronous and seasonally distinct climate responses. Seasonal compensation—where enhanced growing-season uptake is offset by increased fall/winter respiration—complicates detection of long-term climate–carbon feedbacks and may differ among biomes along vegetation and permafrost gradients. This study asks three questions: (1) What are the trends in seasonal and annual net CO2 uptake across the NHL and how do they relate to climate, vegetation (tree cover), and permafrost extent? (2) What mechanisms underlie differences in net CO2 uptake trends—greater temperature sensitivity of productivity (H1) versus seasonal respiratory compensation (H2)? (3) How well do current Dynamic Global Vegetation Models (DGVMs) reproduce NHL seasonal CO2 dynamics? Addressing these questions is critical for assessing whether NHL permafrost regions are becoming net CO2 sources or sinks under warming and for improving projections of climate–carbon feedbacks.

Prior work documents enhanced CO2 seasonal exchange and increasing land carbon uptake in northern regions, often attributed to amplified plant productivity (greening) and earlier spring onset, but also highlights growing winter and fall respiration that can offset gains. Observational and experimental studies have raised concerns that warming-driven permafrost thaw and accelerated soil decomposition could shift Arctic systems from sinks to sources. Studies also report expanding shrub/tree cover and changing moisture regimes, with region-specific drivers such as Eurasian warming, wetting, and high light-use efficiency in larch forests, as well as disturbance (e.g., fire) patterns differing between Eurasia and North America. Model assessments frequently attribute changes to productivity and tend to underestimate the observed atmospheric seasonal amplitude, indicating potential underrepresentation of late-season respiration processes.

Study domain: Northern High Latitudes (NHL, >50°N, excluding Greenland), restricted to natural vegetation (MODIS MCD12Q1 IGBP classes), excluding croplands, cropland/natural mosaics, and urban areas. Regions characterized by percent tree cover (%TC) and percent permafrost extent (%P); permafrost zones categorized as continuous (>90%), discontinuous (10–90%), and non-permafrost (<10%). Data and periods: (1) Atmospheric CO2 inversions (ACIs): ensemble mean of six monthly gridded products at 1° resolution (CAMS 1979–2019; Jena CarboScope s76_v4.2 1976–2017; Jena s85_v4.2 1985–2017; JAMSTEC 1996–2017; CarbonTracker CT2019B 2000–2019; CarbonTracker Europe CTE2020 2000–2019). Analyses use 1980–2017 period where available, aggregated to seasonal (EGS: May–Aug; LGS: Sep–Oct; Winter: Nov–Apr) and annual sums. (2) Eddy Covariance (EC) network: 48 sites, ≥3 consecutive years each, totaling 426 site-years (1990–2017). (3) Productivity proxies: GIMMS NDVI (bimonthly, 1/12°, 1982–2015) and LUE-based GPP (AVHRR-driven MOD17 approach; monthly, 0.5°, 1982–2015; meteorology from NCEP/NCAR, MERRA-2, ECMWF; ensemble mean). TER (respiration) estimated residually as TER = GPP − NEE_ACI (1982–2015). (4) Permafrost maps: ESA CCI Permafrost (1 km; thermal-model derived ground temperature at 2 m depth) to compute %P (averaged 1997–2017) and IPA categorical zones for cross-check. (5) Soil moisture: ESA CCI SM v4.5 (daily, 0.25°), aggregated to monthly/seasonal/annual means (1980–2017). (6) Climate: CRU TS v4.02 air temperature and photosynthetically active radiation (PAR) aggregated seasonally/annually. Analytical design: - Trends: For each %TC (5% bins), %P (5% bins), and temperature (1 °C bins), compute linear trends of net CO2 uptake (NEE_ACI; positive denotes net C uptake) for EGS, LGS, winter, and annual periods (1980–2017). Regional aggregations contrasted short-vegetated permafrost (TC <50% in ConP/DisconP) versus tree-dominated non-permafrost (TC >50% in NoP). - Hypothesis testing: H1 tested via correlations between trends in productivity (NDVI, LUE GPP) and net CO2 uptake along gradients, and pixelwise time-series correlations. H2 tested by decomposing seasonal contributions (EGS uptake vs LGS release) and quantifying seasonal compensation. - Structural Equation Modeling (SEM): Separate SEMs for EGS and LGS relating NEE, GPP, TER to drivers (AirT, PAR, SM, %TC, %P, and preseason productivity PreGPP for LGS). Models fit using lavaan in R; goodness-of-fit assessed by χ2, CFI, RMSEA; standardized coefficients interpreted for direct/indirect effects. - DGVM comparison: TRENDY ensemble (n=10; NBP) analyzed analogously for seasonal and annual trends and gradients; also compared seasonal temperature sensitivities. - Robustness checks: Individual ACI trends; randomized start years/lengths (≥10 years); EC vs ACI site-level comparisons; Generalized Linear Mixed-effects Model (GLMM) incorporating inversion spread, flux partitioning, and time-dependent spread; spatial agreement maps (majority across ACIs); pre/post-2000 sensitivity to the number/length of ACIs. - Temperature sensitivity analyses: Regress seasonal/annual net CO2 uptake against spring (May–Jun) and same-season temperatures to quantify direct and legacy effects across permafrost and non-permafrost regions for EC, ACI, and TRENDY datasets.

- NHL contribution and consistency: The ACI ensemble correlates strongly with independent global land CO2 exchange (r = 0.78, p < 0.001). NHL accounts for ~28.7% of the mean global land CO2 sink (0.67 ± 0.28 Pg C yr−1), ~17% of the global trend, and ~12.5% of interannual variability. - Spatial trends: From 1980–2017, 50.3% of the NHL shows significant increases in annual net CO2 uptake (mainly tundra/permafrost), whereas 4.6% shows significant decreases (mainly forest/non-permafrost). The trend in net CO2 uptake is most strongly associated with tree cover (R2 = 0.90, p < 0.001), followed by mean annual temperature (R2 = 0.77, p < 0.001) and permafrost extent (R2 = 0.27, p = 0.033). - Permafrost vs non-permafrost: Net CO2 uptake increased significantly faster in permafrost (short-vegetated) regions (0.58 ± 0.086 g C m−2 yr−2) than in non-permafrost (tree-dominated) regions (0.13 ± 0.11 g C m−2 yr−2). Permafrost regions shifted from near-neutral (3.16 ± 6.51 g C m−2 yr−1, 1980–2000) to a sink after 2000 (15.01 ± 6.14 g C m−2 yr−1). - Mechanism: Productivity increased broadly with warming and tree cover, and productivity correlates positively with net CO2 uptake; however, trends in productivity did not explain the decreasing trend in net uptake with increasing tree cover. Instead, seasonal compensation is key: similar EGS (May–Aug) increases in uptake across tree cover are increasingly offset by stronger LGS (Sep–Oct) respiration with greater tree cover. Aggregated LGS emissions trends are −1.41 ± 0.15 g C m−2 yr−2 (permafrost) versus −2.37 ± 0.19 g C m−2 yr−2 (non-permafrost). - Temperature sensitivity: Annual net CO2 uptake shows a positive sensitivity to temperature in permafrost regions (101 ± 32 Tg C yr−1 K−1, p < 0.001) but a non-significant negative sensitivity in non-permafrost regions (−14 ± 18 Tg C yr−1 K−1, p = 0.46), reflecting stronger LGS respiratory losses in forests. - SEM insights: In EGS, net uptake is primarily controlled by productivity driven by spring air temperature and tree cover. In LGS, observationally constrained data indicate respiration dominates net CO2 balance; LGS respiration is strongly influenced by EGS productivity (labile carbon carryover) and temperature, and is weaker where permafrost extent is higher. - Model evaluation: TRENDY DGVMs simulate increasing annual uptake but fail to reproduce the observed decrease of net uptake trend with increasing tree cover, largely due to underestimation of increased LGS respiration and seasonal compensation. DGVMs also show unrealistic (non-negative) sensitivities of LGS NBP to spring warming compared with observations.

The findings address the central question of why net CO2 uptake increases more rapidly in permafrost tundra than in forested boreal regions. While warming enhances productivity across the NHL, the balance is seasonally modulated: in forested, non-permafrost regions, stronger late-growing-season respiration erodes annual net gains, leading to slower increases in the annual sink and reduced temperature sensitivity. In contrast, short-vegetated permafrost regions experience smaller increases in LGS respiratory losses, allowing EGS gains to drive a faster-growing sink. These seasonal compensation dynamics explain observed differences along vegetation–permafrost gradients and reconcile debates regarding whether increased atmospheric seasonal amplitude reflects productivity versus respiration changes: both contribute, but LGS respiration in forests exerts a strong compensatory control. The inability of DGVMs to capture enhanced LGS respiration undermines confidence in projections that rely on modelled increases in NHL carbon sinks under warming, indicating a need to refine representation of respiration, carryover effects of productivity, and interactions with moisture and permafrost state.

There is no current evidence that NHL permafrost regions are net CO2 sources; instead, they have transitioned toward a strengthening CO2 sink since the 1980s, with increases outpacing those in tree-dominated non-permafrost regions. The stronger sink trend in permafrost tundra arises from enhanced early-growing-season productivity combined with comparatively smaller late-season respiratory offsets. Conversely, in boreal forests, increasing late-season respiration drives larger seasonal compensation, slowing annual sink growth and amplifying the atmospheric CO2 seasonal cycle. Current DGVMs underestimate late-season respiration and seasonal compensation, limiting their ability to reproduce observed gradients and trends. Future research should improve model process representation of respiration (including temperature and moisture controls), legacy effects of early-season productivity, soil carbon decomposition and turnover, and dynamic vegetation–permafrost interactions, to better project climate–carbon feedbacks and potential thresholds as permafrost degrades and ecosystems reorganize.

- Spatial resolution and attribution: ACI products are coarse (1°), limiting attribution to specific ecosystem types and processes; site-level EC trends are larger due to scale mismatch and local process representation. - Inversion uncertainties: Considerable spread among ACIs, differences in flux partitioning between regions, and time-dependent inversion spread persist despite ensemble averaging and GLMM treatment. - Seasonal uncertainties: Larger uncertainties in LGS and winter fluxes and trends than in EGS; respiration is derived residually (TER = GPP − NEE_ACI) and depends on satellite-based GPP estimates. - Proxy and methodological assumptions: Productivity proxies (NDVI, LUE GPP) and SEM rely on assumed causal structures and may omit unobserved drivers (e.g., disturbances, nutrient dynamics). Some episodic processes (e.g., fires) are not fully captured in EC or ACI datasets. - Model comparison constraints: TRENDY models differ in forcings and process representations; comparisons are not fully independent as some model outputs inform priors in inversions.