The study was conducted at two adjacent sites underlain by permafrost in a subarctic boreal forest, of the discontinuous permafrost region, located at the U.S. Army Cold Regions Research and Engineering Laboratory (CRREL) Permafrost Research Tunnel Facility in Fox, Alaska (64.9507° N–147.6200° W, 248 m a.s.l. (meter above sea level)). The region experiences a continental climate, which is defined by an average annual air temperature of -2.4°C, with average temperatures in July reaching 16 °C and January temperatures averaging -21.9°C; extreme temperatures throughout the year can range from -51 to 38 °C (Jorgenson et al., 2020). The two sites were situated approximately 10 m apart (Fig. 1) and exhibited comparable topography and parent material. The first site consists of an undisturbed black spruce forest ecosystem (Fig. S1 in the Supplement); the second consists of an anthropogenically disturbed area where trails were established, and black spruce tree cover was removed in firewood harvests during mining activities in the region in the 1920s.

The vegetation at the undisturbed site consists of small black spruce (Picea mariana) ranging from densely distributed to tightly spaced. Understory canopy is dominated by marsh and bog Labrador tea (Rhododendron tomentosum; groenlandicum). Forest floor cover is primarily mosses (i.e., feather mosses and Sphagnum spp.) and small shrubs including lowbush cranberry (Vaccinium vitis-idaea). The disturbed site is characterized by scattered birch (Betula neoalaskana) and tall black spruce (Picea mariana) cover. The understory canopy is primarily dwarf shrubs including marsh Labrador tea and bog blueberry (Vaccinium uliginosum). The ground surface cover is dominated by grasses (Poaceae) and sedges (Cyperaceae). The typical undisturbed soil profile at this site consists of a fibric organic layer of variable thickness, which primarily contains undecayed and partially decayed moss and forest litter materials (O horizon), underlain by an organic-rich mineral material layer containing hemic-sapric organic fraction (A/B horizon), and thick accumulations of mineral material at the base (B/C horizon). Disturbance seemingly impacts the thickness of the surface organic layer. Soil sampling at each site targeted soil materials below the organic horizon. The sampled soil material at the two sites were classified as mineral soil material (<20% organic matter [OM]; Soil Survey Staff, 2022) with a subdivision of more organic-rich fractions (10 %–18 % OM) comprising a “topsoil” layer (akin to an A horizon) that starts below the organic layer (O horizon), and more mineral-rich material (<5% OM) comprising a “subsoil” layer. The topsoil textures ranged from loam to silt loam, reflecting a higher proportion of sand particles in the topsoil relative to the silt loam-textured subsoil (Table S1 in the Supplement).

Four plots were established at an undisturbed site, and the other four were located at an adjacent disturbed site (Fig. 1). At each site, sensors were distributed to measure total soil respiration (autotrophic and heterotrophic), temperature, and volumetric water content (VWC) every 30 min from 4 November 2022–9 November 2023. During the same period, air temperature and barometric pressure were recorded every 15 min. Because some of the sensors have robotic arms to collect measurements that would be impeded by snowfall in the winter, Costco folding tables (Columbus, Indiana) with a surface of 91.4cm×91.4cm were situated above the plots, during that period, to prevent snow accumulation (Fig. 2b). The folding tables were removed at the onset of the snow-free season to reduce interference with natural environmental factors such as rain and solar radiation (Fig. 2c).

Soil temperature, VWC, air temperature, and barometric pressure were measured using the following Onset HOBO (Onset, Bourne, Massachusetts, USA) instrumentation: U30 USB Weather Station, S-TMB-M002 12-Bit Temperature Smart Sensor, S-SMC-M005 EC5 Soil Moisture Smart Sensor, S-THC-M002 Temperature/Relative Humidity Smart Sensor, and S-BPB-CM50 Smart Barometric Pressure Sensor. Soil temperature and VWC were recorded in the topsoil and subsoil layers, approximately 30 cm from each soil respiration sensor, at a depth of 18.5±1.19cm in the topsoil layer and 34.5±1.47cm in the subsoil layer for the undisturbed site. At the disturbed site, the measurements were taken at 13.5±2.53cm in the topsoil layer and 35.5±1.85cm in the subsoil layer. These depths were measured from the top of the ground vegetation cover (the moss surface). At both sites, the temperature and VWC sensors of the topsoil were installed below the fibric organic layer that had a thickness of 16.5±1.19cm at the undisturbed site and 9.5±0.87cm at the disturbed site. The differences in sensor placement depth were influenced by the variations in the thickness of the organic layer between the two sites (Fig. S2). VWC measurements were restricted to the summer and autumn seasons due to the sensor's inability to measure below freezing temperatures. Air temperature and barometric pressure were measured at 2 m above the substrate surface.

The depth of thaw in the active layer was assessed at ∼10cm from each of the eight chamber plots during each site visit (n=160) from 12 May–3 October 2023. To determine this depth, a graduated metal rod with a diameter of 1 cm (known as a frost probe) was inserted into the ground, at the same location, until it met resistance, establishing the distance between the ground surface vegetation and the top of the frozen soils (Shiklomanov et al. 2013).

For the soil respiration measurements, which encompasses the overall release of CO2 from the soil into the atmosphere (including both autotrophic and heterotrophic processes), we installed 21.3 cm diameter thick-walled polyvinyl chloride (PVC) collars (LI-COR inc., Lincoln, Nebraska, USA) in the center of each plot. They were a height of 11.4 cm and were inserted 2–3 cm into the soil through the soil vegetation cover. The collars spacing varied from two to four meters at the undisturbed site and from three to eight meters at the disturbed site (Fig. 1). We installed a 8200-104 Opaque Long-Term Chamber (LI-COR inc., Lincoln, Nebraska, USA) above each collar, which yielded a total of eight. The chambers were connected via 15 m long tubing and cable assembly to a LI-8250 multiplexer (LI-COR inc., Lincoln, Nebraska, USA) that linked to a LI-870 CO2/H2O gas analyzer (LI-COR inc., Lincoln, Nebraska, USA). The cable and tubing assembly was wrapped in 1.3 cm thick tubular pipe insulation foam to prevent internal clogging from freezing moisture to ensure data collection in winter.

The measurement process followed a closed-chamber dynamic methodology, which was automated and controlled by the LI-8250 Multiplexer. At 30 min intervals, the multiplexer would initiate a measurement at one of the chamber locations. This involved the 8200-104 chamber automatically rotating and lowering to seal onto the pre-installed PVC soil collar, creating a closed headspace over the soil surface. Once the chamber was sealed, the LI-8250 directed a closed loop of air from the chamber headspace to the LI-870 and LI-7810 analyzers and back. The gas analyzers measured CO2 and H2O concentrations, for a period of 120 s, at a rate of 1 Hz. As CO2 was respired, its concentration within the closed system increased over time. The soil CO2 flux was calculated from the rate of this concentration increase. After each measurement cycle, the chamber automatically opened, and the system proceeded to the next chamber in the sequence. A more detailed description of the soil gas flux system's operation and these winter modifications can be found in Vas et al. (2023).

Throughout the year, we performed site visits on a near-weekly basis to download data and maintain the equipment. Standard maintenance involved visually inspecting each chamber and clearing the seals of any debris. During winter months, this was supplemented by ensuring the protective tables remained free of deep snow, a necessary step to allow for sufficient air exchange to ensure the accuracy of the measurements.

Soil samples were collected from both the topsoil and subsoil layers across all plots in the autumn (September 2022), winter (February 2023), and summer (June 2023) seasons to analyze potential variations in microbial community composition among seasons, disturbance regimes, and soil layers. The samples were collected at identical depths during each season, which coincided with the depth at which the soil temperature and moisture probes were positioned; these depths varied for each plot according to the organic layer thickness. The winter samples were acquired using a gas-powered SIPRE (Snow, Ice, and Permafrost Research Establishment) corer (Jon's Machine Shop, Fairbanks, Alaska, USA). Nitrile gloves were utilized to minimize any potential contamination to the cores. Furthermore, the SIPRE corer and all associated tools were thoroughly sanitized with 70 % isopropyl alcohol, DNA away, and RNase away (Thermo Fisher Scientific in Waltham, MA, USA). The cores were then subsampled into approximately 5 cm long cylinders using a sanitized hammer and chisel and were carefully placed into sterile Nasco™ Whirl-pak bags (Thermo Fisher Scientific, Waltham, MA, USA); further information on this sampling method can be found in Barbato et al. (2022). Summer and autumn soil samples were gathered using a sanitized trowel with 70 % isopropyl alcohol, DNA away, and RNase away. The samples, approximately 5 cm thick, were placed in sterile Nasco™Whirl-pak bags and immediately placed in a cooler with frozen ice packs, then transferred to a freezer upon arrival at the Cold Regions Research and Engineering Laboratory in Fairbanks, Alaska (CRREL-AK). All collected soil samples were kept at a temperature of -25°C until shipped to CRREL in Hanover, New Hampshire (CRREL-NH), where they were stored at -20°C until further processing. Deep freezing the soil samples immediately after collection and maintaining this state until processing halts microbial growth and facilitates the preservation of the microbial community structure at the moment of sampling (Baker et al., 2023; Doherty et al., 2020).

Loss on ignition (LOI) was measured as a proxy for SOM content (Storer, 1984) on all soil samples. Here, LOI is the proportion of mass loss from oven-dried soil (dried at 105 °C for 24 h) following 2 h at 360 °C in a muffle furnace. Soil total carbon and total nitrogen was measured via combustion using a TruSpec C and N Analyzer (LECO, St. Joseph, MI, USA) at the University of Wisconsin Soil and Forage Lab. Soil pH was measured from a 1:1 slurry of soil: CaCl2 solution (0.01 M) using a pH probe (Hanna Instruments, Woonsocket, RI, USA) and a SevenEasy S20 pH meter (Mettler Toledo, Columbus, OH, USA). Soil pH was converted to H+ concentration prior to taking an average or statistical analysis. LOI total carbon, total nitrogen and soil pH were statistically analyzed using ANOVA in R.

Soil was partially defrosted and homogenized in the sample bag prior to subsampling 250 mg into bead beating tubes. Total genomic DNA was extracted using the DNeasy PowerSoil Pro Kit (Catalog No. 47014, Qiagen, Germantown, MD, USA), using a Precellys Evolution Touch homogenizer for the bead beating step (Catalog number P002511-PEVT0-A.0, Bertin Technologies, Montigny-le-Bretonneux, France). Automated DNA extraction was done with a QIAcube Connect (Catalog No. 9002864, Qiagen, Germantown, MD, USA) and each extraction run included a blank. Extracted DNA was held at -20°C.

Library preparation and sequencing was completed at Argonne National Laboratory (Lemont, IL, USA), as follows. For bacterial analysis, the V4 region of the 16S rRNA gene was targeted for PCR amplification with region-specific primers (forward primer 515F and reverse primer 806R); and for fungal analysis, the ITS region was amplified using appropriate barcoded primers (Caporaso et al., 2011, 2012; Apprill et al., 2015; Parada et al., 2016; Smith and Peay, 2014; Walters et al., 2016). Each PCR reaction contained 1 µL template DNA, 12.5 µL AccuStart II PCR ToughMix (Quantabio, Beverly, MA, USA), 1 µL forward primer with Golay barcode (5 µM concentration), 1 µL reverse primer (5 µM concentration), and 9.5 µL DNA-free PCR water. PCR conditions were: 94 °C (3 min to denature the DNA); 35 cycles of 94 °C (45 s), 50 °C (60 s), and 72 °C (90 s); final extension at 72 °C (10 min). PCR product was quantified using Quant-iT PicoGreen (P7589, Invitrogen, Waltham, MA, USA). Equimolar amounts of amplicons were pooled, purified using AMPure XP Beads (A63881, Beckman Coulter, Brea, CA, USA), quantified (Qubit, Invitrogen), and diluted to 6.75 pM using a 10 % PhiX spike. Paired-end 2×251 sequencing was done on a MiSeq (Illumina, San Diego, CA, USA).

Sequencing data were processed in R (R-Core-Team, 2018), using a dada2 v1.18.0 pipeline (Callahan et al., 2016), as in Baker et al. (2023), implemented using Snakemake v7.25.0 (Mölder et al., 2021). Taxonomy assignment was based on the SILVA 138.1 reference database for 16S sequences (Quast et al., 2012; Yilmaz et al., 2014) and the UNITE database (release 25.7.2023) for ITS sequences (Kõljalg et al., 2013; Nilsson et al., 2019). Chloroplasts and mitochondria were excluded from the dataset. A total of 2836 fungal ASVs and 10 608 bacterial ASVs were identified (excluding extraction blanks). Amplicon sequences are in the National Center for Biotechnology Information Sequence Read Archive (NCBI SRA), accession PRJNA1178745. Though our bacterial primers targeted both bacterial and archaeal 16S rRNA, we will refer simply to bacteria, which comprise 99.8 % of total reads. Of archeal reads, 80 % represented the phylum Crenarchaeota. One sample was excluded from analysis because it was mislabeled (2023_Feb, Chamber 3, Organic).

R software (R-Core-Team, 2018), and ggplot2 (Wickham, 2016) were used for data analysis and visualization; the bioinformatic approach followed West and Whitman (2022). Community composition was visualized using principal coordinates analysis (PCoA) of Bray-Curtis dissimilarities (Bray and Curtis, 1957) generated using avgdist from the R package vegan (Oksanen et al., 2024) using rarefaction (999 iterations) to a sampling depth of 16 300 for 16S and 6150 for ITS. A significant effect (p<0.05) of disturbance, sampling date, and interaction of these factors on community composition was tested within each soil layer (topsoil and subsoil), using permutational multivariate analysis of variance (PERMANOVA; adonis2 from vegan) (Anderson, 2001). Richness was evaluated using weighted linear regression (betta function in breakaway R package) (Willis et al., 2017), and a significant effect of disturbance tested via ANOVA. Differential abundance (differentialTest in the corncob package) (Martin et al., 2020) was then used to identify significant enrichment or depletion of individual taxa due to disturbance, after excluding taxa with mean relative abundance<0.00001.

For statistical analysis, data on soil respiration, soil temperature, and VWC, as well as air temperature and barometric pressure from the eight chamber plots (four located at the disturbed site and four at the undisturbed site), were averaged to obtain daily means (n=333 per variable) and seasonal means (n=4 per variable). Seasons were defined according to observed efflux seasonality at the research site, a classification consistent with previous research on boreal forest efflux (Watts et al., 2021): winter (November–March, n=123), spring (April–May, n=57), summer (June–August, n=92), and autumn (September to October, n=61). Using these processed data, we performed a series of linear regression analyses to determine the existence of distinct, linear trends between CO2 efflux and each environmental variable. This analysis was conducted separately for each site and season, and the coefficient of determination (R2) was calculated to quantify the strength of each trend. Additionally, a one-way Analysis of Variance (ANOVA) was used to test for statistically significant differences between the disturbed and undisturbed sites. A p-value of less than 0.05 was considered significant for all tests. All statistical linear regression analyses were conducted using Microsoft Excel.

A regression-based Random Forest (RF) non-linear model developed in R was used to identify the relative importance of the input variables to predict hourly and daily CO2 concentrations. RF was chosen over other algorithms due to its wide and successful application in determining variable importance (Behnamian et al., 2017; Lei et al., 2024). In RF, the supervised non-linear algorithm can combine predictions from hundreds or thousands of individual decision trees via bootstrap aggregation to generate an ideal output (Schonlau and Zou, 2020). Compared with individual decision trees, this results in an increase in generalization accuracy and a reduction in overfitting. A repeated k-fold cross-validation technique was also employed. In this technique, data are randomly separated into k subsets with k-1 used to train the model and the remainder to test the model, which is then repeated a specified amount. We selected a value of 10 for k and a value of 5 for repetition. Input variables were the same for each instance except barometric pressure being dropped for daily CO2 concentrations due to poor importance values. Thaw depth was static in the dataset that the model used from 4 October through 16 May due to the presence of a frozen surface layer preventing thaw depth probing. Organic soil VWC and mineral soil VWC from 1 November–31 March and 1 April–31 May were omitted for machine learning due to the inability of the probes to function properly in subzero temperatures.

Soil temperature fluctuated by a factor of three at the disturbed site and by a factor of 1.9 at the undisturbed site over the course of the year. The undisturbed and disturbed sites exhibited contrasting thermal regimes. At the undisturbed site, the mean annual soil temperatures were below freezing, exhibiting -0.33±0.1°C for topsoil and -0.41±0.07°C for subsoil (Fig. 3). In contrast, the disturbed site experienced positive mean annual soil temperatures, with 0.72±0.2°C for topsoil and 0.48±0.13°C for subsoil. Winter was the only season with warmer topsoil and subsoil temperatures at the undisturbed site. For both the undisturbed and disturbed sites, the subsoil layer was cooler than the topsoil layer in terms of mean annual temperature. Significantly warmer soil temperatures were observed at the disturbed site during the summer (4.04±0.19°C; p<0.001, ANOVA) and autumn (1.88±0.23°C; p<0.001, ANOVA) in comparison to the temperatures recorded at the undisturbed site during the same seasons. Conversely, the mean summer soil temperature at the undisturbed site was 1.15±0.08°C, while the mean autumn soil temperature was 0.38±0.07°C (Fig. 3). Soil temperatures in the shallower topsoil layer exhibited greater variability throughout the year compared to the temperatures in the deeper subsoil layer at both sites (Fig. 3).

VWC values ranged from 0.29±0.00 to 0.47±0.00m3m3 (1 June–31 October 2024) (Fig. S3). Subsoil layer exhibited elevated mean seasonal VWC values, as compared to the topsoil layer, at both locations (p values from 0.04 to <0.001, ANOVA). Average maximum seasonal thaw depth exhibited significant differences between the two locations (p values 0.02, ANOVA), ranging from 58±3cm or 143±29cm at the undisturbed site or disturbed site, respectively (Fig. 4). While the maximum seasonal thaw depth remained relatively consistent across the undisturbed plots, ranging from 50–61 cm, the disturbed plot displayed a larger range in maximum thaw depth, varying from 82–204 cm.

The disturbance had a significant effect on LOI (a proxy for SOM content) in the subsoil layer, but not the topsoil layer (Fig. S4). There was no significant effect of sampling date or significant interaction of sampling date and disturbance on LOI. Mean total C concentration for subsoil at the disturbed site was 2.7 % (averaged across LiCor chamber plots and sampling dates), significantly greater than that of the undisturbed site mean of 1.7 % (p<0.01; ANOVA). Mean LOI for the topsoil layer was 0.122 and 0.142 for disturbed and undisturbed, respectively (not significantly different).

Similarly, the disturbance was a significant factor for soil total C concentration and total N concentration in the subsoil layer (p<0.01 and p<0.05, respectively; ANOVA), but not the topsoil layer (Fig. 5), and there was no significant effect of sampling date or significant interaction of sampling date and disturbance on either C or N. Mean total C concentration for the topsoil layer was 4.7 % and 5.8 % for disturbed and undisturbed, respectively (not significantly different). Mean total N concentration for subsoil at the disturbed site was 0.16 %, significantly greater than that of the undisturbed mean of 0.13 % (p<0.01; ANOVA). Mean total N concentration for the topsoil was 0.19 % and 0.25 % for disturbed and undisturbed, respectively (not significantly different).

Similar to the response of LOI, total C, and total N, soil pH was significantly different in the subsoil layer, depending on the site; the effect of disturbance on the topsoil layer was not significant (Fig. S5). There was no significant effect of sampling date or significant interaction of sampling date and disturbance on soil pH. Mean pH for subsoil in the disturbed site was 4.63 (averaged across chambers and sampling dates), significantly greater than that of the undisturbed site mean of 4.11 (p<0.01; ANOVA). Mean pH for the topsoil layer was 3.78 and 3.79 for disturbed and undisturbed, respectively (not significantly different).

We compared microbial community composition and diversity between the disturbed and undisturbed sites to enhance our understanding of the role that microbes play in soil respiration. By examining the effects of the site disturbance on microbial diversity – quantified through variations in species richness and community structure – we can ascertain which microbial groups exhibit the greatest sensitivity to environmental alterations and how their functional roles might adapt in response. The dominant bacterial phyla included Proteobacteria, Acidobacteria, Actinobacteria, Verrucomicrobia, and Chloroflexi, which together comprised over 75 % of relative abundance (Fig. 6). For fungi, 54 % of relative abundance was comprised of Ascomycota (mostly classes Leotiomycetes and Archaeorhizomycetes), and 40 % was Basidiomycota (mostly class Agaricomycetes) (Fig. 7). Differential abundance testing only identified several dozen taxa (after filtering for somewhat higher abundance taxa) that were either significantly enriched or depleted under disturbance (see Figs. S6 and S7). Notably, several Chloroflexi and Dormibacterota (a newly named phylum, previously identified as Chloroflexi) ASV's are depleted under disturbance relative to the undisturbed condition in subsoil, particularly at the February sampling date (Fig. S7).

Mean estimated bacterial richness across the dataset was 711.5 ASVs, with a significant differences between the two sites (Fig. S8). Overall, the disturbed samples had 30 % higher richness compared to the undisturbed samples (p<0.01, ANOVA); within each date and soil layer combination, only the February 2023 subsoil samples demonstrated a significant effect of disturbance (p<0.01, ANOVA). There was no effect of sampling date or soil layer, or interaction amongst the factors. Mean estimated fungal richness across the dataset was 165 ASVs; there was no effect of disturbance, sampling date, soil layer, or interaction amongst the factors (Fig. S8).

Employing Bray–Curtis dissimilarities to assess beta diversity (Fig. 8), we observed notable differences between the undisturbed and disturbed sites, concerning both bacterial and fungal communities in the topsoil and subsoil layers (p<0.001 for all; R2=0.158, 0.178, 0.186, 0.115 for bacterial topsoil, bacterial subsoil, fungal topsoil, and fungal subsoil communities, respectively; PERMANOVA). Sampling date did not have a significant effect.

Total soil respiration rates showed distinct seasonal changes and were consistently higher, except during spring, at the disturbed site compared to the undisturbed one (Fig. 9). The most significant activity occurred in the summer, with mean daily fluxes reaching 4.01±0.12gC-CO2m-2d-1 at the undisturbed site and peaking at 4.81±0.17gC-CO2m-2d-1 at the disturbed site, a statistically significant difference (p<0.0001, ANOVA). Respiration rates declined in autumn, with mean daily fluxes of 1.79±0.13gC-CO2m-2d-1 (undisturbed) and 1.91±0.15gC-CO2m-2d-1 (disturbed). The mean daily fluxes decreased further in spring, measuring 1.46±0.14gC-CO2m-2d-1 (undisturbed) and 1.41±0.13gC-CO2m-2d-1 (disturbed); during these two shoulder seasons, the differences between the sites were not statistically significant. The lowest mean daily efflux was recorded during winter, with rates dropping to 0.33±0.01gC-CO2m-2d-1 at the undisturbed site and 0.39±0.13gC-CO2m-2d-1 at the disturbed site, which, similar to the summer mean daily efflux, was also a statistically significant difference (p<0.01, ANOVA). Annually, the disturbed site had a higher average soil efflux (2.07±0.11gC-CO2m-2d-1) than the undisturbed site (1.81±0.09gC-CO2m-2d-1), with the highest single mean daily flux of 9.27 gC-CO2m-2d-1 recorded on 20 July at the disturbed site.

A linear regression analysis was performed on the mean daily averages to investigate the seasonal correlation between soil efflux and various factors, including air temperature, barometric pressure, seasonal thaw depth, topsoil and subsoil temperature, and volumetric water content (VWC) at both undisturbed and disturbed sites. The analysis revealed that overall soil and air temperature exhibited the strongest correlation with soil respiration across all sites during different seasons. Soil and air temperatures showed the strongest correlations with soil efflux across all seasons. Seasonal R2 values for soil temperatures ranged from 0.21–0.65 (winter), 0.87–0.92 (spring), 0.52–0.79 (summer), and 0.85–0.93 (autumn). Air temperature also correlated strongly with soil efflux in spring (R2=0.74-0.77) and autumn (R2=0.83-0.87). Moderate correlations were observed between soil efflux and thaw depth at both sites, with R2 values ranging from 0.39–0.50 in spring, 0.41–0.43 in summer, and 0.74 in autumn.

The correlation between soil efflux and topsoil and subsoil VWC varied during the summer and fall seasons, as well as across the two soil layers. At the undisturbed site, in summer, moderate correlation was found in the topsoil (R2=0.38) and strong in the subsoil (R2=0.69). In autumn, correlations were strong in the topsoil (R2=0.62) but weaker in the subsoil (R2=0.22). At the disturbed site, in summer, the correlation was weak in the topsoil (R2=0.09) but strong in the subsoil (R2=0.64). In autumn, weak correlations were observed in both layers (R2=0.28 in topsoil and 0.10 in subsoil). Regression analysis showed weak to no correlation between soil efflux and barometric pressure during winter and autumn (R2=0.02-0.12) and no correlation in spring and summer. A slightly higher correlation was noted in winter when using hourly averages (R2=0.23 undisturbed, 0.18 disturbed).

The non-linear RF model effectively captured the impact of disturbance on the variation in soil respiration at both locations, showing strong confidence levels with high R2 and moderate to low mean absolute error (MAE) values. Specifically, the model's R2 values and corresponding MAE were 0.95 (0.14 MAE) for winter, 0.80 (0.38 MAE) for spring, 0.94 (0.16 MAE) for summer, and 0.82 (0.04 MAE) for autumn at the undisturbed site, and 0.95 (0.12 MAE) for winter, 0.83 (0.52 MAE) for spring, 0.96 (0.13 MAE) for summer, and 0.84 (0.05 MAE) for autumn at the disturbed site. Skewness and kurtosis values from the field data ranged from 0.04–1.24 and 1.61–4.56 at the disturbed site, and from 0.07–1.45 and 1.40–6.22 at the undisturbed site for all seasons, respectively. The ranges indicated that some instances contained a more normal distribution while other instances, especially winter, were notably skewed and tail-light and heavy.

The predictors' relative importance (RI) varied across seasons and sites (Fig. 10). In the winter model at the undisturbed site, air temperature emerged as the most influential predictor, accounting for 53.6 % RI. Conversely, at the disturbed site, the subsoil temperature exhibited the highest predictive power with 49.6 % RI (Fig. 10). For the spring model, soil temperature was the best predictor at both sites, the topsoil layer temperature exhibited a slightly greater predictive power at the undisturbed site, whereas the subsoil layer temperature proved to be more influential at the disturbed site. In the undisturbed site's summer model, the soil moisture and temperature as well as the seasonal thaw depth variables exhibit similar RI values, ranging from 15.7 % to 20.2 %. Among these variables, air temperature has the lowest predictive power, with an RI of 10.3 %. Conversely, at the disturbed site, the topsoil layer temperature stands out as the strongest predictor, with an RI of 44.6 %. It is followed by the topsoil VWC, which has an RI of 13.7 %. The undisturbed site's autumn model indicates that subsoil layer VWC has the highest predictive power at 24.3 % RI, followed by subsoil layer temperature at 18.9 % RI, and topsoil layer temperature at 18.7 % RI. However, the disturbed autumn model shows that subsoil layer temperature is the most influential factor in predicting CO2 efflux with 24.3 % RI and topsoil layer (21.1 % RI) and air (18.9 % RI) also playing significant roles.

Throughout the different seasons, various factors contribute to the varying degrees of relative importance regarding soil respiration. These factors include air temperature, the temperature of both topsoil and subsoil layers, and volumetric water content, which pertains only to subsoil. This information is presented in Table 1.