Effect of climate warming on the timing of autumn leaf senescence reverses after the summer solstice

Climate change is shifting the growing seasons of plants, affecting species performance and biogeochemical cycles. Yet how the timing of autumn leaf senescence in Northern Hemisphere forests will change remains uncertain. Using satellite, ground, carbon flux, and experimental data, we show that early-season and late-season warming have opposite effects on leaf senescence, with a reversal occurring after the year’s longest day (the summer solstice). Across 84% of the northern forest area, increased temperature and vegetation activity before the solstice led to an earlier senescence onset of, on average, 1.9 ± 0.1 days per °C, whereas warmer post-solstice temperatures extended senescence duration by 2.6 ± 0.1 days per °C. The current trajectories toward an earlier onset and slowed progression of senescence affect Northern Hemisphere–wide trends in growing-season length and forest productivity. Description Editor’s summary Global warming is changing the timing of photosynthesis, with leaves emerging earlier in spring in the temperate and boreal zones. A longer growing season could mean greater carbon sequestration in forests, but the timing of leaves falling in autumn depends on multiple cues, making it difficult to predict. Zohner et al. investigated how leaf senescence relates to day length, temperature, and early-season photosynthesis across northern forests using remote sensing, ground observations, and experimental data. They found that warming had opposing effects on senescence dates depending on when it occurred: Warmer springs with higher photosynthesis correlated with earlier senescence, wheras warmer temperatures in autumn delayed senescence. Incorporating this shift in knowledge may improve predictions of vegetation response to climate change. —BEL Early-season warming hastens onset and late-season warming extends duration of autumn leaf senescence in northern forests. INTRODUCTION Ongoing climate change is causing rapid shifts in plant phenology, with far-reaching effects on the terrestrial carbon cycle and biodiversity. While advances in spring leaf-out dates in temperate and boreal forests are well documented, the effects on autumn leaf senescence are less clear. This is because leaf senescence is not only affected by temperature but also by day length and vegetation activity early in the season in ways that are poorly understood. Accurately predicting future growing-season lengths and plant photosynthesis requires a better understanding of these interacting mechanisms at broad spatial scales. RATIONALE Local observations and experiments suggest that early-season warming, causing earlier spring leaf-out and faster plant development, tends to advance autumn senescence dates. Conversely, late-season warming tends to delay autumn senescence. If true more generally, then climate warming has opposing effects at the start and end of the growing season, with a reversal of effects somewhere in between. To test the generality of the opposing effects of climate warming on leaf senescence in Northern Hemisphere forests, we used satellite, ground, and carbon flux data, as well as controlled experiments. RESULTS Our results revealed that warming early and late in the growing season indeed has contrasting effects on leaf senescence, with a reversal occurring after the summer solstice. Across 84% of the northern forest area, we found that warmer temperatures and increased vegetation activity before the solstice advance the onset of senescence by 1.9 ± 0.1 days per °C, whereas warmer post-solstice temperatures slow the progression of senescence by 2.6 ± 0.1 days per °C. Between 1966 and 2015, the earlier onset of senescence has led to advances of 0.20 ± 0.07 days per year of the date at which autumn temperature starts to drive senescence progression. By contrast, mid-senescence continues to occur slightly later by 0.04 ± 0.01 days per year, leading to a lengthening of the autumnal senescence period. In our experiments, warmer pre-solstice temperatures also led to earlier primary growth cessation (bud set), demonstrating that the impact of a warmer pre-solstice period extends beyond leaf development and life span. This highlights the crucial role of overall plant development and sink activity before the summer solstice in determining growing-season length. CONCLUSION We have developed a unified explanatory framework for predicting autumn phenology, showing that leaf senescence now tends to begin earlier, because of increased pre-solstice vegetation activity, but progresses more slowly, such that the end of senescence occurs later. The reversal in trees’ responsiveness to warming after the summer solstice likely is triggered by changes in day length and allows them to initiate the physiological processes of leaf senescence and nutrient resorption in a fine-tuned balance between source and sink dynamics. Our results demonstrate the impact of developmental constraints (from cell and tissue growth) on autumn leaf senescence and forest productivity, affecting trends in growing-season length across the entire Northern Hemisphere. These insights provide a better understanding of the changes in growing seasons and carbon uptake of temperate and boreal forests under climate change. Autumn phenological responses to pre-solstice and post-solstice climate warming. In cold years, slow development before the summer solstice delays the onset of senescence, and cold autumn temperatures accelerate senescence progression. In warm years, fast development before the summer solstice advances senescence onset, and warm autumn temperatures slow senescence progression, delaying the end of senescence.

INTRODUCTION: Ongoing climate change is causing rapid shifts in plant phenology, with far-reaching effects on the terrestrial carbon cycle and biodiversity. While advances in spring leaf-out dates in temperate and boreal forests are well documented, the effects on autumn leaf senescence are less clear. This is because leaf senescence is not only affected by temperature but also by day length and vegetation activity early in the season in ways that are poorly understood. Accurately predicting future growing-season lengths and plant photo-synthesis requires a better understanding of these interacting mechanisms at broad spatial scales.
RATIONALE: Local observations and experiments suggest that early-season warming, causing earlier spring leaf-out and faster plant development, tends to advance autumn senescence dates. Conversely, late-season warming tends to delay autumn senescence. If true more generally, then climate warming has opposing effects at the start and end of the growing season, with a reversal of effects somewhere in between. To test the generality of the opposing effects of climate warming on leaf senescence in Northern Hemisphere forests, we used satellite, ground, and carbon flux data, as well as controlled experiments. RESULTS: Our results revealed that warming early and late in the growing season indeed has contrasting effects on leaf senescence, with a reversal occurring after the summer solstice. Across 84% of the northern forest area, we found that warmer temperatures and increased vegetation activity before the solstice advance the onset of senescence by 1.9 ± 0.1 days per°C, whereas warmer postsolstice temperatures slow the progression of senescence by 2.6 ± 0.1 days per°C. Between 1966 and 2015, the earlier onset of senescence has led to advances of 0.20 ± 0.07 days per year of the date at which autumn temperature starts to drive senescence progression. By contrast, mid-senescence continues to occur slightly later by 0.04 ± 0.01 days per year, leading to a lengthening of the autumnal senescence period.
In our experiments, warmer pre-solstice temperatures also led to earlier primary growth cessation (bud set), demonstrating that the impact of a warmer pre-solstice period extends beyond leaf development and life span. This highlights the crucial role of overall plant development and sink activity before the summer solstice in determining growingseason length. CONCLUSION: We have developed a unified explanatory framework for predicting autumn phenology, showing that leaf senescence now tends to begin earlier, because of increased pre-solstice vegetation activity, but progresses more slowly, such that the end of senescence occurs later. The reversal in trees' responsiveness to warming after the summer solstice likely is triggered by changes in day length and allows them to initiate the physiological processes of leaf senescence and nutrient resorption in a fine-tuned balance between source and sink dynamics. Our results demonstrate the impact of developmental constraints (from cell and tissue growth) on autumn leaf senescence and forest productivity, affecting trends in growingseason length across the entire Northern Hemisphere. These insights provide a better understanding of the changes in growing seasons and carbon uptake of temperate and boreal forests under climate change. ▪ T he phenological cycles of trees exert a strong control on the structure and functioning of ecosystems (1, 2) and global carbon, water, and nutrient cycles (3)(4)(5). Anthropogenic climate change has resulted in shifts in the growing seasons of temperate and boreal trees, with the start of the season today occurring, on average, two weeks earlier than it did during the 19th and 20th centuries (6) and the end of the season (EOS) being delayed (4,7,8). Each day of a longer growing season may increase net ecosystem carbon uptake by 3.0 to 9.8 gC m −2 (4). Yet, owing to the complex and interacting effects of growing-season climate and the annual daylength cycle, the direction of EOS changes in response to climate change and the cascading effects on ecosystem productivity remain highly uncertain (9)(10)(11)(12)(13).
Characterizing the interplay among the environmental drivers of EOS at broad spatial scales is integral to improving our understanding of growing-season length and tree growth. Cell division, tissue formation, and growth in northern trees are highest at the beginning of the season and decline with shortening days (14)(15)(16)(17)(18), the adaptive reason being the limited time remaining for tissue maturation and bud set before the first frost (19). Local observations and experiments have shown that earlyseason warming, causing earlier spring leaf-out and faster growth and tissue maturation, tends to advance EOS dates (9,11,20,21), whereas late-season warming has the opposite effect, delaying the EOS (22)(23)(24). Increased temperatures and physiological activity in the beginning of the season might drive earlier autumn senescence through a variety of possible mechanisms, including developmental and nutrient constraints (9,25,26), seasonal buildup of water stress (27,28), and radiationinduced leaf aging (29). In contrast, later in the season, a direct effect of temperature (cooling) is likely to drive the timing of autumn senescence (10,24,30). If these trends are correct, then climate warming has opposing effects at the start and end of the growing season, with a reversal of effects somewhere in between.
In this study, we tested whether earlyseason temperature and vegetation activity drive an earlier EOS across temperate and boreal forests, with day length providing the linkage between seasonal activity, air warming, and autumn phenology (Fig. 1). We expected that early-season plant development and growth determine the onset of senescence ( Fig. 1, scenario 1 versus 2), while autumn temperature affects senescence progression (toward full dormancy), with faster chlorophyll breakdown in colder autumns than in warmer Conceptual model of autumn phenological responses to pre-solstice and post-solstice growth and temperature (solstice-as-phenology-switch hypothesis). The onset of autumnal senescence was estimated in this study as the date when the greenness index last dropped by >10% of the seasonal maximum (EOS 10 ). In northern forests, stem growth, development rates, and photosynthetic capacity are highest before the summer solstice and decline with shortening days (14)(15)(16). Interannual variation in EOS 10 is a function of pre-solstice growth and development, with later EOS 10 in years with slow development and low temperature before the solstice (scenario 1) and earlier EOS 10 in years with fast development and high temperature (scenario 2). The progression of leaf senescence varies with autumn temperature, with faster chlorophyll breakdown in cold-autumn years (scenario A) than in warm-autumn years (scenario B). The dates of 50% chlorophyll loss (EOS 50 ) are therefore driven by pre-and post-solstice effects, whereas EOS 10 dates are mainly driven by pre-solstice effects. An earlier start of senescence in high-activity years (scenario 2) also predicts that trees will become sensitive to autumn cooling earlier than in low-activity years (see blue arrows).
autumns (30,31) (Fig. 1, scenario A versus B). Continued acceleration of early-season growth and development under climate warming (4,5,32,33) might thus cause an ever earlier EOS onset, whereas the progression of senescence should be slowed down by warmer autumns (22,23), lengthening the overall senescence period. This leads to four fundamental predictions that were tested in this study: (i) Enhanced pre-solstice temperature and vegetation activity drives an earlier senescence onset (Fig. 1, scenario 1 versus 2). (ii) Growth and temperature effects on senescence dates reverse around the time of the summer solstice. (iii) Autumn temperature affects the speed of senescence, delaying its later stages (scenario A versus B), but has little effect on its start. (iv) The date when trees become sensitive to autumn cooling (blue arrows in Fig. 1) has advanced over recent decades because of an earlier onset of senescence.
To test these hypotheses, we combined phenology data from satellite observations across Northern Hemisphere temperate and boreal forests (34), ground observations from European (35) and American (36) deciduous trees, eddy covariance flux tower measurements (37), and controlled experiments on European beech (38). Seasonal vegetation activity was estimated using direct measurements of leaf-level gas exchange and ecosystem carbon fluxes (37), as well as photosynthesis models [satellite-derived gross primary productivity (GPP) (39) and the Lund-Potsdam-Jena General Ecosystem Simulator (LPJ-GUESS) model (9)]. We then ran linear models to test for the monthly and seasonal effects of photosynthesis, temperature, shortwave radiation, and water availability on EOS dates. The satellite data allowed us to differentiate between the onset of senescence and its progression by analyzing the dates when greenness had dropped by 10% (EOS 10 ) or 50% (EOS 50 ) relative to the seasonal maximum. The experiments allowed us to directly test for seasonal variation in the effects of air temperature and radiation. Finally, we mapped the relative effects of early-season veg-etation activity and late-season climate across Northern Hemisphere forests to test for possible biogeographic patterns in the drivers of autumn senescence.

Effect of temperature and vegetation activity on senescence dates reverses after the summer solstice
Satellite-based phenology data (Figs. 2 and 3 and figs. S1 to S3), European (Fig. 4) and American ( fig. S4) ground observations, flux tower measurements (figs. S5 and S6), and experiments ( fig. S7) all revealed a consistent advancing effect of pre-solstice (i.e., before 21 June) temperature and productivity on EOS dates, which declined after the summer solstice. Thus, across 84% of the northern forest area [18% (Fig. 3C) and 22% (fig. S1C) significant at P < 0.05], increased pre-solstice daytime temperature (T day ) and photosynthesis led to an earlier onset of senescence: each 1°C increase in pre-solstice temperature resulted, on average, in 1.9 ± 0.1 (mean ± 2 SE) days earlier EOS 10  10% increase in pre-solstice photosynthesis resulted in 3.6 ± 0.1 days earlier EOS 10 (satellite data; fig. S1) or 3.7 ± 1.2 days earlier PD 25 (date when photosynthesis had dropped by 25% relative to the seasonal maximum according to flux tower data; fig. S6). A significant delaying effect of pre-solstice temperature and photosynthesis was found for <1% of northern forest pixels. Post-solstice temperature had a small effect on the onset of senescence (see Figs. 2, A and B, and 5A for EOS 10  The reversal of the effects of air temperature and vegetation activity on EOS dates after the summer solstice was consistent across (i) all EOS metrics used here, that is, the onset of senescence (EOS 10 or EOS start ; Fig. 2, A to C, and fig. S8) and mid-senescence (EOS 50 ; Fig. 2, D to F); (ii) continents and forest types (fig. S11); and (iii) a set of alternative variables linked to growing-season activity and development, namely T day (Fig. 3 and fig. S2) and photosynthetic activity (figs. S1 and S3). Along the full latitudinal gradient (30°N to 65°N) studied here, the period during which vegetation productivity had an advancing effect on EOS 10 dates consistently ended after the solstice, at~26 June ( fig. S12). The effect reversal after the summer solstice was further supported by an analysis that used 10-day moving steps around the solstice (Fig. 2, C and F). To further test for the importance of separating pre-and post-solstice variables for EOS predictions, we ran leave-one-out cross-validation of models that included (i) only pre-solstice, (ii) only post-solstice, or (iii) both types of variables as predictors of EOS dates (38). In full agreement with our hypothesis (Fig. 1), the onset of senescence (EOS 10 ) was better explained by pre-solstice than by post-solstice variables, and the full model-including both pre-and post-solstice variables-showed only slightly better performance than the presolstice model (fig. S13, A and D). Because autumn temperature modulates the progression of senescence, EOS 50 dates were slightly better explained by post-solstice than by presolstice variables, with the combination of both variables yielding the best predictions ( fig. S13, B, C, E, and F).
In line with the satellite observations, high pre-solstice temperature and productivity correlated with advanced EOS 50 dates in the European and American plot data, across all species (Fig. 4B and fig. S4) and across a set of alternative variables [T day (Fig. 4) and LPJ-GUESS model-derived photosynthesis (fig. S15)]. On the basis of these findings, we ran multivariate mixed models for the European plot data, including pre-solstice and postsolstice (solstice to mean EOS 50 ) temperature or photosynthesis and precipitation, CO 2 levels, and autumn nighttime temperature (autumn T night ) to determine their relative importance. Pre-solstice temperature (or photosynthesis) had the strongest advancing effect on EOS 50 dates, whereas autumn T night had the strongest delaying effect on EOS 50 dates, with both effects being more than three times greater than that of precipitation and atmospheric CO 2 (Fig. 4C). EOS predictions showed that the full model representing both pre-and postsolstice effects captured within-site EOS 50 trends in response to mean annual temperature (advance of 0.4 ± 0.1 days per°C increase in mean annual temperature; fig. S16). In contrast, a model representing only post-solstice temperature and precipitation predicted delays of 0.7 ± 0.1 days per°C, whereas a pre-solstice model predicted advances of 0.9 ± 0.1 days per°C. These findings demonstrate that information on both pre-and post-solstice cli-mate is necessary to reproduce the observed EOS 50 responses to rising temperature.
Ongoing trends toward an earlier start, slowed progression, and later end of senescence Our finding that the onset of senescence is driven by pre-solstice vegetation activity and development, while the speed of its progression depends on autumn temperature ( fig. S9), suggests that global warming leads to an earlier start and slowed progression of senescence (scenario 2B in Fig. 1). Indeed, across all analyzed northern forest pixels, the onset of senescence (EOS 10 date) has advanced by an average of -0.4 ± 0.1 days per decade between 2001 and 2018 (Fig. 3E), in parallel with increased pre-solstice vegetation productivity ( fig. S14, A and B), with the strongest advances in EOS 10 dates found for regions with the largest increase in pre-solstice productivity ( fig. S14K). When removing the effect of presolstice temperature or photosynthesis on senescence onset by including either variable as a fixed effect in addition to year, the model instead predicted delays in EOS 10 dates of 0.8 to 1.4 days per decade ( Fig. 3E and fig. S1E senescence, leading to delays in EOS 50 dates ( fig. S2G), especially in regions with the largest increases in autumn temperature (Fig. S14M). The offsetting effect of pre-solstice vegetation activity on autumn warming-induced delays in EOS 50 is also apparent in regional trends over the past 70 years (time series and species as random effects; fig. S17). On average, European EOS 50 dates have been delayed by only +0.35 ± 0.02 days per decade ( fig. S17B). By contrast, when keeping presolstice productivity constant by including it as a fixed effect in addition to year, the model predicts a delay of +0.81 ± 0.03 days per decade ( fig. S17D), showing that the increase in presolstice vegetation productivity has offset up to~60% of the delay in EOS 50 that would have occurred had pre-solstice productivity not increased. This explains why EOS 50 delays have contributed only~15% (2.4 ± 0.2 days) to the 16.7 ± 0.4 day-long extension of the growing season that has occurred over the past 70 years (fig. S17, A and B).
The longer senescence duration ( fig. S9) and delayed EOS 50 dates (when greenness has dropped by 50%) under warmer autumns reveal how autumn temperature modulates senescence (Figs. 4C and 5, B and C, and fig. S18, C and D). However, if increased presolstice vegetation development ( fig. S14A) is the main driver of an earlier onset of EOS, one should find an ever earlier susceptibility of trees to autumn cooling. We tested this by using temporal moving-window analyses on the European long-term observations and found that the effect reversal dates, at which increased temperature and productivity start to be associated with delayed EOS 50 dates, have indeed advanced, by an average of -0.7 to -1.0 days per year (Fig. 4F and fig. S15F or  fig. S19B for a shorter moving window). This is also reflected in the moving windows of monthly effect sizes, which show that July photosynthesis has become more strongly associated with delayed EOS 50 dates over recent decades (fig. S19C). As an alternative method of determining when autumn cooling starts driving senescence progression, we modeled the autumn period best explaining EOS 50 . S20B). This earlier start of the period when trees react to autumn cooling provides further evidence for an earlier onset of senescence in response to increased early-season development.
An important consideration of our autumn senescence model presented in Fig. 1 is that the effect reversal date-the compensatory point of the advancing and delaying effects-should be flexible despite day length likely triggering the initial decline in the advancing effect after the summer solstice (14,15,17). This is because the actual reversal date of the antagonistic effects, meaning the period when trees become sensitive to cooling in late summer, is a function of warming and development before the summer solstice, which will vary (Fig. 1). Involvement of day length and a flexible effect reversal date are therefore not mutually exclusive ideas but rather necessary components of the same model.

Factors governing the link between pre-solstice temperature and senescence onset
Previous research on European deciduous trees has suggested a negative feedback between growing-season productivity and autumn phenology, with increased productivity driving earlier senescence (9). The data analyzed here now reveal that the productivity before the summer solstice is indeed linked to earlier senescence dates and that this holds across the entire Northern Hemisphere temperate and boreal forest biome, implying a widespread constraint on future growing-season extensions in response to global warming. To disentangle the environmental drivers of this feedback, we ran multiple linear regression models that included air temperature, solar radiation, water availability, and spring leafout dates as predictor variables, all of which have been shown to affect leaf senescence dates (9,11,20,26,28,40) (Fig. 5 and figs. S21 or S22 using soil moisture instead of precipitation to represent water availability). Across the majority of pixels or sites, pre-solstice temperature was the main driver of EOS 10 dates (Fig. 5A), whereas EOS 50 dates were driven by both pre-and post-solstice temperatures, with opposite effect directions (Fig. 5, B and C). This suggests that temperature-driven development and growth-rather than radiation-induced leaf aging, drought, or leaf-out per se (i.e., constrained leaf life span)-are driving the advancing effect of early-season vegetation activity on autumn phenology.
To further isolate the mechanisms driving the discovered reversal of the effects of global warming around the summer solstice, we conducted an experiment on a dominant European tree species (Fagus sylvatica), in which we cooled (chamber temperature set to 10°C during the day and 5°C at night) and shaded (~84% light reduction) saplings during different times of the season. Pre-solstice temperature again had a strong advancing effect on autumn phenology, with cooling of trees in June causing a delay in EOS 10 and EOS 50 dates of +16.5 ± 6.6 days and +10.2 ± 2.5 days (mean ± SE), respectively. Cooling in July had no effect, and August cooling tended to advance EOS dates (figs. S7A and S23A), in full agreement with the global-scale remote sensing data and the ground observations. The effect of shading was small before the summer solstice and most pronounced during July-the month with the highest mean daily radiation and temperature-with EOS 50 delayed by +6.5 ± 2.8 days under shade conditions. Radiation effects thus followed a different seasonal pattern than temperature, supporting a direct effect of radiation on leaf aging (26,29). Summer photosynthesis was equally reduced in both the shade and the cooling treatments by 52 to 72% compared with the control (fig. S24). That presolstice temperature, but not pre-solstice light availability, affected EOS dates provides further support for the idea that accelerated development and growth under warmer temperatures, rather than photosynthesis-or radiationinduced leaf aging, are the main drivers of the pre-solstice effects on senescence dates. Yet given that both plant developmental speed (tissue expansion and meristematic activity) and photosynthesis early in the season are strongly driven by temperature (41), both processes are linked, and pre-solstice GPP therefore appears to be a reliable proxy of this effect ( fig. S1). An important future avenue will be to disentangle the relative roles of developmental processes within leaves (42) versus growth processes in the rest of the plant (25,41) in driving earlier senescence under increased pre-solstice temperatures. That pre-solstice temperatures and productivity advanced senescence even in species with continuous leaf production throughout the season, such as Betula pendula (Fig. 4B and fig. S15B), suggests that leaf development and longevity cannot be the sole drivers, implying involvement of growth (sink) processes in other organs, including wood, roots, and buds. This idea was further supported by a second experiment in which we tested the response of primary growth cessation (bud set) to pre-and post-solstice warming. The results showed that pre-solstice warming (22 May to 21 June) advanced bud set in F. sylvatica by 4.4 ± 1.8 days, whereas post-solstice warming (22 June to 21 July) delayed bud set by 4.9 ± 1.8 days ( fig. S25), consistent with an effect reversal after the summer solstice. The negative effect of warmer presolstice temperatures on growing-season duration is therefore not limited to leaf development and life span (23,42,43), suggesting an important role of overall plant growth and sink activity before the solstice in determining autumn growth cessation and leaf senescence (14,16,17,41).
The observed EOS advances under elevated pre-solstice temperature and growth may additionally be linked to availability of soil resources throughout the growing season. Optimal growth conditions in spring (high temperatures and sufficient precipitation) may deplete nitrogen availability to trees, which in turn may drive early bud set and senescence (25,26,44). This effect may be further enhanced by increased C:N ratios of leaves under rising atmospheric CO 2 , causing microbial immobilization and reduced nitrogen accessibility (45,46).

Conclusions
Our investigation of the seasonal drivers of autumn phenology shows a consistent reversal after the summer solstice in the effects of climate warming on leaf senescence (i) across large biogeographic ranges with varying presolstice growth (satellite data); (ii) in multiple, broad-ranged tree species with different spring phenologies (ground data); (iii) in eddy covariance measurements; and (iv) under controlled experimental conditions. These findings reveal important constraints on growing-season length and photosynthetic activity, whereby earlier and enhanced activity before the summer solstice drives earlier photosynthetic declines, primary growth cessation, and senescence in autumn. The seasonal control on EOS by air temperature and vegetation activity may be mediated by the annual day-length cycle (15,17) and nutrient availability (26,44). Year-to-year differences in the onset and progression of autumn senescence thus emerge as the result of a complex synchronization between trees' developmental states, seasonal variation in the circadian rhythm, and climate fluctuations. This mediation by the annual day-length cycle provides a unifying framework to explain previous results, in which the magnitude and direction of climate warming effects on autumn phenology varied (9,12,22,23,33,(47)(48)(49)(50), largely because studies did not disentangle pre-and post-solstice variables. It now is clear that warmer temperatures and increased vegetation activity before the summer solstice drive an earlier onset of senescence, whereas, in agreement with previous studies (12,47,51), warmer temperatures after the solstice slow down the progression of senescence, predicting that senescence will continue to start earlier but progress more slowly.
The reversal in how trees respond to temperature during the summer likely evolved as an adaptation to seasonal climates with harsh winters because it allows plants to reliably anticipate the progression of the seasons and prepare for winter dormancy long before the temperature starts dropping (19). The solstice switch in trees' physiological responsiveness to temperature calibrates their seasonal rhythms and mediates how they react to warm or cool temperatures, now and in the future. This enables trees to initiate tissue maturation and the physiological processes of leaf senescence and nutrient resorption (15) in a fine-tuned balance between source and sink dynamics. Improved models of plant development and growth under climate change will need to incorporate the reversal of warming effects after the summer solstice.

Methods summary Satellite observations
Start-of-season (SOS) and EOS dates for Northern Hemisphere forests from 2001 to 2018 came from the MODIS Global Vegetation Phenology product (34). SOS was defined as the date when satellite-derived greenness had reached 15% of its annual maximum. EOS was defined as the date when greenness had decreased by 10% (EOS 10 ), 50% (EOS 50 ), or 85% (EOS 85 ), representing the start of senescence, mid-greendown, and dormancy, respectively. As an alternative measure of the start of senescence, we used the VIIRS/NPP Land Cover Dynamics product (52), which uses a breakpoint algorithm to define the onset of greendown. Information on climate came from (53), and GPP came from the MODIS GPP product (39). We included GPP, T day , T night , shortwave radiation, CO 2 levels, precipitation, soil moisture, and SOS dates as covariates in our analyses.
To investigate the seasonal effects of temperature, photosynthesis, shortwave radiation, and water availability on the timing of EOS 50 dates, we aggregated data for different periods before and after the summer solstice. For each time series, we also determined the optimal autumn interval for which temperature explains most of the variation in EOS 50 dates. We quantified the importance of each covariate in driving interannual variation in EOS dates with pixel-level linear models (Figs. 2 and 3). To approximate the end date of the period during which early-season vegetation activity had an advancing effect on the onset of senescence, we tested the correlation between different GPP periods (always beginning at the SOS date) and EOS 10 dates (fig. S12). To test for decadal-scale temporal trends in EOS dates, we ran mixed-effects models including only year or year and pre-solstice temperature (Fig. 3, D and E) or GPP (fig. S1, D and E) as fixed effects and treating pixels as grouping variables of random intercepts.

Ground observations
Ground data on 396,411 EOS dates of four dominant tree species in central Europe between 1951 and 2015 came from the Pan European Phenology Project (35). Climate data for the same period were obtained from the Global Land Data Assimilation System (53). The EOS 50 corresponded to the date when 50% of leaves had lost their green color. We included photosynthesis, T day , T night , shortwave radiation, CO 2 levels, precipitation, soil moisture, and spring leaf-out dates as covariates in our analyses. Daily photosynthesis was derived from the LPJ-GUESS photosynthesis model (54). As for the satellite observations, data were aggregated for monthly and longer periods before and after the summer solstice, and we also determined the optimal autumn interval for which temperature explains most of the variation in EOS 50 dates.
The importance of these variables in driving EOS dates was then evaluated with linear mixed models, including time series (unique site and species combination) and species random effects. Our final model included fixed effects of pre-and post-solstice T day (Fig. 4C) or photosynthesis ( fig. S15C), as these variables emerged as the strongest drivers of EOS dates.
To test whether the relative effects of variables have been changing over the past decades, we ran the mixed models separately for each 20-year (or 15-year) time period from 1966 to 2015 (or 1980 to 2015). To estimate the date at which the effect of temperature and photosynthesis reverses, we conducted movingwindow analyses of the monthly temperature and photosynthesis effects. Additionally, we estimated the day at which autumn temperature starts driving senescence progression by calculating the autumn period for which temperature best explained variation in EOS 50 dates for each moving window ( fig. S20). We additionally analyzed direct observations of EOS 50 dates for 10 North American tree species from 1991 to 2019 from the Phenology of Woody Species at Harvard Forest dataset (36) (fig. S4).

Flux tower measurements
Eddy covariance data from 10 temperate forest (broadleaf deciduous/mixed) sites with >9 years of data were gathered from the FLUXNET2015: CC-BY-4.0 dataset (37). To represent the onset dates of photosynthesis declines (PD) in late summer, we extracted the day of year on which GPP last dropped below 10% (PD 10 ; fig. S5) or 25% (PD 25 ; fig. S6) of maximum annual GPP. We then tested for the relationships between monthly (April to September) GPP estimates and PD 10 or PD 25 dates by running mixed-effects models, including site as a random effect.

Experiments
In the first experiment, we tested for the effects of seasonal variation in temperature and light on EOS dates. We exposed 3-year-old F. sylvatica trees to cooling or shading conditions during four 1-month-long periods (from May to August). In the cooling treatments, trees experienced a nighttime temperature of 5°C and a daytime temperature of 10°C. In the shading treatments, trees were exposed to an~84% reduction in incoming radiation. Individual leaf senescence dates were defined as the day of year when chlorophyll content last dropped by 10% (EOS 10 ) or 50% (EOS 50 ) of the maximum chlorophyll content in summer. Leaf net photosynthesis was measured with a portable infrared gas analyzer (LI-6800, LI-COR Biosciences, Lincoln, NE). To test for differences in leaf senescence dates among treatments, we ran multivariate linear models including temperature and shade treatments as categorical variables.
In the second experiment, we tested the effects of pre-and post-solstice warming on bud set dates by exposing 4-year-old F. sylvatica trees to cooling conditions (8°C during day and night) during 1-month-long periods before and after the solstice. Bud set was defined as the date when the buds of an individual had reached >90% of their final length (23), indicating aboveground primary growth cessation (23,55). To test for differences in bud set dates and autumn bud growth rates among treatments, we ran mixed-effects models including treatment as categorical fixed effect and bud types (apical versus lateral) as a random effect.