Greening of the land surface has attracted attention as the enhanced vegetation productivity may have contributed to 12.4% of the accumulated terrestrial carbon sink1. Atmospheric CO2 concentration, climate change, atmospheric nitrogen (N) deposition (Ndeposition), and land-use change have been identified as the main drivers of vegetation greening2. Of these factors, climate warming is known to dominate the increase in plant productivity in the boreal biome3. However, recent studies have found that the traditional positive relationship between warming and plant productivity has been decoupled4,5. Specifically, changes in plant productivity cannot be fully explained by these factors in the sub-regions of the boreal biome3.

Permafrost degradation and related ecological processes may drive the decoupling between climate warming and vegetation greening in the northern permafrost region but remain theoretical considerations4. Permafrost degradation associated with summer warming and changes in the thickness and duration of snow cover caused by winter warming6,7 also play an essential role8,9. Variations in snow conditions could affect plant growth by altering soil thermodynamics and biogeochemistry10,11. There is considerable observational12,13,14,15 and experimental16,17,18,19 evidence for the effect of snow on spring greening and plant productivity by affecting snowmelt dates (SMD) and soil water availability, individually or in combination. In addition, changes in snow would affect soil thermodynamics and permafrost thaw and, in turn, associated soil biogeochemical cycles20,21, indirectly affecting plant growth22,23,24. The two ecological processes of the effect of snow-associated SMD and soil thermodynamics on plant productivity may always be interlinked. The timing and depth of soil thaw may always be determined by SMD, which changes the surface energy balance and soil thermodynamics14,21. Relationships between spring greening and SMD may be affected by the timing and depth of soil thaw, possibly because they determine the timing and amount of water and nutrients available to plants after snowmelt25,26. However, the cascading effects of snow on plant productivity by affecting permafrost thaw and the associated availability of water and nutrients to plants and spring greening are generally neglected (Fig. 1).

Fig. 1: Schematic path diagram showing the effect of shortened snow cover duration on plant productivity.
Fig. 1: Schematic path diagram showing the effect of shortened snow cover duration on plant productivity.
Full size image

Cascading relationships of the three pathways include six processes: snow cover duration length (SCD) in the snow hydrological year; snowmelt dates (SMD); thermal regime of the soil ground after snowmelt (soil temperature, Tsoil); soil thaw including start of soil thaw (SOT) and active layer thickness (ALT); N-uptake (Nplant), water availability (soil water content, SWC) and plant phenology (start of season, SOS). Climate factors (air temperature, precipitation, solar radiation), atmospheric N deposition, and runoff are considered in the cascading relationships.

There is compelling evidence of a trend towards less snow in the northern permafrost region in recent warm decades6,7,9,27. However, a few studies have examined the effects of reduced snow on plant productivity16, and the effects have been positive9,28 or negative24. The contrasting effects may be due to soil thermodynamics, which are not currently understood28. Warm soils with reduced snow could promote plant growth by accelerating permafrost thaw, increasing plant water and nutrient availability. In contrast, cold soils with reduced snow always suppress plant productivity, possibly by limiting permafrost thaw and its associated positive effect of increasing water availability and nutrient uptake for plants24 or by increasing the risk of frost damage10. It is still a mystery how reduced snow affects plant productivity through these cascading relationships and whether they are related to soil thermodynamics. Furthermore, less snow is always associated with a warming climate6. It will be valuable to disentangle the contribution of reduced snow to plant productivity from climate warming.

Here, we compiled regional-scale remoting sensing datasets to investigate the effects of warming-independent shortened snow cover duration on plant productivity across the northern permafrost region. The snow cover duration was defined as the cumulative number of days with at least 10 cm of snow on the ground in a snow hydrological year (September to May of the following year27). Years with shortened snow cover duration were those being below the lower limit of the 90% confidence interval of the snow cover duration time series. Specifically, we quantified changes in soil temperature (Tsoil), plant productivity, and their associated drivers in years with shortened snow cover duration compared to the long-term mean and parsed their differences between the warm and cold soil regions. Although there was no direct relationship between snow and plant productivity12, shortened snow cover duration can have cascading effects on growing season productivity by altering the snowmelt dates (SMD), Tsoil28,29, start of soil thaw (SOT), active layer thickness (ALT)20,24 and in turn, the known drivers of the start of season (SOS)12,26, soil water content (SWC)12,26,30 and plant N uptake (Nplant)22,31 (Fig. 1). Three pathways were summed by these cascading relationships. The first one is referenced as the “SOS” pathway, which mainly works through changing Tsoil, SOT, and, in turn SOS and plant productivity (Fig. 1). The other two are identified as the “N-uptake” (and “SWC”) pathways, which operate by altering Tsoil, ALT and, in turn, Nplant (and SWC) and plant productivity (Fig. 1). The three pathways were estimated by the mediation effect models when climate factors (temperature, precipitation, and radiation), runoff, and Ndeposition were taken into account.

Our data suggested that plant productivity increased in years with shortened snow cover duration independent of climate factors in the warm soil region. The “N-uptake”, “SOS”, and “SWC” pathways were the main factors explaining the increase in plant productivity, accounting for half of the increase. In contrast, shortened snow cover duration decreased plant productivity in the cold soil region of the highlands, as the beneficial effects of the three pathways on plant productivity were weakened or turned negative.

Results

Shortened snow cover duration altered soil temperature

Warm soils with shortened snow cover duration prevailed in the northern permafrost region. There were \(14.66^{12.81}_{17.59}\) days (\({{\rm{median}}}^{25th}_{75th}\)) fewer days of snow cover duration in years with shortened snow cover duration than the long-term mean (Supplementary Fig. 1). Changes in Tsoil were not uniform across the study region. Cold soil region was always distributed in the Ural Mountains and East Siberian highlands, whereas the warm soils predominated in the lowlands (Fig. 2a–f). Overall, the increases in Tsoil were observed in 79.7 ± 2.0% (mean ± 1 standard error; three Tsoil datasets) of the study region (56.3 ± 2.3% with a significant increase, p < 0.05) (Fig. 2a–f). Compared to the long-term mean, Tsoil during the early growing season increased by \(0.17^{0.05}_{0.28}\)°C for every 10-day shortened snow cover duration in the topsoil and by \(0.14^{0.03}_{0.24}\)°C in the subsoil (Fig. 2a–f).

Fig. 2: Spatial distributions of changes in Tsoil in years with shortened snow cover duration compared to the long-term mean.
Fig. 2: Spatial distributions of changes in Tsoil in years with shortened snow cover duration compared to the long-term mean.
Full size image

The spatial distributions of the differences in Tsoil between years with shortened snow cover duration and long-term mean (ΔTsoil for every 10-day shortened snow cover duration). a Topsoil 0–7 cm of ERA data; b subsoil 7–28 cm of ERA data; c topsoil 0–10 cm of MERRA data; d subsoil 10–20 cm of MERRA data; e topsoil 0–5 cm of GLDAS data; f subsoil 5–20 cm of GLDAS data. Black points indicate the significant changes in Tsoil (p < 0.05). The relative frequency (%) distribution of ΔTsoil in the center of the insets: significant decreases (Neg*), significant increases (Pos*), insignificant decreases (Neg), insignificant increases (Pos). ERA ERA5-Land European Centre for Medium-Range Weather Forecasts Reanalysis v5, MERRA MERRA-2, Modern-Era Retrospective Analysis for Research and Applications version 2, GLDAS global land data assimilation system.

Shortened snow cover duration increased plant productivity

Plant productivity increased in years with shortened snow cover duration. Compared to the long-term mean, increases in gross primary productivity (GPP) and normalized difference vegetation index (NDVI) in years with shortened snow cover duration were observed in more than 60% of the study region (more than 30% with a significant increase, p < 0.05) (Fig. 3a–c). Changes in plant productivity tended to increase with increasing Tsoil (Fig. 3a, b). Increases in GPP and NDVI were observed in the warm soil region, but decreases predominated in the cold soil region (Fig. 3d; Table 1). To further assess the robustness of our results, we repeated our analyses by selecting the years with shortened snow cover duration but standard climate factors, which yielded results consistent with those above (Supplementary Table 1). Overall, these findings consistently indicated that plant productivity tended to increase in years with shortened snow cover duration in the warm soil region but not in the cold soil region.

Fig. 3: Changes in plant productivity in years with shortened snow cover duration.
Fig. 3: Changes in plant productivity in years with shortened snow cover duration.
Full size image

a, b Spatial distributions of changes in GPP (ΔGPP) and NDVI (ΔNDVI) for every 10-day shortened snow cover duration compared to the long-term mean. Black points indicate the significant changes (p < 0.05). The horizontal axis of the color legend is ΔTsoil, and the vertical axis of the color legend is ΔNDVI and ΔGPP. In this legend, the green color indicates ΔNDVI and ΔGPP are positive, and ΔTsoil is positive; the yellow color indicates ΔNDVI and ΔGPP are negative, and ΔTsoil is negative; the blue color indicates ΔNDVI and ΔGPP are negative, and ΔTsoil is positive; the red color indicates ΔNDVI and ΔGPP are positive, and ΔTsoil is negative. c The relative frequency (%) distribution of ΔNDVI and ΔGPP is significant negative (Neg*), significant positive (Pos*), non-significant negative (Neg), non-significant positive (Pos), sum of the relative frequency of the significant and non-significant positive (T_Pos), and the sum of the relative frequency of the significant and non-significant negative (T_Neg). d Density distributions of NDVI and GPP between warm soil and cold soil regions in years with shortened snow cover duration only (“Shortened snow cover duration”) and in combination with normal climates (“Shortened snow cover duration +Climate normal”). The extent of the density distributions indicates the 25th and 75th percentiles, and the dots in the violins and black lines in the boxes indicate the median of ΔNDVI and ΔGPP for every 10-day shortened snow cover duration.

Table 1 Summary of changes in variables (for every 10-day shortened snow cover duration) and the results of the three pathways in years with shortened snow cover duration
Full size table

Biogeochemical responses to shortened snow cover duration

Earlier and deeper soil thaw, earlier spring greening, and greater plant N uptake were observed in years with shortened snow cover duration. Compared to the long-term mean, the SMD advance was observed in 78.8% of the study region (p < 0.05), the SOT advance in 68.4% of the study region (p < 0.05), and the SOS advance in 74.1% of the study region (p < 0.05) (Supplementary Fig. 2a–c; Table 1). In years with shortened snow cover duration, ALT was above the long-term mean in 76.0% of the study region (45.7% with a significant increase, p < 0.05), and increases in Nplant were observed in 87.5% of the study region (55.2% with a significant increase, p < 0.05) (Supplementary Fig. 2d, e; Table 1). There was a decrease in SWC in 69.0% of the study region (35.63% with a significant decrease, p < 0.05) (Supplementary Fig. 2f; Table 1).

Changes in these variables showed substantial spatial heterogeneity. Earlier SMD, SOT, and SOS in years with shortened snow cover duration were found in the warm soil region than those in the cold soil region (Supplementary Fig. 2a–c; Table 1). There were greater increases in ALT and Nplant in the warm soil region than in the cold soil region (Supplementary Fig. 2d, e; Table 1). Declines in SWC were greater in the warm soil region than in the cold soil region (Supplementary Fig. 2f; Table 1). In years with shortened snow cover duration but standard climate factors, changes in these variables between the warm and cold soil regions were generally in line with the years with shortened snow cover duration alone (Supplementary Table 1).

Mechanisms of shortened snow cover duration caused vegetation greening

We used the mediation effect model to quantify the contribution of the three mechanisms to vegetation greenness when the climate factors, Ndeposition, and runoff were considered. Shortened snow cover duration advanced snowmelt and accelerated permafrost thaw, increasing plant water and nutrient uptake and bringing forward spring greening, which in turn increased plant productivity. The positive relationship between shortened snow cover duration and SMD was observed in 58.9% (35.0% with a significant positive relationship, p < 0.05) of the study region (Supplementary Fig. 3a). Early SMD increased Tsoil in 64.6% (41.6% with a significant negative relationship, p < 0.05) of the study region (Supplementary Fig. 3b). Increases in Tsoil advanced SOT in 68.9% (50.2% with a significant negative relationship, p < 0.05) of the study region and increased ALT in 57.8% (36.5% with a significant positive relationship, p < 0.05) of the study region (Supplementary Fig. 3c, d). SOS advanced with early SOT in 85.3% (67.9% with a significant positive relationship, p < 0.05) of the study region and SWC and Nplant increased with increasing ALT in more than 55% (more than 30% with a significant positive relationship, p < 0.05) of the study region (Supplementary Fig. 3e–g). Advances of SOS and increases in SWC and Nplant increased GPP were observed in more than 60% (more than 39% with a significant relationship, p < 0.05) of the study region (Supplementary Fig. 3h–j).

The negative indirect correlations between shortened snow cover duration and GPP, NDVI through the “SOS”, “SWC”, and “N-uptake” pathways were observed in over 55% of the study region (more than 25% with a significant negative correlation, p < 0.05) (Fig. 4a–f). The negative standardized indirect coefficients (CoefIndirect) of the “SOS”, “SWC”, and “N-uptake” pathways indicated that shortened snow cover duration increased GPP and NDVI through the three pathways (Fig. 4a–f; Table 1; Methods). The contributions of the three pathways to changes in plant productivity varied across the space. Compared to the cold soil region, more increases in GPP and NDVI were observed in the warm soil region (Fig. 4a–f; Table 1). Overall, the three pathways accounted for 55.6% of the increase in GPP in years with shortened snow cover duration and 56.7% of the increase in NDVI (Table 1). These results suggested that the shortened snow cover duration, which drove soil thermodynamics, permafrost thaw, and its associated effects on water and nutrients for plants and spring greening, was crucial in explaining vegetation greening in the northern permafrost region.

Fig. 4: Spatial distributions of standardized indirect coefficients of the “SWC”, “SOS”, and “N-uptake” pathways.
Fig. 4: Spatial distributions of standardized indirect coefficients of the “SWC”, “SOS”, and “N-uptake” pathways.
Full size image

a–f Spatial distributions of the standardized indirect coefficients (CoefIndirect) of the three pathways. Black points indicate the significant standardized indirect effect coefficients (p < 0.05). The horizontal axis of the color legend is the changes in Tsoil in years with shortened snow cover duration compared to the long-term mean, and the vertical axis of the color legend is the standardized indirect coefficients of the “SWC”, “SOS” and “N-uptake” pathways. In this legend, the green color indicates CoefIndirect is positive, and ΔTsoil is positive; the yellow color indicates CoefIndirect is negative, and ΔTsoil is negative; the blue color indicates CoefIndirect is negative, and ΔTsoil is positive; the red color indicates CoefIndirect is positive, and ΔTsoil is negative. g The relative frequency (%) distribution of CoefIndirect is significant negative (Neg*), significant positive (Pos*), non-significant negative (Neg), non-significant positive (Pos), sum of the relative frequency of the significant and non-significant positive (T_Pos), and the sum of the relative frequency of the significant and non-significant negative (T_Neg).

To further assess the robustness of the results, we used other datasets to estimate the three pathways. We used the plant N uptake from the Coupled Model Inter-comparison Project Phase 6 (CMIP6) to replace that from the Trends in net land-atmosphere carbon exchange dynamic global vegetation models (TRENDY DGVMs) and estimated the “N-uptake” pathway. The CoefIndirect of the “N-uptake” pathway estimated by the CMIP6 was consistent with the TRENDY DGVMs, with correlation coefficients close to 0.70 (Supplementary Fig. 4). In addition, climatic water deficit (CWD) was used to replace the Global Land Evaporation Amsterdam Model (GLEAM) SWC to estimate the “SWC” pathway. The correlation coefficients of the CoefIndirect of the “SWC” pathway between the two datasets were >0.70 (Supplementary Fig. 4).

Discussion

Reduced snow depth and duration have become increasingly prevalent in recent decades under a warming climate6,7,9. Our study is one of the attempts to develop a framework to isolate the effects of climate warming and to build the cascading relationships between shortened snow cover duration and plant productivity (Fig. 5a). In years with shortened snow cover duration, increased plant productivity was due to earlier and deeper soil thaw, resulting in earlier spring greening and greater N and water availability to plants in the warm soil region (Fig. 5b, c). Permafrost thaw and its associated ecological effects are driven by shortened snow cover duration, which has not received much attention from the permafrost research community12, yet they play a key role in the carbon balance in the permafrost region considering the warming climate.

Fig. 5: Schematic diagram of the magnitude and mechanism of changes in plant productivity in years with shortened snow cover duration.
Fig. 5: Schematic diagram of the magnitude and mechanism of changes in plant productivity in years with shortened snow cover duration.
Full size image

a The cascading relationship between changes in snow cover duration and GPP through the “SWC”, “SOS”, and “N-uptake” pathways in years with shortened snow cover duration. To avoid the confusing lines in the cascading relationships, we use “f()” to express the relationships between the study variables in parentheses. b Comparison of changes in the study variables between the years with shortened snow cover duration and the long-term mean in the warm and cold soil regions. Increases (decreases and advances) are indicated by the position of the variable boxes above (below) the red dotted lines. c Contributions of three pathways to changes in GPP in years with shortened snow cover duration in the cold and warm soil regions. SCD snow cover duration, SMD snowmelt dates, Tsoil soil temperature, SOT start of soil thaw, ALT active layer thickness, Nplant plant N uptake, SWC soil water content, SOS start of season, GPP gross primary productivity.

The warm soils with shortened snow cover duration led to earlier and deeper soil thaw. Snow can strongly influence soil thermodynamics by altering surface albedo, emissivity and absorptivity, thermal conductivity, and latent heat32. The effects of reduced snow on soil thermodynamics vary. Both cold29,33 and warm soils28,34 have been reported in association with reduced snow. The soil thermodynamics of reduced snow could be determined by the freezing strength of the soil in winter and the heat uptake after snowmelt in spring. Observational and experimental studies have reported that reduced snow and increased35 or slightly decreased36 winter Tsoil always synergistically occur under climate warming. In spring, reduced snow is expected to melt more quickly and early due to less heat requirements16,19. Early melting allows the soil to absorb radiation earlier and warms the soil faster, thereby accelerating soil thawing9,19,32,34,37.

Plant productivity is strongly influenced by the onset of spring greening after the soil thawing and is enhanced by the “SOS” pathway. Water and nutrients released from the melting snow are available to plants so that SMD may dominate SOS in the boreal biomes12,13,14,38. However, the synchronization between SMD and SOS is reported to be decoupled39,40. The onset of plant activity in spring is strongly influenced by the timing of the transition from frozen to thawed soil and the transition from solid to liquid soil water25,26. As soils begin to thaw, fine root activity is triggered, increasing water and nutrient availability to plants19,34,41. In addition, the CO2 pulse caused by soil thawing could enhance the activity of a key enzyme in photosynthesis and accelerate the onset of spring greening42. The SOT could be a signal and a key to restoring physiological activity43 and dominate the SOS but not the climate factors26. Early SOS resulting from increased Tsoil and early SOT in years with shortened snow cover duration extends the plants available for resource allocation of water and nutrients, promoting plant productivity in the early growing season28.

Plant growth could benefit from permafrost-released N and the “N-uptake” pathway as permafrost thaws in deeper soils. Large amounts of plant-available N have been stored in the permafrost31. As permafrost thaws, the newly available N would be released and would be of the same order of magnitude as Ndeposition31. Stable isotope labeling has shown that the new N source would be available for plants20,22,25,31,44,45, although mismatch in the timing of new N release and root uptake46, high new N leaching in autumn47, and competition between plants and soil microorganisms45 may limit the new N uptake in space and time. Especially for the deep-rooting species, deeper roots allow them to proliferate and take up N in newly available soil space, even after above-ground senescence25. In the longer term, the new N would be available to shallower rooting species as the new N is redistributed in the active layers44. New N available to plants would relieve N limitation and increase plant productivity in the northern permafrost region45,48. As previously analyzed, a favorable soil thermal environment due to shortened snow cover duration promotes permafrost thaw, and the associated increase in N availability is important for enhancing plant productivity.

Increases in plant productivity were attributed to the “SWC” pathway. Water availability from snowmelt is vital in influencing plant growth in the boreal biome14,38. Reduced snow always results in less snowmelt, directly exacerbating water limitations for plant growth12,14. Additionally, early spring greening in years with shortened snow cover duration may be expected to advance soil moisture depletion and amplify plant water stress12,49,50, possibly leading to the decreases in SWC in the warm soil region being greater than that in the cold soil region. Although snow depth and duration have shown a decreasing trend in the past decades6,7,27, the northern permafrost region is still a “vegetation water surplus region”51. This may be mainly attributable to the water release from thawing permafrost52,53. Permafrost degeneration would amplify the sensitivity of plants to permafrost thaw-induced changes in hydrology54. Recent studies further confirm that increasing soil thaw depth can improve soil water availability, advance spring greening26, and promote plant productivity30. Overall, shortened snow cover duration could suppress plant growth by exacerbating water limitations, but this suppression could be partially offset by the positive effect of the “SWC” pathway.

Cold soils with shortened snow cover duration in the highlands could limit soil thaw and sequentially reduce the positive effects of the “SOS”, “N-uptake”, and “SWC” pathways. Shortened snow cover duration, combined with lower ambient temperatures, dramatically reduces the snow insulation, increasing the freezing strength of the ground in winter55. Additionally, snow in the mountains can be densely distributed in the valley due to strong winter winds, and this uneven distribution can further increase ground freezing. Insufficient heat accumulation after snowmelt may not be able to compensate for the strong soil freeze in cold winter, leading to cold soil with less snow, as shown by our findings and a previous study28. Cold soils can limit permafrost thaw, reducing plant N and water availability and plant productivity24, i.e., reducing the beneficial effects of the “SWC” and the “N-uptake” pathways. N availability may be further limited by weakened microbial activity and N mineralization resulting from low Tsoil in winter and early growing season20. In addition, a large drop in elevation in the highlands could increase the runoff as the permafrost thaws, reducing the availability of water and nutrients to plants, which would become an additional explanation for the decrease in plant productivity. For the “SOS” pathway, although SOS was advanced and increased plant productivity was contributed to the “SOS” pathway in these areas, this positive effect may be limited by early exposure to cold temperatures and frost damage to plants16,17,38.

Limitations should be acknowledged when we explain changes in plant productivity in years with shortened snow cover duration. Although we considered the “N-uptake” pathway as expected by the snow research community12, it is important to consider the limitations of the data. At the in-situ scale, the positive effect of the permafrost-released N on plant productivity is always examined using the stable isotope labeling in the high- and sub-Arctic20,22,25,31,44, but not in the southern of the northern permafrost region. At the continental scale, a long-term observation of plant N uptake has mostly come from model projections, such as the CMIP6 and TRENDY DGVMs. The data limitations may lead to biases in the estimation of the “N-uptake” pathway at the large scale, and the biases may be exacerbated by the uneven spatial distribution in the in-situ scale observations. In addition, the active layer can function as a near-surface aquifer to regulate runoff and streamflow response to snowmelt as well as nutrient and water transport and cycling, but these processes are not explicitly considered in the three pathways largely because of the data scarcity. These limitations warrant further studies to validate the results presented here.

Conclusion

The effects of shortened snow cover duration remain poorly characterized due to its association with warming’s impacts. Our integrative analyses demonstrated increased plant productivity in the early growing season in years with shortened snow cover duration in the warm soil region. This positive effect was due to the cascading effects of the shortened snow cover duration accelerating permafrost thaw, which increased water and N availability to plants and advanced spring greening. In years with shortened snow cover duration, the cascading effects contributed ~1/2 of the increase in plant productivity. Future shallower and more transient snowfall, combined with a warmer climate, can be expected to increase vegetation greenness by accelerating permafrost thaw. This could address concerns about uncertainties in the warming-greenness relationship and accurately quantify and project the land carbon sink in the northern permafrost region.

Methods

Study domain

Permafrost is defined as the ground that remains at or below 0 °C for at least two consecutive years56. The northern permafrost region, as our study region, is mainly located in Siberia, Alaska, the Canadian Arctic, and high mountain regions. We removed pixels with the mean annual NDVI < 0.1 to eliminate the noise of unvegetated and sparse vegetation areas and removed all cultivated land to eliminate the effects of human activity26.

Plant productivity

Satellite-based GPP and NDVI were used to estimate changes in plant productivity. This is because the biotic and abiotic factors considered in our analyses are not used in producing both datasets, thus avoiding potential inherent bias. Satellite-based near-infrared reflectance (NIRv) is used as a proxy of GPP57. GPPNIRv is based on the robust empirical NIRv-GPP relationships with R2 > 0.70 using eddy-covariance flux data57,58. Global annual GPPNIRv and inter-annual trends in GPPNIRv fall between estimates from the machine learning, light use efficiency models, and process-based models58. Recently, there has been evidence that GPPNIRv performs well in estimating the response of plant productivity to climate change57,59.

NDVI was widely used as a proxy for plant productivity to investigate the response of plant productivity to changes in snow12. We used a new version of the GIMMS NDVI product (PKU GIMMS NDVI) to characterize changes in plant productivity in response to shortened snow cover duration. PKU GIMMS NDVI is produced by massive, high-quality global Landsat NDVI samples and a data consolidation method60. It overcomes several limitations (e.g., orbital drift and sensor degradation) compared to previous versions60.

Environmental variables

Changes in climate factors and Tsoil were used to attribute changes in plant productivity. Monthly gridded temperature, precipitation, and downward shortwave radiation data were from a high-resolution (1/24°) climate dataset from TerraClimate61. The CWD from the TerraCliamte is estimated by a water balance model that incorporates reference evapotranspiration, precipitation, temperature, and interpolated plant extractable soil water capacity61. Monthly root-zone SWC is calculated using a multi-layer running-water balance model at a spatial resolution of 0.25° from GLEAM v3.8a dataset62. Monthly Tsoil derived from the European Centre for Medium-Range Weather Forecasts Reanalysis v5 (ERA5-Land; 0.1° × 0.1°; 0–7 cm and 7–28 cm)63, Modern-Era Retrospective Analysis for Research and Applications version 2 (MERRA-2; 0.5° × 0.625°; 0–10 cm and 10–20 cm; M2TMNXLND)64, and Global Land Data Assimilation System (GLDAS; 0.25° × 0.25°; 0–5 cm and 5–20 cm, GLDAS_NOAH025_M)65.

Soil thaw and plant N uptake

Soil thaw was measured by SOT and ALT. SOT is estimated from the satellite soil freeze/thaw Earth system dataset (FT-ESDR)8. It is generated by calibrating 37 GHz brightness temperature retrievals from the Scanning Multichannel Microwave Radiometer (SMMR), Special Sensor Microwave Imager (SSM/I), and SSM/I Sounder (SSMIS). The FT-ESDR contains a global record of the daily freeze/thaw status (F/T) of the landscape. Daily AM and PM F/T status indicates whether the cell is frozen (0) or thawed (1) during the corresponding morning or afternoon overpass. Daily F/T status reports a single, combined daily value for the cell using the following scheme:

• AM and PM frozen (0)

• AM and PM thawed (1)

• AM frozen, PM thawed (2)

• AM thawed, PM frozen (3)

The SOT is determined by the date with a transition of state code from 0 to 1 (i.e., from frozen to thawed)8,26,66. Both ALT datasets were used in this study. The first one is calculated by the Stefan equation with a spatial resolution of 0.08°26 (ALTs, Eq. 1). Another one is produced by the machine learning models, and climate, plant, and soil factors are considered in the models (ALTm)67. ALT in the study was used as the mean of both datasets. A detailed description of the SOT can be found in ref. 8 and of the ALTs in ref. 26, and of the ALTm in ref. 67.

$${ALT}=\sqrt{\frac{2\lambda {ns}}{Q} \times {DDT}}$$
(1)

DDT: degree-days of thaw, as the sum of the daily mean temperature above 0 °C; \(\lambda\), the thermal conductivity of thawing soil; s, time conversion factor; Q: the volumetric latent heat of fusion; n, the ratio of degree-day sums of the ground surface to air temperatures when both are above 0 °C.

Plant N uptake was estimated based on the model outputs. Monthly N mass in leaves, roots, and stems were retrieved from six Earth system models (ESMs) in the CMIP6 in the historical scenario. Plant N content was calculated as the sum of N mass in leaves, roots, and stems. The six ESMs are Community Earth System Model version 2 (CESM2), CESM2-finite volume 2° (CESM2-FV2), CESM2-Whole Atmosphere Community Climate Model (CESM2-WACCM), Norwegian Earth System Model-Low-Resolution Model (NorESM2-LM), NorESM2-Medium Resolution Model (NorESM2-MM), and Taiwan Earth System Model version 1 (TaiESM1). In addition, the N uptake of vegetation was from TRENDY DGVMs v1268. For the TRENDY DGVMs, we selected the S3 simulation scenario that included CO2, climate, and land use (all forcing time-varying)68. The six TRENDY DGVMs are Commonwealth Atmosphere Biosphere Land Exchange model (CABLE), Jena Scheme for Biosphere-Atmosphere Coupling in Hamburg (JSBACH), Joint UK Land Environment Simulator (JULES), Lund-Potsdam-Jena managed Land (LPJmL), Lund-Potsdam-Jena Phenology and productivity model (LPX), and Organizing Carbon and Hydrology In Dynamic Ecosystems Europe (ORCHIDEE).

Snow cover duration

Snow cover duration in our study was derived from the daily snow depth product estimated by a novel data fusion framework based on machine learning models (fused snow depth dataset). The algorithm integrates snow depth products from the Global Snow Monitoring for Climate Research (GlobSnow), Advanced Microwave Scanning Radiometer-Earth Observing System (AMSR-E), AMSR-2, National Snow Depth Database, ERA-Interim, and MERRA-2, incorporating geolocation and topographic data27. The models are trained and validated using over 30000 ground observations and outperform the best original GlobSnow gridded snow depth product, with an R2 improvement of 0.58 (0.23 vs. 0.81) and a Root Mean Squared Error reduction of 8.17 cm (15.86 vs. 7.69 cm)27. Most of the bias (≈90%) between the fused snow depth and in-situ observations is distributed from −10 cm to 10 cm27. Snow cover duration was defined as the duration of daily snow depths of greater than 10 cm covering the ground in a snow hydrological year from September to May of the following year.

SOS and SMD

SOS is estimated based on the latest GIMMS NDVI dataset. Abnormal NDVI data is excluded by a modified Savitzky-Golay filter69. The SOS is calculated by five widely used phenology retrieval methods (e.g., double logistic function)69. The dataset has been shown to ensure the quality of SOS variability under climate change in the boreal biomes69. In this study, we used the mean of five methods to analyze changes in SOS.

The fixed threshold12, i.e., the SMD is defined when the daily snow depth is less than the set fixed threshold, and the dynamic threshold14 has always been used to estimate the SMD. Compared to the dynamic threshold, the fixed threshold may be more susceptible to outliers, and the threshold selection may be more complex. Recently, the Normalized Difference Snow Index (NDSI) has been used to estimate the SMD in the northern permafrost region using a dynamic threshold method, with an R2 of the correlation between observations and estimates close to 0.6014. The NDSI dataset of the MODIS snow product (MOD10A1 Collection 6) after 2000 does not cover our study period (1982–2014). In this study, we used the fused snow depth dataset instead of NDSI because both datasets showed the same trend of variations14,27. Regarding estimating the SMD26, we used a Savitzky-Golay filter to exclude abnormal daily snow depth data in a snow hydrological year. Then, the dynamic threshold method was used to extract the SMD from the time series of daily snow depth, as Eq. 2:

$${{Snow}{{\_}}{depth}}_{{threhold}}=\left(\frac{{{Snow}{{\_}}{depth}}_{{DOY}}-{{Snow}{{\_}}{depth}}_{\min }}{{{Snow}{{\_}}{depth}}_{\max }-{{Snow}{{\_}}{depth}}_{\min }}\right)\times 10 \%$$
(2)

Where Snow_depthDOY, is the daily snow depth on a Julian date of year. Snow_depthmin, and Snow_depthmax are the minimum and maximum of the daily snow depth in a snow hydrological year, respectively. When the Snow_depthDOY was firstly less than the Snow_depththreshold, the corresponding Julian date of year was defined as SDM. The SMD, as estimated by our study, was in good agreement with the NDSI estimate14, with a mean correlation coefficient of 0.90 (Supplementary Fig. 5).

Statistical analyses

All the statistical analyses in this study were performed using R 4.3.0. The statistical differences in GPP, NDVI and their biotic and abiotic drivers between years with shortened snow cover duration and the long-term mean were estimated by the one-sample t-test using t.test function. Shapiro-Wilk was used to test the normal distributions of the variables using the shapiro.test function. If study variables did not fit the normal distributions, we used a non-parametric test by the wilcox.test function.

Mediation effect models estimated with the sem function in the lavaan package were used to investigate the cascading effects of the “SWC”, “SOS”, and “N-uptake” pathways on changes in plant productivity. The mediation effect model can test a hypothetical causal chain where one variable X (independent variable) affects a second variable M (mediators), which in turn affects a third variable Y (dependent variable). In our study, the independent variable was Δ snow cover duration ΔSCD, and the dependent variables were ΔGPP and ΔNDVI. Mediators were ΔSMD, ΔTsoil, ΔALT, and ΔNplant (ΔSWC in the “SWC” pathway) in the “N-uptake” pathway. The “SOS” pathway highlighted the regulations of ΔSMD, ΔTsoil, ΔSOT, and ΔSOS. The covariates in the three pathways were Δair temperature, Δprecipitation, Δsolar radiation, Δrunoff, and ΔNdeposition. In the mediation models, CoefIndirect of the three pathways suggested that a unit change of one standard deviation (sd) in ΔSCD caused a CoefIndirect unit change in sd of ΔGPP and ΔNDVI. We further quantified the contributions of the three pathways to changes in plant productivity (ConΔGPP and ΔNDVI) in each pixel based on Eq. 3. The satellite-derived datasets were aggregated into a spatial resolution of 0.5° during the early growing season (April–July) from 1982 to 201412. The satellite-derived analysis was applied over 2.5° by 2.5° moving windows centered on each pixel to satisfy the sample size requirements12.

$${{Con}}_{\Delta {GPP\; or\; NDVI}}=\frac{\Delta {SCD}}{{sd}(\Delta {SCD})}\times {{Coef}}_{{Indirect}}\times {sd}(\Delta {GPP\; and }\;\Delta {NDVI})$$
(3)