Phenological transition dates are extracted from the GLASS (Global Land Surface Satellite) LAI dataset , which is primarily based on MODIS observations. The original 8-daily, 0.05° LAI data are linearly interpolated in space to 0.25° and in time to daily resolution.

SOS and EOS are determined following the methodology of . For each grid cell, the daily LAI time series is first smoothed using a spline with 10 nodes implemented in the EasyPhenology software package . SOS and EOS are then defined as the day of year (DOY) when LAI exceeds and falls below 30 % of the annual maximum LAI, respectively (Fig. ). The 30 % threshold was selected as a widely used and methodologically consistent choice within the accepted 20 %–50 % range , where no universally agreed standard exists, and is applied uniformly across all vegetation types to ensure comparability. To test the influence of the “30 % threshold” choice on our results, we conduct and present a sensitivity analysis for different thresholds. Regional growing-season dynamics that do not align with the calendar year can result in SOS or EOS values outside the range DOY 1–365. In these cases, we add or subtract 365 d, respectively, to obtain transition dates within a single calendar year.

Land surface temperature is obtained from the MODIS MOD11C1 Version 6 product . The data are coarsened to 0.25° spatial resolution by averaging valid pixels and resampled to an 8-daily temporal resolution.

Land cover information is derived from the European Space Agency Climate Change Initiative datasets ESA-CCI;, using annual maps from versions v2.0.7 and v2.1.1. Fractional land cover within each 0.25° grid cell is computed from the relative frequency of classes and aggregated to evergreen needleleaf forest (ENF), deciduous needleleaf forest (DNF), deciduous broadleaf forest (DBF), mixed forest (MF), grassland and shrubland (GRSH), and cropland (CRO). The most frequent class over the study period is used as a static land cover map. Arctic tundra (TUN) is additionally separated using the ESA-CCI CALU dataset .

We use ERA5 reanalysis data at 0.25° resolution to obtain evaporation, sensible heat, shortwave albedo, and skin temperature. Daily totals of evaporation and sensible heat are derived from hourly values and multiplied by -1 to harmonise sign conventions such that positive fluxes indicate transfer from surface to atmosphere. ERA5 evaporation includes transpiration, evaporation from bare soil and water bodies, and evaporation of intercepted water. Sensible heat represents turbulent heat exchange driven by temperature gradients, wind speed, and surface roughness. Shortwave albedo is calculated from ERA5 radiation fluxes; grid cells with albedo values above 0.3 are masked to exclude snow-dominated conditions. To minimise potential circularity, we do not use MODIS-derived albedo, as both MODIS albedo and LAI rely on the same multispectral observations. Skin temperature, representing the theoretical temperature of the uppermost surface layer, is analysed as it responds instantaneously to changes in surface energy fluxes.

Latent heat and sensible heat are additionally obtained from the FLUXCOM product , while evaporation is taken from FLUXCOM-X-Base . As FLUXCOM-X-Base provides only evaporation, sensible heat for EF calculations is taken from the precursor FLUXCOM dataset. Both products upscale eddy covariance measurements using machine learning models driven by meteorological data and satellite-based surface reflectances, land surface temperature, and land cover.

FLUXCOM-X-Base evaporation is aggregated from the original hourly, 0.05° resolution to daily values at 0.25°. The original FLUXCOM LE and H datasets cover the period 2001–2020 only, limiting analyses based on these variables to 20 years.

Evaporation and sensible heat are further obtained from GLEAM4.1a (Global Land Evaporation Amsterdam; ), which integrates observational, modelled, and satellite data at daily and 0.25° resolution. In GLEAM, actual evaporation comprises contributions from bare soil evaporation, open water, snow sublimation, interception loss, and transpiration.

Sensitivities of surface energy balance variables to changes in SOS and EOS are quantified per grid cell as the slope of bivariate linear regressions between annual time series of SOS or EOS (explanatory variable) and spring or autumn mean values of each surface energy balance variable (response variable; Fig. ).

Prior to calculating spring or autumn means, long-term trends in the surface energy balance variables are removed to isolate interannual variability. Specifically, for each grid cell and calendar day, we fit a linear trend across years and subtract this trend from the original daily values. The detrended daily data are then used to compute spring or autumn averages given the defined window.

Spring and autumn windows are defined as ±20 d around the temporal mean SOS and EOS, respectively. For example, average albedo within ±20 d of the mean SOS and EOS is referred to as spring and autumn albedo. The ± 20 d window was chosen to capture the surface energy balance response over the period surrounding the phenological transition, accounting for the fact that biophysical responses may not be instantaneous at the exact transition date, while maintaining a consistent analysis window across years. Regression slopes are estimated from the resulting time series of interannual variability. A negative slope in the albedo–SOS relationship indicates increasing albedo with earlier SOS, whereas a positive slope with EOS indicates higher albedo with later EOS. Statistical significance of regression slopes is assessed using the Wald Chi-Squared test with a significance threshold of p≤0.05.

To account for potential confounding effects of temperature, we additionally perform a partial correlation analysis between SOS/EOS and each surface energy balance variable, using ERA5 surface temperature as a covariate. Partial correlation coefficients and associated p values are computed using Pearson correlation applied to the residuals obtained by regressing both SOS or EOS and the surface energy balance variable on surface temperature.

Grid cells with very low vegetation cover or dominated by urban and barren areas, water bodies, or permanent snow and ice are excluded from both analyses using the static plant functional type (PFT) map derived from ESA-CCI land cover data and by masking grid cells with an annual maximum LAI below 0.5 m³ m⁻³.

Sensitivities and partial correlations derived from the individual datasets are combined into composite maps. For each variable, we assess whether the majority of datasets with statistically significant values (Wald Chi-Squared test; p≤0.05 for sensitivities, corresponding thresholds for partial correlations) agree on the sign of the relationship. Composite maps show the mean slope (or partial correlation coefficient) across datasets that share the same dominant sign, calculated separately for positive and negative relationships. Grid cells are classified as “no trend” (white) when the majority of datasets are non-significant or when there is no convergence in sign. The color bar is divided into discrete bins, each containing an equal number of positive or negative grid cells, with zero serving as the boundary between the two color scales.

To quantify agreement among datasets, we also generate maps of dataset counts. Each slope or partial correlation coefficient is assigned a flag: -1 for negative values, +1 for positive values, and NaN for non-significant values and missing data. These flags are summed across datasets for each grid cell. For example, with three datasets, a count of -3 indicates unanimous agreement on a negative relationship, whereas -1 indicates partial disagreement (e.g., two negative and one positive). A value of 0 arises from opposing signs cancelling each other (e.g., two positive and two negative relationships when four datasets are available) or NaN if the majority of the datasets have no value. The number of contributing datasets varies by variable.

Finally, we present boxplots of sensitivity slopes for each dataset and PFT, including both significant and non-significant values, to illustrate the statistical distributions underlying the composite patterns. Additionally, we assess the robustness of the presented sensitivity slopes by comparing the mean, median, and standard deviation of the reference analysis (30 % LAI threshold, ±20 d window) with those obtained from alternative analysis settings. The alternative settings include a 20 % threshold, a 40 % threshold, a ±15 d window, and a ±25 d window, where only one parameter (threshold or window width) is varied, while the other is kept identical to the reference analysis. The robustness assessment is based only on significant spatial sensitivities of the surface energy balance variables. Therefore, we calculate the average grid-cell slopes for locations where the majority of datasets show statistically significant relationships and agree on the sign of the response. Evaporative fraction is omitted from this analysis because it is derived from latent and sensible heat, which are evaluated separately. Results are presented by PFT for each surface energy balance variable (Fig. ), while the corresponding analysis aggregated across all PFTs (Fig. ), and a comparison of the average slope and standard deviation across all analysis settings with those of the reference (Table ) is provided in Appendix A.

Spring albedo shows widespread significant positive sensitivities to SOS over boreal and Arctic regions. For example, sensitivities reach up to 0.0040.0020.006 (unitless; median with 25th and 75th percentiles) in deciduous needleleaf forests from the ERA5 albedo dataset, indicating that an earlier SOS is associated with declining albedo (Fig. a, b). Strong positive sensitivities also occur in evergreen needleleaf forests and tundra, as well as in high mountain regions such as the European Alps, Pyrenees, Caucasus, Altai, and Rocky Mountains. In contrast, mid-latitude regions in Europe, Central and Eastern Asia, and eastern North America exhibit negative sensitivities, implying increasing albedo with earlier SOS.

Positive sensitivities in tundra and boreal forests are likely driven by spring green-up masking residual snow cover, thereby reducing surface albedo. Conversely, negative sensitivities in croplands and deciduous broadleaf forests likely reflect canopy development shielding darker soils during spring green-up, particularly under moist soil conditions . In agricultural systems, increased albedo may additionally arise from earlier sowing of brighter crops .

Partial correlation analyses confirm these regional sensitivities of spring albedo to SOS, although areas of positive relationships are less spatially continuous and correlation magnitudes are more uniform across regions than the corresponding regression slopes (Fig. a).

Autumn albedo sensitivities to EOS are generally weaker and less spatially coherent than in spring. High-latitude regions, particularly tundra and evergreen needleleaf forests, tend to show negative sensitivities, indicating decreasing albedo with delayed EOS, likely due to prolonged canopy cover over bright soils or early snow. Deciduous broadleaf and mixed forests show on average positive sensitivities, consistent with extended leaf cover over relatively dark soils. Partial correlations reproduce these autumn patterns with similar spatial coverage and trend directions, although correlations at high latitudes are not markedly stronger than at lower latitudes (Fig. b).

As stated above, phenological shifts toward an earlier SOS result in the largest albedo responses in evergreen needleleaf forest, deciduous needleleaf forest, and tundra, with changes of 0.003–0.004 in albedo per day of SOS advance (Fig. b). To give a sense of magnitude in the context of the surface energy balance, we present a first-order estimate, based on the following assumptions: (i) the multi-decadal ERA5-Land average downward shortwave radiation for May–June over northern lands (>30° N) is used as the representative flux value; (ii) this downward flux is treated as approximately time-invariant over the study period; and (iii) the albedo change is assumed to act uniformly across the analysis domain. Using the ERA5-Land average of 257±18 W m⁻² (the uncertainty is the spatially aggregated standard deviation of daily estimates for those months; May–June marks the SOS period, see Fig. a), the albedo change converts to a change in reflected shortwave energy. An albedo decrease of 0.003–0.004 per day thus corresponds to a reduction in reflected shortwave of about 0.77±0.05 to 1.01±0.07 Wm-2d-1 for the regions considered. This additional shortwave absorption at the surface increases available energy potentially fueling turbulent heat fluxes, which we discuss next.

Spring latent heat fluxes exhibit widespread significant negative sensitivities to SOS (Fig. a–c), ranging from -0.68-0.85-0.52 Wm-2d-1 in mixed forests to -0.15-0.35-0.03 Wm-2d-1 in croplands. This indicates enhanced latent heat fluxes with earlier SOS. Sensitivities are negative across all vegetation types and strongest over forests. The four datasets show consistent large-scale patterns across most northern lands.

These results align with increased energy supply through shortwave radiation absorption (Sect. ) and reported evidence that increasing vegetation cover during spring enhances transpiration , primarily regulated by photosynthetic activity via stomatal conductance . Partial correlations support this relationship, showing overall significant negative associations between spring LE and SOS, albeit with reduced spatial coverage compared to the sensitivity analysis (Fig. a, b).

Autumn latent heat generally exhibits positive sensitivities to EOS, reaching up to 0.140.0040.329 Wm-2d-1 in mixed forests (Fig. d–f), indicating increasing LE with delayed EOS. Compared to SOS, EOS sensitivities are weaker, more spatially heterogeneous, and less consistent across datasets. In particular, GLEAM shows regionally negative sensitivities in forested areas. Partial correlations reveal strong associations in regions such as Central North America and eastern Europe and indicate comparable or slightly stronger LE–EOS correlations than LE–SOS correlations (Fig. c, d).

This variability likely reflects constraints imposed by soil moisture availability and seasonal meteorological conditions. While delayed EOS may enhance transpiration, limited water availability towards the end of the growing season can decouple phenology from evaporation . An additional and potentially active confounding factor is that an earlier SOS may progressively deplete soil moisture reserves through extended transpiration and canopy interception over summer, directly suppressing autumn latent heat flux independently of EOS timing. Where soil moisture is limiting at the start of autumn, reduced stomatal conductance and transpiration would dampen the expected positive EOS–LE response, or even reverse it. This is consistent with the weaker and more spatially heterogeneous EOS–LE sensitivities we report compared to SOS, and with the regionally negative EOS–LE sensitivities found in some datasets – particularly GLEAM, where explicit soil moisture constraints on evaporation are represented. The negative sensitivities that occur in some regions support the hypothesis of limited water reserves with an elongation of the growing season , but the location and abundance varies between the different datasets.

Sensitivities of sensible heat to SOS and EOS show no coherent large-scale pattern for northern lands and exhibit strong regional variability and dataset divergence (Fig. ). However, smaller regions do exhibit coherent responses: spring H sensitivities to SOS are significantly negative across parts of northern Eurasia, particularly Siberia (Fig. a). Deciduous needleleaf forests show sensitivities of -0.24-0.6430.005 Wm-2d-1, whereas other vegetation types display weaker and often mixed responses. Differences between datasets are especially pronounced in temperate and boreal Eurasia, with FLUXCOM generally indicating more negative or neutral sensitivities than ERA5 and GLEAM.

Partial correlations reveal little spatially extensive significant association between H and SOS, showing only weak tendencies towards negative correlations at some high latitude regions and positive correlations at some low latitude regions (Fig. a, b), reinforcing the absence of a robust spring signal.

Autumn sensible heat sensitivities to EOS are spatially mixed but show greater agreement between datasets than in spring (Fig. e, f). Central and south-eastern North America consistently exhibit negative sensitivities (e.g. -0.08-0.24-0.02 Wm-2d-1 in croplands). Partial correlations coincide with these regions of significant sensitivities, again mainly showing negative relationships in mid- and lower-latitude regions (Fig. c, d).

Overall, these results show substantial uncertainty in how phenological shifts affect sensible heat. However, in high northern latitudes – particularly for deciduous and evergreen needleleaf forests and tundra (Fig. a, c) – a pattern of increased sensible heat flux with earlier SOS is emerging, mirroring the increase in latent heat flux (see Sect. ). With both turbulent heat fluxes increasing, additional energy is required, consistent with the larger shortwave absorption from reduced albedo (Sect. ). In the next section we examine how the partitioning of turbulent heat fluxes, expressed as the evaporative fraction (EF), responds to phenological shifts.

Generally, sensitivities of EF integrate concurrent changes in LE and H (Fig. ). Spring EF largely mirrors LE patterns, showing widespread negative sensitivities to SOS, indicating increasing EF with earlier SOS (Fig. a). Sensitivities reach -0.002-0.00480.0012 in deciduous needleleaf forests (Fig. c). ERA5 and GLEAM exhibit substantially larger variability than FLUXCOM-derived estimates. Regional positive sensitivities occur in Central Asia and the western United States, but datasets show limited agreement in these areas. Partial correlations reproduce the dominant spatial patterns but with reduced coverage, particularly at high latitudes (Fig. ).

Autumn EF generally increases with delayed EOS in temperate regions (Fig. d), reaching 0.0011-0.00080.0054 in croplands. Agreement among datasets is found across large parts of the Mediterranean, Central and Eastern Asia, and North America. In boreal and Arctic regions, however, EF sensitivities are spatially heterogeneous and datasets frequently disagree, particularly in boreal forests and tundra where ERA5 and GLEAM show much larger magnitudes than FLUXCOM. These discrepancies indicate substantial uncertainty in EF responses to phenological shifts. The potential confounding by SOS-driven soil moisture depletion (see Sect. ) may likewise partially explain the dataset disagreement and spatial heterogeneity in autumn EF sensitivities: where soil moisture is limiting, suppressed transpiration reduces EF independently of EOS timing, masking or reversing the expected positive EOS–EF relationship.

Since rising surface temperature – primarily driven by increased downwelling longwave radiation from greenhouse gases over the long term – is the key driver of phenological changes, one cannot estimate the effect of phenological changes back on surface temperature. In other words, surface temperature both influences and responds to phenology through coupled energy-balance processes, and observational analyses cannot disentangle these feedbacks or provide quantitative causal attribution. Thus, we can only speculate about emerging effects linked to phenology based on observations.

The sensitivity analyses reveal negative associations between spring surface temperature and SOS and predominantly positive associations between autumn temperature and EOS (Fig. ), consistent with earlier SOS and delayed EOS under warmer conditions. Warmer conditions also drive increases in both H and LE; however, the concurrent increases in EF suggest enhanced latent cooling, pointing toward a potential negative phenological feedback on surface warming. The negative partial correlations of LE and EF with SOS (Figs.  and ) – i.e., statistically controlling for the effect of surface temperature – lend support to phenology-driven latent cooling. We therefore hypothesize that phenological shifts may partially offset GHG-induced surface warming through enhanced evaporation.

Overall, the estimated sensitivities are robust to the choice of analysis settings. Across all surface energy balance variables and PFTs, the spatially averaged reference slopes (30 % threshold, ±20 d window) generally fall near the centre of the distributions obtained from the alternative threshold and window width settings, and within one standard deviation of their mean (Figs. , , Table ). The sign of the response is preserved in almost all cases; apparent sign changes occur only for H in MF, where mean sensitivities are close to zero across all settings and therefore do not represent a reversal of a meaningful trend.

The reference setting tends to produce slightly larger absolute slopes than the mean across all alternative settings and is most pronounced for SOS, where the ratio of larger to smaller reference slopes is roughly 8:1. EOS sensitivities show the greatest absolute deviation from the cross-settings mean, though the differences remain within the reported standard deviation.

Among the surface energy balance variables, sensible heat is the most robust: mean slopes and standard deviations are consistent across settings and PFTs. By contrast, autumn surface temperature shows the greatest sensitivity to the choice of analysis setting, with somewhat broader inter-setting distributions, particularly for GRSH and MF.

Among PFTs, mixed forest exhibits the highest sensitivity to the choice of settings, with large inter-setting variability. Croplands, in contrast, show the most compact and consistent distributions across all settings, with comparatively small sensitivities. This pattern that may partly reflect the managed nature of agricultural ecosystems.

Overall, the variability in sensitivities across PFTs and surface energy balance variables substantially exceeds the variability introduced by the tested methodological choices, suggesting that the observed responses primarily reflect underlying ecosystem characteristics rather than the specific choice of threshold or window width.