Phenological fluctuations induce seasonal asymmetric warming in boreal forests

Abstract

Phenological shifts in boreal forests alter land-atmosphere energy exchange, but annual averages mask pronounced seasonal contrasts in the magnitude and pathways of this feedback. Here we show that climate-driven phenological instability generates a systematic warming asymmetry: surface warming from leaf area loss exceeds the cooling from an equivalent gain. During the growing season, additional foliage yields diminishing cooling returns as transpiration approaches its physiological limits, constraining latent heat flux. During the dormant season, the canopy-snow albedo contrast makes shortwave radiation dominant; because this radiative perturbation is converted into surface temperature highly non-linearly, foliage loss still produces a net warming exceeding the cooling from an equivalent gain. Per unit leaf area change, the median asymmetry index indicates warming roughly 2.0-2.3 times stronger than cooling in spring and autumn, and about an order of magnitude stronger in winter, where a very small greening-side sensitivity inflates the ratio; in summer both responses approach the detection limit and the asymmetry is not directionally robust. Phenological instability therefore acts as a seasonally structured positive biophysical feedback in northern ecosystems.

This study uses satellite data and shows that in boreal forests the warming from seasonal leaf area loss outweighs the cooling from an equivalent gain. This asymmetry is driven by transpiration limits in the growing season, and by canopy-snow albedo contrasts in winter.

Introduction

Boreal forests exert a disproportionate influence on the global climate system by regulating surface energy exchange through albedo, evapotranspiration (ET), and aerodynamic roughness^1–3. In these high-latitude ecosystems, land surface temperature (LST) is tightly coupled with the seasonal timing and magnitude of vegetation activity^4,5. The boreal region is warming at nearly four times the global mean rate⁶, and its seasonal dynamics are becoming increasingly unstable. Emerging evidence suggests that these changes cannot be interpreted simply as a linear extension of the growing season^7,8. Instead, intensifying heatwaves, droughts, and abrupt spring cold events^9–11 are reshaping phenological trajectories in ways that are nonlinear, seasonally structured, and at times opposing^12,13. A salient manifestation of this instability is the emergence of bidirectional anomalies in the leaf area index (LAI); within the same seasonal window, some years exhibit foliage expansion whereas others exhibit contraction^14–16. Such intraseasonal deviations perturb the canopy structure and surface energy partitioning, potentially altering the LST. However, the magnitude, direction, and mechanistic pathways of LST responses to phenology-driven leaf area variability remain unresolved.

This gap is consequential for climate assessments. Many biophysical analyses and Earth system projections rely, explicitly or implicitly, on symmetry assumptions, whereby thermal responses to gains and losses in leaf area are treated as equivalent and seasonally invariant^5,17–19. Such representations overlook the state dependence of the high-latitude energy balance. Transitional seasons characterized by strong insolation and snowmelt, metabolically active summers, and snow-dominated autumn–winter periods each establish distinct thermodynamic baselines^20–24. If surface–atmosphere coupling varies across these states, identical phenological perturbations may produce divergent temperature responses. Two questions therefore arise: within a given season, can the surface cooling and surface warming caused by phenological fluctuations truly offset each other? And under equivalent magnitudes of leaf area change, do LST responses remain consistent across seasons? Addressing these uncertainties to determine whether these biophysical feedbacks exhibit directional or seasonal asymmetry is essential for constraining vegetation–climate interactions in a rapidly warming boreal system.

To answer the above questions, we integrated long-term satellite observations with a structural equation modeling (SEM) framework across the boreal forest region. We utilized the seasonal dynamics of leaf area (indicated by the LAI) as a robust proxy for phenological shifts^25,26, allowing us to quantify the independent land surface temperature responses to phenology-driven and stress-driven leaf area variations. Specifically, we identified forest grid cells that exhibited significant changes in leaf area while maintaining consistent land cover types from 2005 to 2020, based on the Moderate Resolution Imaging Spectroradiometer (MODIS) land cover product²⁷ and LAI product²⁸. We then derived the seasonal land surface temperature response for each target grid cell using the 0.05° resolution MOD11C3 v.061 LST product²⁹. To isolate direct biophysical impacts (${\mathrm{\Delta }\mathrm{LST}}_{\mathrm{LAI}}$), the background land surface temperature signal driven by climate variability (${\mathrm{\Delta }\mathrm{LST}}_{\mathrm{CV}}$) was derived from neighboring stable forest grid cells and removed from the target signals. Finally, we employed the SEM framework to quantify the distinct contributing pathways, offering a mechanistic explanation for the observed feedback^30,31.

Results

Asymmetric impact of increasing and decreasing seasonal leaf area on land surface temperature

In spring, the distribution of the asymmetry index (AI) is skewed toward low values, with a median of 0.44 ± 0.02 (Fig. 1a). Across all four seasons the median AI is significantly below the symmetry value of 1 (6° spatial block bootstrap; Supplementary Fig. 26). This finding indicates that although the increase in leaf area caused by earlier phenology is associated with a decrease in land surface temperature (−0.037 ± 0.002 °C; Supplementary Fig. 1a), this cooling magnitude is lower than the temperature increase associated with a decrease in leaf area caused by delayed phenology (0.045 ± 0.004 °C, Supplementary Fig. 1b). Sensitivity analysis reveals that the land surface temperature increase associated with a decrease per unit of leaf area (2.548 ± 0.002 °C per unit of leaf area) is approximately 2.5 times greater than the land surface temperature decrease associated with an increase in equivalent leaf area (−1.008 ± 0.004 °C per unit of leaf area; Supplementary Fig. 2a, b). This pooled-slope ratio is slightly larger than the median-based value (1/AI = 2.27), because pooling across windows gives disproportionate weight to the small number of windows with a near-zero gain-side slope; consistent with the truncation analysis (Supplementary Fig. 25), we quote the median-based value throughout. This pattern of higher warming risk is concentrated in regions with frequent phenological fluctuations, such as Eastern Siberia and northwestern North America (Supplementary Fig. 3a, b).

Fig. 1. Spatial pattern of land surface temperature asymmetry across boreal forests.

a–d Spatial distribution of the asymmetry index (AI) during spring, summer, autumn, and winter (Methods). The AI is defined as the signed ratio of the land surface temperature sensitivity coefficient for an increase in leaf area to that for a decrease in leaf area (Eq. (11)); the value annotated in each histogram inset is the median across moving windows. Color coding illustrates the directional consistency and magnitude of the asymmetry: positive AI values indicate a consistent biophysical feedback direction (i.e., greening-induced cooling and browning-induced warming), where red tones (0 < AI < 1) represent a warming-dominant regime (the warming penalty from leaf area loss exceeds the cooling benefit from leaf area gain) and blue tones (AI > 1) represent a cooling-dominant regime. Gray tones (AI < 0) indicate mechanistic anomalies, where land surface temperature and leaf area changes are positively coupled (e.g., greening-induced warming and browning-induced cooling). A value of 1.0 represents a perfectly symmetric, linear response. Circle size denotes the magnitude of the change in leaf area (ΔLAI). The histogram insets display the moving-window frequency (%) distribution of AI values; the dashed red line and annotated number indicate the median, and the gray curve shows a smoothed density estimate. Only moving windows with ≥ 30 matched grid pairs in both the increase and decrease categories are shown; windows with fewer than 30 pairs are excluded from display and from domain-level summary statistics.

This asymmetry weakens markedly in the peak growing season (summer; Fig. 1b). The absolute impact of leaf area change on land surface temperature is small (cooling of −0.005 ± 0.002 °C for a leaf area increase and warming of 0.010 ± 0.003 °C for a decrease; Supplementary Fig. 1c, d). The per-unit sensitivity ratio still favors warming (median AI = 0.51 ± 0.04; Fig. 1b); however, because both responses are close to the noise level, the direction of the summer asymmetry is not robust at the window scale: the actual warming and cooling magnitudes are comparable (response-magnitude ratio ≈ 0.5; Supplementary Fig. 28), and only a minority of windows show the expected greening-cooling/browning-warming pattern (Supplementary Fig. 29). Summer is therefore best characterized as near-symmetric rather than warming-dominant.

In the dormant season, although delayed autumn phenology leads to increased leaf area, its mitigating effect on surface warming is minimal (the cooling sensitivity is −0.155 ± 0.01 °C per unit leaf area, whereas the warming sensitivity is 0.274 ± 0.005 °C per unit leaf area; Supplementary Fig. 2e, f), and the autumn asymmetry index remains well below symmetry (median AI = 0.51 ± 0.03; Fig. 1c).

The per-unit asymmetry index reaches its lowest value in winter (median AI = 0.07 ± 0.01; Fig. 1d), because the greening-side sensitivity approaches zero (−0.109 ± 0.01 °C per unit leaf area) while the browning-side sensitivity remains finite (0.419 ± 0.09 °C per unit leaf area; Supplementary Fig. 2g, h). This very small denominator, however, inflates the sensitivity ratio: the corresponding difference in the actual temperature response is modest (response-magnitude ratio ≈ 1.6; Supplementary Fig. 28), and roughly one-quarter of winter windows show a positive LST–LAI coupling (AI < 0; Supplementary Fig. 24), reflecting the weak and variable winter signal rather than a proportionally extreme warming effect. The domain-median AI nonetheless remains well below the symmetry line in every season (Fig. 1).

Asymmetry across different magnitudes of leaf area change

To capture the dynamic response of land surface temperature to varying intensities of phenological shifts, we performed a binned analysis of changes in leaf area (ΔLAI) (Fig. 2). Bins are defined on the relative change rate ΔLAI_rate = ΔLAI/LAI_baseline (dimensionless; Methods), which normalizes the perturbation by the local baseline canopy state. In spring, land surface temperature response scales monotonically with the magnitude of the phenological fluctuation, without the mid-range peak seen in the other seasons. The land surface warming associated with leaf area loss increased from 0.066 ± 0.001 °C at low magnitudes (ΔLAI_rate < 0.2) to 3.018 ± 0.374 °C in the highest populated spring bin (0.6 ≤ ΔLAI_rate < 0.8; Fig. 2a). In contrast, the land surface cooling associated with an equivalent leaf area increase reached only −1.459 ± 0.184 °C in the same magnitude bin and was significantly weaker than the corresponding warming response. Within this bin, daytime warming (3.515 ± 0.465 °C) significantly exceeded nighttime warming (2.522 ± 0.284 °C; Supplementary Fig. 7a).

Fig. 2. Magnitude-dependent land surface temperature sensitivity and asymmetry across different intensities of change in leaf area.

a–d Comparison of daily mean land surface temperature changes (ΔLST, °C) induced by increases in leaf area (blue bars) versus decreases (red bars) across five bins of relative leaf-area change (ΔLAI_rate: <0.2, 0.2–0.4, 0.4–0.6, 0.6–0.8 and ≥0.8) for spring, summer, autumn, and winter. The bars represent domain means, error bars indicate ± 1 standard error of the spatial mean (Methods), and the overlaid points show the corresponding 6° spatial-block values. Asterisks denote significant differences between the increase and decrease scenarios, assessed using two-sided spatial-block bootstrap tests (***P < 0.001, **P < 0.01, *P < 0.05). e Distribution of the signed AI across the same magnitude bins for all four seasons. Data are presented as raincloud plots combining violin plots (probability density), box plots (center line, median; box limits, upper and lower quartiles; whiskers, 1.5× interquartile range) and scatter points (individual moving-window values). The colors represent the seasons: spring (blue), summer (brown), autumn (yellow), and winter (orange). The horizontal dashed red line at AI = 1.0 denotes a symmetric thermal response in which warming and cooling effects are balanced. Panels a–d are based on 2.7 × 10⁵ (spring), 5.1 × 10⁵ (summer), 2.2 × 10⁶ (autumn) and 7.0 × 10⁵ (winter) matched grid pairs; panel e is based on n = 367, 665, 1470 and 671 moving windows, respectively.

In contrast, the other seasons exhibit distinct nonlinear patterns. Unlike the monotonic trend observed in spring, the land surface temperature variations in summer peak at intermediate magnitudes (0.4−0.6), where the warming magnitude (0.523 ± 0.076 °C) exceeds the cooling magnitude (−0.305 ± 0.121 °C; Fig. 2b), before both decrease considerably at high magnitudes. Diurnal analysis indicates that the cooling that accompanies leaf area gain is concentrated during the day (peak of −0.762 ± 0.046 °C), significantly exceeding the nighttime magnitude (Supplementary Fig. 7b). Autumn displays similar nonlinear dynamics, with cooling magnitudes appearing to plateau; it also peaks at intermediate magnitudes (0.4 − 0.6), with 0.158 ± 0.021 °C warming versus −0.092 ± 0.011 °C cooling (Fig. 2c). Winter presents a complex nonlinear trend, peaking at intermediate magnitudes (warming 0.295 ± 0.017 °C vs. −0.189 ± 0.007 °C; Fig. 2d) but exhibiting a notable inversion at extremely high magnitudes (ΔLAI_rate ≥ 0.8) where leaf area loss is associated with slight cooling (−0.066 ± 0.007 °C) and leaf area gain is associated with slight warming (0.078 ± 0.013 °C).

These divergent land surface temperature response patterns drive the significant evolution of the asymmetry index (AI) across disturbance intensities (Fig. 2e). At low magnitude (ΔLAI_rate < 0.2), the AI values diverge seasonally: spring is characterized by a cooling-dominant feature (AI = 1.012 ± 0.058), whereas winter is already suppressed to extremely low levels (AI = 0.021 ± 0.004), indicating inherent warming dominance. However, as phenological fluctuations intensify, this seasonal divergence rapidly leads to homogenization. In the medium–low magnitude interval (0.2−0.4), the AI value for spring decreases significantly to 0.483 ± 0.034. This trajectory, from divergence at low intensities to convergence at high intensities, demonstrates that the biophysical cooling advantage of the growing season is restricted to mild disturbances. Under strong phenological perturbations, the growing-season cooling advantage disappears and spring and autumn converge toward a warming-amplified state; the summer and winter responses, by contrast, remain weak and directionally mixed even at large magnitudes (Supplementary Fig. 29).

Mechanisms underlying the seasonal land surface temperature response

To elucidate the biophysical mechanisms underlying the observed magnitude-dependent asymmetry, we integrate the statistical characteristics of the surface energy balance with the directional associations inferred from SEM. Our analysis reveals that the basis of the asymmetric surface temperature response lies in a seasonal shift in dominant biophysical processes (Fig. 3): a transition from physiological regulation associated with transpiration during the growing season to physical radiative control dominated by albedo feedback during the dormant season. The corresponding seasonal SEM path structures are shown in Fig. 4.

Fig. 3. Biophysical mechanisms and energy-flux pathways regulating seasonal land surface temperature responses.

Conceptual diagram summarizing the seasonal reorganization of mechanisms linking leaf-area changes to land surface temperature responses. The left and right halves summarize growing-season physiological constraints (spring and summer) and dormant-season physical control (autumn and winter), respectively. In summer, the nonradiative sensible-plus-ground-heat pathway (H + G) becomes dominant, as described in the main text, even though physiological saturation constrains the response. The blue and orange curves qualitatively indicate the declining relative importance of latent heat and the increasing relative importance of shortwave radiation through the seasonal cycle. Blue upward arrows denote latent-heat flux, and orange V-shaped arrows denote shortwave-radiation effects; their number and width are qualitative and are not scaled to pathway contributions. Red thermometers and upward arrows denote warming associated with leaf-area decrease, whereas blue thermometers and downward arrows denote cooling associated with leaf-area increase. The displayed values are the corresponding domain-mean land surface temperature changes (ΔLST, °C).

Fig. 4. Decoupling of seasonal energy pathways via structural equation modeling.

a–d Path diagrams representing the SEM for spring, summer, autumn, and winter. The models separate the biophysical regulation of land surface temperature (LST) by leaf area index (LAI) into three pathways: latent heat flux (LE; physiological), reflected shortwave radiation (SW; the LAI-induced change defined in Eq. (15), for which negative values denote reduced reflection and increased shortwave absorption), and the nonradiative sensible-plus-ground-heat pathway (H + G). Blue paths represent LAI-increase scenarios and red paths represent LAI-decrease scenarios. Solid and dashed rendering separates overlapping paths from the two directional models, and pale curved arrows show the direct LAI-to-LST pathways. Arrow width is proportional to the absolute magnitude of the displayed standardized coefficient. Positive coefficients indicate that an increase in the source variable is associated with an increase in the target variable, whereas negative coefficients indicate an inverse association. Asterisks denote two-sided Wald z tests of the path coefficients (***P < 0.001, **P < 0.01, *P < 0.05). Panel a shows spring, dominated by evaporative cooling through LAI → LE → LST. Panel b shows summer, in which the nonradiative H + G pathway is strongest while evaporative cooling persists. Panel c shows autumn, with generally weaker transitional pathways. Panel d shows winter, which shifts to a radiative-dominant regime driven by the LAI → SW → LST pathway, with minimal physiological contribution. Model-fit statistics are detailed in Supplementary Table 1.

In the early growing season (spring), energy partitioning is linked to biological processes. SEM analysis reveals a strong positive correlation between spring leaf area and latent heat flux (LE) (coefficient of 0.156 ± 0.012; Fig. 4a). The statistical results reveal that under typical leaf area increase scenarios (ΔLAI_rate > 0.2), the latent heat flux increases by 0.440 ± 0.012 W m⁻² (Supplementary Fig. 9a). The contribution of evaporation is quantified via contribution analysis (Fig. 3 and Supplementary Fig. 14): the evaporation component accounts for 66.9% of the land surface temperature variation in spring, far exceeding the combined contributions of solar radiation (~11%) and sensible heat. This observed asymmetry corresponds to a quantitative difference in the flux response: while the increase in leaf area is associated with substantial cooling via transpiration, this statistical link becomes weaker as the leaf area decreases, when the surface is more exposed to direct radiative heating.

As the ecosystem progresses into the peak growing season (summer), the dominant process shifts to the unresolved nonradiative pathway (mainly sensible and ground heat, H + G). Unlike spring, the summer canopy is dense, and the SEM indicates that land surface temperature sensitivity is highly responsive to the nonradiative H + G component. In scenarios in which the leaf area decreases, the contribution of the nonradiative H + G component increases to 53.7%, surpassing the contribution of the evaporation component (34.1%) (Supplementary Fig. 14). The data reveal a nonlinear response pattern: further increases in leaf area are associated with limited cooling gains via turbulent fluxes. Conversely, a decrease in leaf area is accompanied by a disproportionate increase in the H + G component. This explains why the warming associated with leaf area loss is significant, whereas the cooling efficiency associated with leaf area gain tends to plateau.

In the dormant season, the mechanism shifts predominantly toward physical processes governed by albedo feedback. This transition begins in autumn, as the response of latent heat flux decays and its contribution to temperature drops to negligible levels (<2%, Supplementary Fig. 14). This physical regulation culminates in winter, where thermal dynamics are uniquely dominated by the stark contrast between a dark canopy and bright snow cover. During this extreme winter phase, the role of shortwave radiation (SW) increases sharply. Here SW↑ denotes the LAI-induced change in reflected (upward) shortwave radiation (Eq. (15)), so that a negative value indicates reduced reflection and hence increased shortwave absorption. SEM analysis reveals that winter leaf area is strongly correlated with this reflected shortwave term (the standardized coefficient is −0.145 ± 0.011 for the increase in leaf area and 0.152 ± 0.005 for the decrease in leaf area; Fig. 4d), consistent with a canopy that masks the underlying snow: densification reduces the reflected flux (SW ↑ = −0.761 ± 0.022 W/m²), whereas thinning increases it (SW ↑ = 0.485 ± 0.022 W/m²) by exposing snow. The shortwave component accounts for ~94% of the total pathway contribution to the land surface temperature response (Fig. 3 and Supplementary Fig. 14); because contribution rates are computed on absolute indirect effects (Eq. (26)), this figure quantifies the relative magnitude of the shortwave pathway and does not, by itself, indicate the direction of the temperature response. We note, however, that our albedo-based SW (Eq. (15)) cannot fully separate LAI-driven albedo change from co-varying snow-cover variability, so the winter shortwave attribution should be read as an upper bound on the LAI-specific effect. The efficiency with which these radiative perturbations are translated into actual surface temperature changes is highly non-linear and direction-dependent, so the sign of the net temperature response does not follow directly from the sign of the shortwave anomaly. Because the loss of canopy buffering during leaf area decline alters the local thermodynamic baseline by removing canopy insulation, weakening the coupling between the surface and the overlying air, and exposing the surface to stronger radiative and turbulent exchange, the resulting per-unit warming penalty continues to exceed the cooling benefit of an equivalent canopy gain, even though the shortwave anomaly itself acts in the opposite direction.

Discussion

Quantifying the biophysical impacts of vegetation dynamics is essential for resolving climate–vegetation feedback. Previous research has documented asymmetric land surface temperature responses to forest cover gain and loss at annual scales^22,23. However, even structurally stable boreal forests are dynamically regulated by climate-driven phenological fluctuations. By resolving these internal dynamics seasonally, our results demonstrate that the biophysical regulation of land surface temperature (LST) differs fundamentally across the annual cycle. Annual-mean asymmetries therefore represent the net outcome of distinct, and at times opposing, energy pathways operating in spring, summer, autumn and winter. A mechanistic understanding of boreal climate regulation thus requires explicit consideration of seasonal leaf area dynamics rather than reliance on aggregated annual metrics.

A central limitation in existing biophysical assessments is the implicit assumption that vegetation-induced temperature responses arise from a uniform mechanism throughout the year. Many evaluations emphasize the annual cooling potential of greening or restoration^5,16, while underrepresenting seasonal shifts in dominant constraints. Our findings reveal a clear seasonal transition in governing controls. During the growing season, the regulation of the LST is primarily constrained by physiological saturation. Mature boreal canopies frequently operate near maximum conductance thresholds^7,15,32, such that additional leaf area yields diminishing increases in latent heat flux. Cooling efficiency therefore saturates, whereas a reduction in leaf area weakens evaporative buffering and induces disproportionately strong surface warming. In contrast, during the dormant season the surface energy balance is governed predominantly by the canopy-snow albedo contrast, which makes shortwave radiation the dominant pathway by magnitude. The temperature consequence, however, is not proportional to the radiative perturbation: the conversion of a given shortwave anomaly into a surface temperature change is markedly asymmetric between the gain and loss directions, so that foliage loss nevertheless yields a net warming that exceeds the cooling associated with an equivalent gain^5,14. This seasonal alternation between physiological saturation and radiative amplification provides a mechanistic basis for the observed asymmetric temperature response.

The MODIS LAI product used here has been validated over needleleaf and boreal forests, with an overall ground-validation RMSE of ~0.66 LAI units (improved from 0.80 in the previous collection) and R² ≈ 0.77³³; agreement approaches the GCOS target accuracy over sparsely vegetated surfaces but is lower over dense and high-latitude needleleaf canopies. Retrieval quality in these forests is strongly seasonal: the main-algorithm retrieval index falls from ~90% in boreal summer to below 30% in boreal winter, because solar-zenith angles exceeding 52.5° together with snow and cloud contamination force the lower-quality backup algorithm³⁴. Two features of our design limit the impact of this uncertainty on our results. First, we use only main-algorithm, quality-screened retrievals and form seasonal composites only where at least 50% of observations pass quality filters, which excludes most low-illumination and high-zenith retrievals. Second, our analysis is differential: it compares ΔLAI between adjacent years within the same pixel and season, with a paired climate-background removal, so slowly varying retrieval biases, whose temporal stability has been demonstrated over the full MODIS record²⁶, largely cancel rather than propagating into the reported sensitivities. Winter remains the most uncertain season, and we interpret the winter results qualitatively.

Our results have implications for both climate modeling and management. Although the latest Earth System Models (ESMs) have significantly advanced in representing vegetation phenology, capturing the high-frequency, non-linear thermal coupling between canopy structure and turbulent fluxes remains an ongoing challenge^17,18. Our purely observation-based findings provide a highly resolved empirical benchmark. Integrating these seasonally varying, intensity-dependent biophysical sensitivities could offer valuable pathways for further refining high-latitude feedback parameterizations in future model developments. From a management perspective, the identified asymmetry indicates that cooling gains from additional leaf areas are physiologically bounded, whereas warming from canopy degradation can be immediate and amplified. Under such constraints, conserving a mature canopy structure may provide a more reliable biophysical buffer against regional warming than strategies that rely solely on incremental increases in leaf area. In boreal regions, the climatic effectiveness of new plantations should thus be evaluated in light of diminishing cooling returns, with priority given to maintaining existing forest integrity. Beyond diagnosing this challenge for Earth-system models, the season- and direction-resolved biophysical sensitivities derived here (Fig. 1 and Supplementary Fig. 2) provide observation-based constraints that could be used to benchmark or parameterize the vegetation–temperature coupling in land-surface models.

Several aspects of our analysis warrant caution. First, the asymmetry index is a ratio of directional sensitivities and is heavy-tailed when the greening-side slope approaches zero, particularly in winter; we therefore summarize it by the median across moving windows and confirm that it is insensitive to the outlier-truncation threshold (Supplementary Fig. 25) and that its confidence intervals, obtained from a 6° spatial block bootstrap that accounts for spatial autocorrelation and pseudo-replication among overlapping windows, exclude 1 in every season (Supplementary Fig. 26). Second, the directional pattern of cooling under greening and warming under browning is most robust in spring and, to a lesser degree, autumn; in summer and winter the per-pixel temperature responses are small and the direction is more variable (Supplementary Fig. 29), so the seasonal conclusions for these two seasons should be read qualitatively. Third, the seasonal boundaries are climatological means carrying an ± 8–15-day spread (Methods), and a boundary shift of this order affects only the small number of low-signal, near-transition pixels; the median-based AI is correspondingly robust to such shifts. Finally, the domain is delineated by the Köppen–Geiger cold-climate zone and therefore includes both subarctic and some cold-temperate continental forests, which contributes to a minority of sign-anomalous (AI < 0) windows along the southern margin.

Methods

Seasonal demarcation and detection of phenological shifts in the boreal zone

To accurately quantify seasonal feedback, we established a robust phenological framework specific to boreal ecosystems. We developed a multi-method ensemble detection algorithm based on 448 site-years (1991–2014) (Supplementary Table 2) of FLUXNET eddy covariance observations³⁵. Unlike static calendar definitions, this algorithm identifies seasonal transitions by integrating independent physiological and meteorological drivers (Supplementary Fig. 15). Specifically, spring onset was determined by synthesizing temperature thresholds (≥3 days with mean Ta > 5 °C within a 5-day window)³⁶, carbon flux transitions (NEE shifting to net sink)³⁷, and growing degree-day accumulation³⁸. Summer and autumn transitions were identified using gross primary productivity (GPP) inflection points and temperature gradients^39,40, whereas winter onset was defined on the basis of frost frequency and metabolic dormancy signals⁴¹. The results from these diverse methods were fused using a confidence-weighted averaging scheme to minimize detection uncertainty.

To minimize uncertainty arising from individual methods, we developed a confidence-weighted integration scheme. First, we calculated the standard deviation (${\sigma }_{\mathrm{DOY}}$) of the detected dates across all n successful methods to quantify divergence:

$${\sigma }_{\mathrm{DOY}}=\sqrt{\left(\frac{1}{n-1}\right){\sum }_{i=1}^n{\left(\mathrm{DO}{Y}_i-\overline{\mathrm{DOY}}\right)}^2}$$ 1

where $\mathrm{DO}{Y}_i$ denotes the day of year detected by the i-th method, $\overline{\mathrm{DOY}}$ represents the mean value across all successfully detected methods, and $n$ is the number of successful detections.

$$C_{\mathrm{consistency}}=\exp \left(-\frac{{\sigma }_{\mathrm{DOY}}}{10}\right)$$ 2

$$\mathrm{D}{\mathrm{O}Y}_{\mathrm{ensemble}}=\frac{\sum {\left(i=1\right)}^n{w}_i\times DO{Y}_i}{\sum {\left(i=1\right)}^n{w}_i}$$ 3

where $w_i$ represents the confidence weight of detection method i, defined as its per-method confidence score $C_i$ (0–1), which quantifies how unambiguously that method’s driver crossed its detection criterion (e.g., the margin above the temperature/GDD threshold or the sharpness of the GPP/NEE inflection); methods that fail to detect a transition are assigned w_i = 0 and excluded. The weights are used un-normalized in Eq. (3) (the denominator ${\sum w}_i$ performs the normalization).

$$\mathrm{Confidenc}{\mathrm{e}}_{\mathrm{final}}=\overline{\mathrm{Confidence}}\times C_{\mathrm{consistency}}$$ 4

where $\overline{\mathrm{Confidence}}$ is the average confidence score across the methods.

The uncertainty range for each transition was estimated as follows:

$$\mathrm{Uncertaint}{\mathrm{y}}_{\mathrm{range}}=\left[\min \left(\mathrm{DO}{Y}_i\right)-5,\max \left(\mathrm{DO}{Y}_i\right)+5\right]$$ 5

The algorithm demonstrated high robustness, achieving detection success rates >98% across all seasons (Supplementary Fig. 17) with high confidence scores (Supplementary Fig. 18). Leave-one-out cross-validation confirmed its accuracy, yielding a root mean square error (RMSE) of only 5–8 days (Supplementary Fig. 19). Based on these detections, we determined the climatological mean transition dates: spring onset at DOY 99 ± 15 (early April), summer at DOY 165 ± 8 (mid-June), autumn at DOY 253 ± 12 (early September), and winter at DOY 313 ± 11 (early November) (Supplementary Fig. 20); the corresponding coefficients of variation are given in Supplementary Fig. 21. These dates align closely with independent global seasonal classifications⁴², and the resulting mean growing season length of 153 ± 20 days (Supplementary Fig. 22) establishes a precise temporal baseline for our subsequent biophysical analysis.

Calculation of the variation in leaf area

Following the establishment of the seasonal boundaries, we leveraged the Google Earth Engine (GEE) platform to process the MODIS MOD15A2H data²⁸ spanning the study period (2005–2020) to generate composite images representative of distinct seasons. To isolate stable forest areas, a persistent forest mask was created using the MODIS MCD12Q1 land cover type product²⁷. This mask strictly retained pixels that were consistently classified as forest annually throughout the study period, ensuring spatiotemporal consistency in vegetation type. The analysis was restricted to the Köppen–Geiger cold-climate zone of the Northern Hemisphere⁴³, delineated in ArcGIS and hereafter referred to as boreal forest following recent usage⁴⁴; this zone spans subarctic and cold-temperate continental forests.

For each season, all available MODIS LAI images within the corresponding annual and monthly seasonal ranges were initially extracted. A temporal averaging composite was then applied to the filtered image collection to generate a single, representative cloud-minimized LAI composite characterizing the overall condition of that season. These composite images were chronologically labeled as LAI_baseline and LAI_end (e.g., Spring 2005 as LAI_baseline and Spring 2006 as LAI_end), representing the mean LAI values for the corresponding season in the starting and ending years, respectively. Finally, both the absolute change in LAI (ΔLAI) and the relative change rate (ΔLAI_rate) were computed at the pixel level using the following formulas:

$$\mathrm{\Delta }\mathrm{LAI}={\mathrm{LAI}}_{\mathrm{end}}-{\mathrm{LAI}}_{\mathrm{baseline}}$$ 6

$${\mathrm{\Delta }\mathrm{LAI}}_{\mathrm{rate}}=\frac{\mathrm{\Delta }\mathrm{LAI}}{{\mathrm{LAI}}_{\mathrm{baseline}}}$$ 7

where LAI_baseline and LAI_end represent the mean LAI values for the corresponding season in the starting and ending years, respectively.

To mitigate the influence of ceiling and floor effects on the analysis of LAI dynamics, an adaptive filtering strategy based on quartiles was implemented. Specifically, to assess the potential increase in LAI, pixels that satisfied LAI_baseline ≤ Q₂₅ (the 25^th percentile) and exhibited a positive rate of change in LAI (ΔLAI > 0) were selected to ensure sufficient upward potential for pixels with low initial values. Conversely, to evaluate potential decreases in LAI, pixels meeting LAI_baseline ≥ Q₇₅ (the 75^th percentile) with a negative rate of change in LAI (ΔLAI < 0) were identified to guarantee adequate downward capacity for pixels with high initial values. This data quality control procedure effectively enhanced the reliability and robustness of the subsequent biophysical effect analyses.

Based on the relative change rate (ΔLAI_rate), each pixel was subsequently categorized into one of three LAI change types: pixels with ΔLAI_rate > 15% were classified as increased LAI; those with ΔLAI_rate < −15% were assigned to decreased LAI; and the remainder were designated as no significant change. Crucially, within the context of stable forest cover, these categories served as direct proxies for phenological timing: an increase in leaf area in spring signified phenological advancement (earlier green-up), whereas in autumn, it represented phenological delay (extended growing season). Conversely, a decrease in leaf area reflected phenological contraction (e.g., spring phenological events were delayed, or autumn phenological events occurred earlier). This classification result was ultimately exported as a separate band within the image to facilitate subsequent spatial statistical analysis.

Estimating the biophysical effects on land surface temperature induced by changes in leaf area

To disentangle the direct biophysical temperature variations driven by phenologically-induced vegetation anomalies (ΔLST_LAI) from broad-scale climate variability (ΔLST_CV), we utilized MOD11C3 v.061 Land surface temperature (LST) products (0.05° spatial resolution), which captured the skin temperature of Earth’s surface. The daily mean LSTs were derived by averaging the daytime and nighttime observations^29,45,46. Crucially, to ensure that the observed LAI dynamics reflect genuine phenological shifts (e.g., interannual variations in spring onset or autumn senescence) within stable ecosystems rather than artifacts of land cover conversion (e.g., deforestation or urbanization), we applied a strict persistent vegetation mask during preprocessing. Using the MODIS land cover product (MCD12Q1)²⁷, we retained only those pixels that maintained an identical vegetation classification code consistently throughout the study period (2005–2020). Pixels exhibiting land cover transitions were rigorously excluded. This quality control step guaranteed that the substantial LAI fluctuations analyzed in this study represented reversible phenological variability within a consistent forest type, thereby ruling out land-use change as a confounding factor.

A paired-grid signal separation approach, combined with a multilevel spatial comparative analysis framework, was employed to strictly exclude the interference of environmental heterogeneity. This analytical framework was operated within a two-tiered grid system, consisting of 50 km parent grids and nested 5 km child grids (Supplementary Fig. 23). We first identified 50 km grids that contained coexisting pixels of all three LAI change types (phenological expansion/increase, phenological contraction/decrease, and stable) within their boundaries. Crucially, to rule out confounding effects from slow-varying structural changes such as tree height growth or stand aging, the pairing of target and reference grids was strictly restricted to data from temporally adjacent years. These grids were designated valid analysis units. This coexistence criterion and temporal adjacency were fundamental; they ensured that the comparative analysis was restricted to a localized domain with a homogeneous climatic background and comparable vegetation structural maturity¹⁹.

Within each valid unit, 5 km child grids were classified based on the dominant trajectory of their constituent pixels. A strict 70% purity threshold was applied: a child grid was classified as an increased grid only if more than 70% of its valid pixels belonged to the increased category (ΔLAI_rate > 15%, serving as a proxy for phenological advancement/delay). Similarly, decreased grids and stable grids were identified based on the >70% dominance of their respective pixel types. Increased and decreased grids served as the target experimental units, whereas stable grids functioned as the local phenological reference units. By selecting reference units exclusively from undisturbed 5 km grids within the same 50 km radius, we ensured that the control group shared nearly identical environmental attributes with the target group²². Grids straddling the domain boundary or containing mixed land cover were excluded by the persistent-forest mask and the >70% purity threshold, so that every analysis grid represents a single, temporally stable forest type.

The core methodological innovation lies in the reconstruction of a counterfactual climate baseline (i.e., a hypothetical climate reference state) using a locally adaptive inverse distance weighting (IDW) technique⁴⁷. The fundamental premise is that stable grids act as thermometers for the regional climate, recording temperature variations driven solely by atmospheric forcing (ΔLST_CV) without significant phenological feedback. Specifically, for each target grid $i$, the climate background ΔLST_CV was calculated as the distance-weighted average of all stable grids $k$ within the 50 km radius, as shown in Eq. (8):

$${\mathrm{\Delta }\mathrm{LST}}_{\mathrm{CV}}=\frac{{\sum }_{k=1}^n\frac{{\mathrm{\Delta }\mathrm{LST}}_k}{d_k}}{{\sum }_{k=1}^n\frac{1}{d_k}}$$ 8

where ΔLST_k is the observed change in stable grid $k$ and $d_k$ is the Euclidean distance between grid centers. A lower bound of d_min = 500 m is imposed in Eq. (8) purely to avoid a singularity when a query point coincides with a data point; because the analysis is performed on a 5 km grid, all pairwise distances between grid centers exceed this bound and d_min has no practical effect on the interpolated values. The pure land surface temperature response (ΔLST_LAI) was subsequently isolated by subtracting this reconstructed climate baseline from the total observed land surface temperature change, as shown in Eq. (9):

$${\mathrm{\Delta }\mathrm{LST}}_{\mathrm{LAI}}={\mathrm{\Delta }\mathrm{LST}}_{\mathrm{total}}-{\mathrm{\Delta }\mathrm{LST}}_{\mathrm{CV}}$$ 9

This subtraction functioned as an environmental normalization: by removing the background signal derived from local neighbors (which implicitly encode local elevation and soil properties), we filtered out static environmental confounders, rendering spatially distinct pixels that were scientifically comparable.

To quantify asymmetric responses, we developed a systematic grid-matching algorithm that paired spatially distinct 5 km grids that exhibited opposite LAI trends within the same season and climate zone. To control for the nonlinear scaling of temperature responses, matched pairs were strictly constrained to have similar perturbation magnitudes (||ΔLAI_rate, _increase|-|ΔLAI_rate, _decrease|| < 0.2 (dimensionless)). During the matching process, the rate of change in the LAI (ΔLAI_rate = ΔLAI/LAI_baseline) was categorized into five standardized intervals: [0.0, 0.2), [0.2, 0.4), [0.4, 0.6), [0.6, 0.8), and [0.8, +∞). This ensured that all matched grid pairs fell within the same change magnitude category, thereby controlling for the nonlinear influence of the change in the LAI on the land surface temperature response⁴⁸.

Following the IDW decomposition and isolation of ΔLST_LAI values, we employed a moving window approach to investigate the spatial heterogeneity of biophysical effects and their sensitivity to changes in leaf area across different geographic and climatic contexts⁴⁹. Specifically, a 6° × 6° analysis window was applied, moving systematically across the study area at 2° intervals in terms of both latitude and longitude. This spatial scale was selected through preliminary analysis to ensure (i) sufficient sample points within each window for robust statistical inference (minimum n = 30 grid pairs) and (ii) effective capture of regional-scale spatial variability without excessive smoothing of local patterns.

Within each moving window, we performed separate analyses for leaf area increase and leaf area decrease scenarios to quantify directional asymmetry. Huber robust regression analysis⁵⁰, which was chosen specifically to mitigate the influence of statistical outliers while maintaining efficiency for normally distributed data, was applied to establish the quantitative relationship between the rate of change in the leaf area and the corresponding land surface temperature variation as Eq. (10):

$${\mathrm{\Delta }\mathrm{LST}}_{\mathrm{LAI}}={\beta }_0+{\beta }_1\times {\mathrm{\Delta }\mathrm{LAI}}_{\mathrm{rate}}+\epsilon$$ 10

where β₁ represents the sensitivity coefficient, that is, the land surface temperature response per unit relative change in leaf area. Because ΔLAI_rate is dimensionless, β₁ has units of °C. The Huber M-estimator was used to down-weight observations with large residuals (|ε| > 1.345σ) using an iteratively reweighted least squares algorithm⁵¹, providing robust parameter estimates even when 5–10% of the data points were outliers. Statistical significance was evaluated using a spatial block bootstrap (10,000 iterations)⁵¹ in which 6° spatial blocks were resampled to preserve within-block dependence, thereby accounting for spatial autocorrelation and for pseudo-replication among overlapping moving windows (Supplementary Fig. 26).

To systematically evaluate the biophysical feedback, we calculated the asymmetry index (AI) for each moving window and seasonal-climate zone combination. While our broader analysis compared cross-seasonal differences, the AI served as a critical diagnostic metric for the mechanistic linearity and reversibility of feedback within a given season. Following previous analyses of asymmetric biophysical responses to forest-cover change^22,23, the AI was defined as the signed ratio of the land surface temperature sensitivities to leaf-area gain and loss (Eq. (11)), and for each season and climate zone it was summarized as the median across all moving windows. The signed formulation preserves the direction of the LST–LAI coupling (so that anomalous, positively coupled windows appear as AI < 0), and the median is used because the ratio is heavy-tailed and unstable when the greening-side slope approaches zero (Supplementary Figs. 24, 25). Unless otherwise stated, uncertainties are reported as X ± 1 standard error (s.e.). For spatial-mean quantities (the domain-mean temperature and flux changes in Fig. 2, Supplementary Figs. 1 and 7–9), this is the standard error of the spatial mean; for regression sensitivities and SEM path coefficients it is the standard error of the estimated coefficient. Phenological transition dates and growing-season length are reported as X ± 1 standard deviation (s.d.) of the climatological distribution. Because the paired-difference fields are spatially autocorrelated, naive standard errors underestimate the true sampling uncertainty; we therefore base all significance testing, and the uncertainty of the asymmetry index (AI), on a 6° spatial block bootstrap (Supplementary Fig. 26), and the ± values reported for the AI are its block-bootstrap standard errors.

$$\mathrm{AI}=\frac{{\beta }_{1,\mathrm{increase}}}{{\beta }_{2,\mathrm{decrease}}}$$ 11

where ${\beta }_{1,\mathrm{increase}}$ and ${\beta }_{2,\mathrm{decrease}}$ represent the sensitivity coefficients (slopes) from the Huber regression analysis for the LAI-increase (phenological advancement/delay) and LAI-decrease (phenological contraction) grids, respectively. An AI value greater than 1 indicates cooling-dominated asymmetry, where phenological expansion yields disproportionately strong cooling benefits. Conversely, an AI < 1 indicates a warming-dominated asymmetry or nonlinear regime. An AI = 1 indicates a symmetric, linear feedback mechanism, suggesting that the biophysical driver (e.g., transpiration in spring) is robust and reversible. An AI < 0 indicates a positive correlation between changes in land surface temperature and leaf area; that is, an increase in leaf area is accompanied by land surface warming, and vice versa.

Throughout, the per-unit magnitude quoted in the Abstract and Discussion is the reciprocal of the seasonal median AI (1/AI), i.e. the median across windows of the ratio of the loss-side to the gain-side per-unit sensitivity. This differs from the ratio of the domain-pooled sensitivity slopes, which is computed on aggregated rather than per-window data and is therefore not median-robust; the two are reported separately where both appear. In addition to the per-unit asymmetry index, we report a response-magnitude ratio, defined for each season as the ratio of the domain-mean |ΔLST| under leaf area loss to the domain-mean |ΔLST| under an equivalent leaf area gain, computed within matched ΔLAI_rate bins and averaged across bins with sample-size weighting (Supplementary Fig. 28). Unlike the AI, which compares per-unit slopes, this ratio compares the actual temperature responses and is therefore not inflated by a small gain-side slope.

Assessment of changes in the closed surface energy balance

To understand the biophysical mechanisms through which LAI changes affect land surface temperature, we examined alterations in the complete surface energy budget. At the land surface, incoming solar energy is redistributed among multiple pathways: reflection back to space, emission of thermal radiation, evaporative cooling through latent heat transfer, convective warming via sensible heat exchange, and conduction into subsurface soil layers. This energy partitioning can be represented by the surface energy balance (Eq. (12))⁵²:

$${\mathrm{SW}}_{\downarrow }-{\mathrm{SW}}_{\uparrow }+{\mathrm{LW}}_{\downarrow }-{\mathrm{LW}}_{\uparrow }=\mathrm{LE}+H+G$$ 12

where the left-hand side represents net radiation (R_n). ${\mathrm{SW}}_{\downarrow }$ is the total incident solar radiation reaching the surface, ${\mathrm{SW}}_{\uparrow }$ is the portion of solar radiation reflected away, ${\mathrm{LW}}_{\downarrow }$ denotes the downwelling atmospheric thermal emission incident on the surface, and ${\mathrm{LW}}_{\uparrow }$ is the thermal radiation emitted upward by the land surface. On the right-hand side, LE (latent heat flux) represents the energy consumed during water phase transitions (primarily evapotranspiration), H (sensible heat flux) represents turbulent heat transfer between the surface and overlying air, and G (ground heat flux) represents thermal conduction into the soil column^53,54.

Our analytical approach focused on identifying how each term in the energy balance responded to structural changes in vegetation. By distinguishing the energy balance equation between periods with distinct LAI states as Eq. (13)^46,55:

$$\mathrm{\Delta }{\mathrm{SW}}_{\downarrow }-{\mathrm{\Delta }\mathrm{SW}}_{\uparrow }+{\mathrm{\Delta }\mathrm{LW}}_{\downarrow }-{\mathrm{\Delta }\mathrm{LW}}_{\uparrow }=\mathrm{\Delta }\mathrm{LE}+\mathrm{\Delta }H+\mathrm{\Delta }G$$ 13

where Δ indicates temporal changes in each flux component. Following established methodologies for satellite-based energy studies^55,56, we assumed that LAI modification at a 500-m pixel resolution was insufficiently extensive to induce large-scale atmospheric feedback, particularly cloud formation or dissipation, that would alter incident radiation fields. Under this local-scale approximation, changes in downwelling radiation streams are negligible (i.e., $\mathrm{\Delta }{\mathrm{SW}}_{\downarrow }\approx 0$ and ${\mathrm{\Delta }\mathrm{LW}}_{\downarrow }\approx 0$). This assumption has been validated in previous studies for forest cover changes up to several square kilometers^19,57. With this simplification, the combined changes in sensible and ground heat fluxes attributable to LAI variations could be derived from other energy balance components (Eq. (14)):

$${\mathrm{\Delta }\left(\mathrm{H}+\mathrm{G}\right)}_{\mathrm{LAI}}={-\mathrm{\Delta }\mathrm{SW}}_{\uparrow ,\mathrm{LAI}}-{\mathrm{\Delta }\mathrm{LW}}_{\uparrow ,\mathrm{LAI}}-{\mathrm{\Delta }\mathrm{LE}}_{\mathrm{LAI}}$$ 14

where the subscript LAI explicitly denotes that the flux change is attributed to phenological state anomalies, as isolated using the IDW decomposition method described above.

Vegetation canopy architecture directly affects surface reflectivity (albedo), and the change in reflected solar radiation induced by variations in the LAI is calculated as the product of changes in albedo and incident solar flux^58,59 (Eq. (15)):

$${\mathrm{\Delta }\mathrm{SW}}_{\uparrow ,\mathrm{LAI}}=\mathrm{\Delta }\mathrm{albedo}\times {\mathrm{SW}}_{\downarrow }$$ 15

where ∆albedo represents the change in albedo between the two analysis periods and SW↓ is the climatological mean downwelling shortwave radiation. Equation 15 is applied across all four seasons. In the snow-free seasons, the persistent-forest mask and the paired-difference design make leaf area the dominant control on the albedo change, so ∆albedo can be attributed primarily to the leaf area perturbation^60,61. In winter, the same term additionally captures the contrast between the dark boreal canopy and the bright snow surface that is central to the cold-season mechanism; because snow depth and duration co-vary with leaf area, ∆SW↑ in winter cannot be uniquely partitioned into a leaf area specific component and a co-varying snow-cover component with the present data. The winter shortwave attribution should therefore be read as an upper bound on the leaf area specific effect, and the winter results are interpreted qualitatively throughout.

Land surface temperature modifications directly alter thermal emission from the surface according to the Stefan–Boltzmann law. The change in upwelling longwave radiation associated with LAI-driven LST changes was computed by linearizing the Stefan–Boltzmann relationship around the baseline temperature state^62,63 (Eq. (16)):

$${\mathrm{\Delta }\mathrm{LW}}_{\uparrow ,\mathrm{LAI}}=4\epsilon \sigma {\mathrm{LST}}^3{\mathrm{\Delta }\mathrm{LST}}_{\mathrm{LAI}}$$ 16

where $\epsilon$ is the broadband emissivity (dimensionless), $\sigma$ is the Stefan–Boltzmann constant (${5.67\times 10}^{-8}\mathrm{W}{\mathrm{m}}^{-2}{\mathrm{K}}^{-4}$), LST is the baseline land surface temperature (K), and ΔLST_LAI is the LAI-induced temperature change (K) isolated via the IDW method. Surface emissivity was retrieved from the MOD11C3 product²⁹ using a weighted combination of MODIS thermal infrared bands⁶⁴ (Eq. (17)):

$$ϵ=0.2122{ϵ}_{29}+0.3859{ϵ}_{31}+0.4029{ϵ}_{32}$$ 17

where ${ϵ}_{29}$, ${ϵ}_{31}$, and ${ϵ}_{32}$ denote the emissivity values in MODIS bands 29 (8400–8700 nm), 31 (10780–11280 nm), and 32 (11770–12270 nm), respectively. This band-weighted emissivity provides improved accuracy for broadband thermal flux calculations compared with single-band retrievals⁶⁵.

The data sources of the components in the surface energy balance are as follows: The LE data used to calculate ΔLE were derived from the MOD16A2GF Version 6.1 product⁶⁶ at a spatial resolution of 500 m. The albedo and SW↓ data used to calculate ΔSW↑_LAI were obtained from the MCD43A3 Version 6.1 product⁶⁷ at a 500 m resolution and the MCD18A1 Version 6.2 product⁶⁸ at a 1 km resolution, respectively. The daytime and nighttime LST data were derived from the MOD11C3 LST product²⁹, while the daily mean ΔLST_LAI used to estimate ΔLW↑_LAI was calculated as the average of the daytime and nighttime ΔLST_LAI. The Δ(H + G)_LAI component was calculated from ΔSW↑_LAI, ΔLW↑_LAI and ΔLE_LAI using Eq. (14), with the two radiative terms obtained from Eqs. (15, 16), respectively.

All the remote sensing products were subjected to quality screening using product-specific quality assurance flags. Pixels with poor retrieval quality, cloud contamination, or missing data were excluded from the analysis. Both the albedo and the shortwave radiation datasets were temporally aggregated to seasonal means and spatially harmonized to match the 50 km analysis grid defined above. Seasonal aggregates were computed only when at least 50% of the observations within a season passed quality filters, ensuring robust statistical estimates⁶⁹.

Mechanistic decomposition of vegetation‒climate feedback pathways

To elucidate the mechanistic pathways through which LAI changes were associated with LST via distinct energy flux components, and to quantify the relative contribution of each pathway, we employed structural equation modeling (SEM) for systematic path analysis. The SEM is an advanced multivariate statistical framework that enables the simultaneous estimation of multiple hypothesized dependence relationships while distinguishing direct from indirect effects and has been extensively applied to complex system analyses in ecological and climate change research^30,31.

We constructed a multilevel structural equation modeling framework in which ΔLAI served as the exogenous variable; ΔLE, ΔSW, and Δ(H + G) functioned as the mediating variables; and ΔLST represented the endogenous variable. The model structure was specified based on surface energy balance principles^70,71; LAI anomalies first modulate energy flux components by regulating seasonally dependent canopy properties (e.g., stomatal conductance, albedo, and surface roughness)^3,19, and these flux modifications subsequently collectively determine LST responses. Notably, although ΔLW constitutes a component of the surface energy balance, according to the Stefan–Boltzmann law, LW scales with the fourth power of the LST (LW↑ ∝ T⁴). Therefore, ΔLW is fundamentally a secondary variable dependent on ΔLST rather than an independent driver. To avoid model misspecification arising from reverse causality, we excluded ΔLW from the mediating variables, as its effect was implicitly captured in the Δ(H + G) term through energy balance closure. Specifically, the structural equations of the model can be expressed as follows:

$$\mathrm{\Delta }\mathrm{LE}=a_1\cdot \mathrm{\Delta }\mathrm{LAI}+{\epsilon }_1$$ 18

$$\mathrm{\Delta }\mathrm{SW}=a_2\cdot \mathrm{\Delta }\mathrm{LAI}+{\epsilon }_2$$ 19

$$\mathrm{\Delta }\left(H+G\right)=a_3\cdot \mathrm{\Delta }\mathrm{LAI}+{\epsilon }_3$$ 20

$$\mathrm{\Delta }\mathrm{LST}=b_1\cdot \mathrm{\Delta }\mathrm{LE}+b_2\cdot \mathrm{\Delta }\mathrm{SW}+b_3\cdot \mathrm{\Delta }\left(H+G\right)+{\beta }_{\mathrm{direct}}\cdot \mathrm{\Delta }\mathrm{LAI}+{\epsilon }_4$$ 21

where the coefficients $a_1-a_3$ quantify the sensitivity of each energy component to LAI changes, $b_1-b_3$ represent the conversion efficiency of energy flux to LST responses, ${\beta }_{\mathrm{direct}}$ captures residual direct effects not fully explained by measured energy fluxes, and ${\epsilon }_1-{\epsilon }_4$ are error terms. This model incorporates three indirect pathways (ΔLAI → ΔLE → ΔLST, ΔLAI → ΔSW → ΔLST, and ΔLAI → Δ(H + G) → ΔLST) alongside a direct pathway (ΔLAI → ΔLST). To avoid model misspecification, we allowed covariances among the mediating variables to reflect their physical interdependencies in energy partitioning processes⁵⁵.

During data preprocessing, we implemented rigorous quality control procedures for all the variables. Outliers were identified and removed using the mean ±1 standard deviation criterion. The retained samples were then stratified by season and LAI direction (increase/decrease), with z-score standardization applied to the core analytical variables within each stratum:

$$Z=\frac{X-\mu }{\sigma }$$ 22

where X represents the original variable value, $\mu$ is the within-group mean, and $\sigma$ is the within-group standard deviation. This standardization rendered variables with different units comparable, and the resulting standardized path coefficients could be directly interpreted as effect sizes, facilitating cross-seasonal and cross-directional comparisons⁷².

Model parameters were estimated using maximum likelihood (ML) estimation implemented through the lavaan package in R⁷³. To handle missing data, we used full information maximum likelihood (FIML), which uses all available information more efficiently than listwise deletion and reduces estimation bias⁷⁴. Model fit quality was assessed through multiple indices: the comparative fit index (CFI > 0.95 indicates excellent fit), root mean square error of approximation (RMSEA < 0.06), and standardized root mean square residual (SRMR < 0.08)⁷⁵. This combination of indices provided a comprehensive evaluation of model–data correspondence.

Based on estimated path coefficients, we calculated the indirect effect of each energy flux pathway and its contribution rates to the changes in the LST. Indirect effects were obtained as products of consecutive path coefficients:

$$\mathrm{Indirect}{\mathrm{effect}}_j=a_j\times b_j$$ 23

For instance, the indirect effect via the LE pathway was computed as ($a_1\times b_1$). The total indirect effect summed across all pathways is as follows:

$$\mathrm{Total}\mathrm{Indirect}{\mathrm{effect}}_j={\sum }_{j=1}^3{a}_j\times b_j$$ 24

The total effect combined both indirect and direct pathways is as follows:

$$\mathrm{Total}\mathrm{effect}={\beta }_{direct}+{\sum }_{j=1}^3{a}_j\times b_j$$ 25

To prevent spurious contribution rates arising from cancellation between positive and negative effects, we applied an absolute value approach to calculate the contribution rate of each pathway:

$${\mathrm{Contribution}}_{\mathrm{j}}\left(\%\right)=\frac{∣{Indirecteffect}_j∣}{{\sum }_{k=1}^3∣{Indirecteffect}_k∣}\times 100\%$$ 26

Additionally, we calculated contribution rates relative to the total effect:

$${\mathrm{Contribution}}_{\mathrm{j},\mathrm{total}}\left(\%\right)=\frac{∣{\mathrm{Indirecteffect}}_j∣}{∣\mathrm{Totaleffect}∣}\times 100\%$$ 27

to comprehensively assess the importance of each pathway in the overall LST response.

Author contributions

K.Y. and P.L. conceived the study, and drafted the manuscript. P.L. performed the data analysis, and coordinated its preparation. K.Y. and Q.W. supervised the research. C.H., J.C., Y.S., S.G., X.C., X.M., L.F., G.Y., W.F., L.L., G.W., and Q.W. contributed to interpreting the results and revised the manuscript.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Funding

This work was supported by the National Natural Science Foundation of China Major Program (42192580), the National Natural Science Foundation of China (42271356), and the Fundamental and Interdisciplinary Disciplines Breakthrough Plan of the Ministry of Education of China (JYB2025XDXM905), awarded to K.Y.

Data availability

All the remote sensing data products used in this study are publicly available. The MODIS leaf area index product (MOD15A2H, Version 6.1) was obtained from NASA’s Land Processes Distributed Active Archive Center (LP DAAC; https://lpdaac.usgs.gov/products/mod15a2hv061/). Land surface temperature data were derived from the MODIS MOD11C3 Version 6.1 product (https://lpdaac.usgs.gov/products/mod11c3v061/). Land cover classification was based on the MODIS MCD12Q1 Version 6.1 product (https://lpdaac.usgs.gov/products/mcd12q1v061/). Evapotranspiration estimates were obtained from the MODIS MOD16A2GF Version 6.1 product (https://lpdaac.usgs.gov/products/mod16a2gfv061/). Surface albedo data were derived from the MODIS MCD43A3 Version 6.1 product (https://lpdaac.usgs.gov/products/mcd43a3v061/). Downwelling shortwave radiation data were obtained from the MODIS MCD18A1 Version 6.2 product (https://lpdaac.usgs.gov/products/mcd18a1v062/). The FLUXNET2015 dataset used for phenological algorithm development and validation is available at https://fluxnet.org/data/fluxnet2015-dataset/. The Köppen-Geiger climate classification map is available at http://koeppen-geiger.vu-wien.ac.at/. Initial remote sensing preprocessing and seasonal compositing were performed in Google Earth Engine. Subsequent statistical analyses and figure generation were conducted in Python, and the structural equation models were fitted in R.

Code availability

The code used to generate the results of this study is available at Zenodo (https://doi.org/10.5281/zenodo.21911456) under the MIT license. It comprises the Google Earth Engine scripts for seasonal compositing of the MODIS products, quality masking, the persistent-forest mask and computation of the leaf area change and change-rate layers; the Python scripts for the paired-grid matching, the inverse distance weighting climate-background removal, the moving-window Huber regression, the asymmetry index, the response-magnitude ratio and the robustness analyses (spatial block bootstrap, baseline stratification and truncation sensitivity); the R scripts for the structural equation modeling; and scripts generating Figs. 1, 2 and 4 and the code-generated Supplementary Figs. Delineation of the study domain from the Köppen-Geiger cold-climate zone was performed manually in ArcGIS.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77464-6.