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

Climate change is causing shifts in the growing seasons of plants 1,2 , affecting species performance and 26 interactions 3,4 as well as global carbon, water and nutrient cycles 5,6 . How the timing of autumn leaf 27 senescence in extra-tropical forests will change remains unclear because of the complex seasonal 28 interaction of climate warming, earlier and enhanced vegetation activity, and the annual day-length cycle 7– 29 10 . Here we use experiments, long-term ground observations, and satellite-based vegetation monitoring to 30 show that early-season and late-season warming have opposite effects on the onset and progression of 31 leaf senescence, with a reversal occurring at the year’s longest day (summer solstice). Across ~84% of the 32 northern forest area, increased temperature and vegetation activity before the solstice led to an earlier 33 senescence onset (10% greenness loss) of, on average, –1.6 ± 0.1 days-per-°C, while warmer post-solstice 34 temperatures did not affect senescence onset but reduced its speed (progression to 50% greenness loss) 35 by 0.8 ± 0.1 days-per-°C. Due to the earlier senescence onset, the day at which autumn temperature starts 36 driving senescence progression has been shifting to ever earlier dates, between 1951—2015 at a rate of – 37 0.20 ± 0.07 days per year. These developmental constraints suggest that senescence will start earlier but 38 progress more slowly in the future, revealing Northern Hemisphere-wide compensation effects on trends in 39 growing-season length, caused by enhanced pre-solstice vegetation activity. This new mechanistic insight 40 improves our ability to model carbon uptake by extra-tropical forests under climate change.

To test these hypotheses, we combined phenology data from i) satellite observations across 79 Northern Hemisphere temperate and boreal forests, ii) ground-sourced European observations from 80 widespread deciduous trees 32 , and iii) controlled experiments on European beech. As a proxy for vegetation 81 activity, we used three photosynthesis models (satellite-derived gross-primary productivity [GPP] 33 , LPJ 9 82 and p model 34 ). We then ran linear models to test for the monthly and seasonal effects of photosynthesis, 83 temperature, short-wave radiation and water availability on EOS dates. The satellite data allowed us to 84 differentiate between the onset of senescence and its progression, by analysing the dates when greenness 85 had dropped by 10% (EOS10) or 50% (EOS50) relative to the seasonal maximum. The experiments allowed 86 us to directly test for seasonal variation in the effects of day length, air temperature, radiation, water, and 87 nutrient availability. Finally, we mapped the relative effects of early-season vegetation activity and late-88 season climate across Northern Hemisphere temperate and boreal forests to test for possible historic-89 biogeographic patterns in the drivers of autumn senescence.

103
The reversal at the summer solstice of the effects that air temperature and vegetation activity have 104 on EOS dates was consistent across i) both EOS metrics used here, i.e., the onset of senescence (EOS10;

113
Our finding that the onset of senescence is driven by pre-solstice activity and development, while 114 its rapid or slow progression depends on autumn temperature (Fig. 5b,c and Extended Data Fig. 2) suggests 115 that, under global warming, senescence will start earlier but progress more slowly (scenario 2B in Fig. 1).

116
Indeed, analysis of the temporal trends in remotely-sensed EOS dates and pre-solstice vegetation activity 117 showed that across all analysed northern forest pixels, the onset of senescence (EOS10 date) has advanced we ran multivariate mixed models, including or excluding the effects of pre-solstice or post-solstice (solstice 133 to mean EOS50) activity and precipitation to determine the relative importance of seasonal activity, 134 precipitation, CO2 levels, and autumn night-time temperature (Autumn Tnight). Pre-solstice activity and 135 autumn Tnight had the strongest effects on EOS50 dates, with the effect of pre-solstice photosynthesis being 136 ~3 times greater than that of precipitation and atmospheric CO2 (Fig. 3h). EOS predictions from these 137 models show that the model representing both pre-and post-solstice effects adequately captures within-138 site EOS50 trends in response to mean annual temperature (advance of -0.4 days per each °C increase in 139 mean annual temperature, Fig. 3g). In contrast, the post-solstice model representing only post-solstice 140 activity and precipitation, predicts delays of +0.8 days per °C, while the pre-solstice model predicts advances 141 of -1.0 days per °C, demonstrating that information on both pre-and post-solstice climate is necessary to 142 reproduce the observed EOS50 responses to rising temperature.

143
The offsetting effect of pre-solstice development on autumn-warming-induced delays in EOS50 also 144 is apparent when quantifying either regional trends over the past 70 years (time series and species as 145 random effects; Fig. S8) or spatial patterns in EOS50 (year and species as random effects; Fig. S9

152
S8a, b). The advancing effect of pre-solstice productivity on EOS dates is also evident across spatial 153 gradients (Fig. S9 a,c), such that EOS50 dates occur only +0.85 ± 0.03 days later for each 1°C increase in 154 that region's autumn temperature, while they occur +1.40 ± 0.04 days °C -1 later if the advancing effect of 155 pre-solstice productivity is removed (39% reduction of the geographic autumn temperature response; Fig.   156 S9 b,d).

157
The progression of senescence is modulated by autumn air temperature, as is evident from the 158 delaying effect of warm autumns on EOS50 dates (Figs. 2, 3 and 5 and Extended Data Fig. 6b,c). However,

159
if increased pre-solstice vegetation activity and development (Extended Data Fig. 5a-b) indeed is the main 160 driver of an earlier onset of EOS, one should find an ever earlier susceptibility of trees to autumn cooling.

161
To test this, we conducted temporal moving-window analyses on the European long-term observations

162
showing whether the reversal date -at which increased productivity and temperature start to be associated 163 with delayed EOS50 dates -has shifted over recent decades. The results reveal that this is the case, with

187
To further isolate the mechanisms driving the discovered reversal of the effects of global warming 188 around the June solstice, we designed two experiments using a dominant European tree (Fagus sylvatica).

189
In a first experiment, we cooled (day-/ night-time temperature = 10°C / 5°C) and shaded (~84% light 190 reduction) saplings during different times of the season. Pre-solstice temperature again had a strong 191 advancing effect on autumn phenology, with cooling of trees in June causing a delay in EOS10 and EOS50 192 dates of +16.5 ± 6.6 days and +10.2 ± 2.5 days (mean ± standard error), respectively, whereas cooling in 193 July had no effect and August cooling tended to advance EOS dates ( Fig. 4a and S12a), in full agreement

194
with the global-scale remote sensing data and the European ground observations. The effect of shading 195 was small before the June solstice and most pronounced during July -the month with the highest mean 196 daily radiation and temperature -with +6.5 ± 2.8 days later EOS50 under shade conditions. Radiation effects 197 thus followed a different seasonal pattern than temperature, supporting a direct effect of radiation on leaf 198 ageing 22,36 . Summer photosynthesis was equally reduced in both the shade and the temperature treatments 199 by 52-72% compared to the control (Fig. S13). That pre-solstice temperature, but not pre-solstice light

203
In a second experiment, we tested the effects of elevated sink strength (nutrient addition) and soil 204 moisture. In agreement with the idea that nitrogen supply governs the sink control of leaf development and 205 senescence 19 , there was a strong delaying effect of extra nutrients on EOS dates, with +22 ± 6-day (mean 206 ± standard error) later EOS10 and +14 ± 5-day later EOS50 in trees grown in nutrient-rich soils compared to 207 trees grown in NPK-poor soils (Extended Data Fig. 10 and Fig. S14). Reduced soil moisture slightly delayed

301
Seasonal and monthly photosynthesis/climate: To obtain monthly photosynthesis, Tday, and short-wave 302 radiation values, we took the sums of daily photosynthesis and radiation values and the means of daily Tday 303 values for each month (January to October) [see e.g., Fig. 3a, b]. Similarly, we aggregated values of these 304 variables for six 30-day intervals before, during, and after the June solstice (with 10-day steps), i.e., from 305 May 13 to June 11, May 23 to June 21, June 2 to July 1, June 12 to July 11, June 22 to July 21, and July 2 306 to July 31 (see e.g., Fig. 3d). In addition, we summed the daily photosynthesis values for eight periods (SOS

312
Autumn temperature intervals: It is well known that cold temperatures at the end of the season accelerate 313 the senescence process 9,17,18 , and, for each time series, we therefore determined the optimal autumn 314 interval for which temperature explains most of the variation in EOS50 dates. To do so, we ran linear 315 regressions between EOS50 dates and the temperature interval (Tday or Tnight) 10 to 120 days before the 316 average EOS50 date of each time series (with 10-day steps). Relationships were evaluated using both the 317 coefficient of determination (R 2 ) values and the standardized coefficients (Extended Data Fig. 6c). These 318 analyses showed slightly higher R 2 s and standardized coefficients for Tnight than for Tday, and, for each time 319 series, we therefore included the respective Tnight interval with the highest R 2 in the analyses (hereafter 320 referred to as Autumn Tnight, see e.g., Fig. 3h).

322
Models: To test for the effects of photosynthesis and climate parameters during different times of the year 323 on EOS50 dates, we ran linear mixed models, implemented in the R package lme4 53 . All models included  (Fig. S6a,b). To characterize the effects of photosynthesis, Tday, and short-wave radiation within

346
S7g). To test for multicollinearity among covariates, we calculated variance-inflation factors -using the vif 347 function implemented in the R package car 54 , for all models that contained multiple variables, i.e., the 348 monthly and the full models. All VIFs were < 2, indicating sufficient independence among predictors.

350
Moving-window analysis: To test whether the relative effects of variables have been changing over the past 351 decades, we additionally ran all above-mentioned mixed models separately for each 20-year time period 352 from 1966-2015. To ensure that the long-term average climate of sites did not differ across the moving-353 window periods as a result of site-level differences, we excluded high-elevation sites > 600 m a.s.l. as they 354 were underrepresented in earlier years. We then tested whether the average long-term (1948-2015) climate 355 of sites included in each moving-window period differed between years and found no trend (Fig. S18),

362
To estimate the reversal date of the climate-autumn phenology relationship at which increased 363 temperature and productivity start to be associated with delayed EOS50 dates, we conducted moving-

371
As an alternative to estimating the day at which autumn temperature starts driving senescence 372 progression, for each moving window, we estimated the autumn temperature-sensitive period, based on the 373 autumn period for which temperature best explained variation in EOS50 dates (see above paragraph on

387
Data cleaning. We extracted phenology information for all Northern Hemisphere pixels classified as mixed, 388 broadleaf deciduous, evergreen needleleaf, or deciduous needleleaf forest (tree cover >60%) by the MODIS 389 land cover type map (MCD12Q1 V6) 56 . We removed pixels (i) for which phenology information for fewer 390 than 15 years was available and (ii) for which the mean SOS occurred before March 1 or after May 31 and 391 for which the mean EOS50 occurred before July 18 and after November 30. We then aggregated the data 392 to 0.25 arc-degree (27.8 km at the Equator) spatial resolution to match with the resolution of the climate Analysis. As for the ground-sourced phenology data, climate and soil moisture information at a spatial 398 resolution of 0.25 arc degrees were derived from the Global Land Data Assimilation System (GLDAS) 41 .

399
Information on gross primary production (GPP) was extracted from the MODIS Gross Primary Productivity 400 product (MOD17A2H V6) 33 , which provides an 8-day composite dataset at 500 m spatial resolution. In total, 401 we included eight variables in our analyses: GPP, LPJ model-derived Anetday, Tday and Tnight, short-wave 402 radiation, CO2 levels, precipitation, and soil moisture (at 0-40 cm depth).

404
Seasonal and monthly photosynthesis/climate: To obtain monthly photosynthesis, Tday, and short-wave

409
Autumn temperature intervals: The optimal autumn intervals at which temperature has the strongest effect 410 on senescence dates were calculated as described in the above section Analysis of ground-sourced 411 European phenology observations -Autumn temperature intervals. For EOS50 dates, these analyses 412 showed slightly higher R 2 s and standardized coefficients for Tday than for Tnight (Extended Data Fig. 6b), and,

418
Models: To test for the effects of photosynthesis and climate parameters during different times of the year 419 on autumn senescence dates, we ran pixel-level linear models. Both predictor and dependent variables 420 were standardized to obtain relative effect sizes. To run the monthly models, we included the relative 421 photosynthesis/climate value of each month (January to October) in a multivariate model. For the monthly 422 effects of photosynthesis and short-wave radiation on EOS10 dates, only April to September values were 423 included as predictors in the model, with April values representing the sums of January to April (Fig. 2g). To 424 characterize the effects of photosynthesis, Tday, and short-wave radiation within 30-day-long intervals 425 around the June solstice on EOS10 dates, we included the variable value within the respective interval as 426 single fixed effect (see e.g., Fig. 2h), whereas we additionally included Autumn Tday (to control for autumn 427 temperature) as fixed effect when testing for the effects on EOS50 dates (see e.g., Extended Data Fig. 2).

428
We then ran models including information on pre-solstice (sum of photosynthesis from leaf-out to June 21   and light intensities. In the shading treatments, trees were exposed to shade conditions using a shading net 467 that intercepted c. 84% (± 10%; mean ± SD) of the PAR experienced by the control treatment. All trees were 468 watered regularly to keep soil moisture constant.

469
To quantify seasonal changes in leaf chlorophyll content, we measured the relative chlorophyll

494
To test for differences in leaf senescence dates among treatments, we ran multivariate linear 495 models including temperature and shade treatments as categorical variables.

497
Experiment 2 -Nutrient and soil moisture manipulation. The experiment was conducted in Zurich,

498
Switzerland (450 m a.s.l.) between June and December 2020. Sixty three-year old Fagus sylvatica trees 499 were obtained from a local nursery in February 2020, transferred to 20 l plastic pots containing a 50/50 The experiment followed a 2 x 4 full-factorial design with two nutrient treatments (high versus low 505 nitrogen, phosphorous, and potassium) and four irrigation treatments (High, Intermediate, Low, and No 506 irrigation), resulting in a total of eight treatment combinations (Extended Data Fig. 10). The experimental 507 and observational unit was a pot with a single individual. Each irrigation treatment consisted of 14 individuals 508 of which eight individuals were exposed to high-nutrient and seven to low-nutrient conditions. In the High

519
Leaf chlorophyll content measurements and autumn phenology scoring were done as in Experiment 520

683
According to the model, inter-annual variation in EOS10 should be a function of pre-solstice growth due to 684 developmental constraints on leaf longevity, with later EOS10 in years with slow development / low activity 685 before the solstice (scenario 1) and earlier EOS10 in years with fast development / high activity (scenario 2).

686
The progression of leaf senescence varies with autumn temperature, with faster chlorophyll breakdown in 687 cold-autumn years (scenario A) than in warm-autumn years (scenario B), and the dates of 50% chlorophyll 688 loss (EOS50) are therefore the combined result of pre-and post-solstice effects. An earlier start of 689 senescence in high-activity years (scenario 2) also predicts that trees will become sensitive to autumn 690 cooling earlier than in low-activity years (see blue arrows