QUINCY is a land surface model that simulates a fully coupled carbon (C), nitrogen (N), phosphorus (P) and water and energy cycle as well as tracking ¹³C, ¹⁴C and ¹⁵N isotopes. The full model description can be found in and in follow-up studies introducing processes such as permafrost freeze-thaw cycles in the Arctic and variable leaf N content . QUINCY has also been evaluated in model inter-comparison projects, such as for surface turbulent heat flux predictions and in for N isotopic processes. We use QUINCY because its modular design allows for easy hypothesis testing and implementation of different representation of processes. Below we provide a short description of the model structure as well as details relevant for leaf turnover and end of season dynamics. In this study, we use the terms “leaf turnover” and “senescence” interchangeably. In terms of model processes, this describes the process of living leaf tissue becoming (“turning over into”) leaf litter. For the purpose of this study, we operate QUINCY at half-hourly time-scale (denoted as dt) and use it as a C-N only model, with no impact of P availability to reduce complexity for model testing introduced by P dynamics. The QUINCY model version used in this study is revision 974a6b7f (Git commit hash).

QUINCY has 14 PFTs , including two non-managed, herbaceous PFTs (C₃ and C₄ grasses). While the default and dynamic model are general grassland phenology models, we maintain the distinction between C₃ and C₄ grasses in this section because other biogeochemical processes and parameters remain PFT-specific.

Plants are comprised of pools of C, N and P and are split into leaf, fine root, coarse root, sapwood, heartwood, fruit, seedbed, labile and reserve pools, where grass PFTs do not have woody pools, i.e. heartwood and coarse roots. The non-structural labile C pool is a short-term storage pool, where newly assimilated C from photosynthesis and nutrients taken up by the roots are stored prior to allocation to respiration, growth and storage. The non-structural longer term reserve pool stores resources for the next growing season and provides resilience against inter-annual variability. C and N can flow between these two storage pools seasonally and based on resource availability and demand. At the start of and during the whole growing season, necessary resources can be pulled out of the reserve pool into the labile pool for growth and maintenance of new and existing tissue. Resources are moved into the reserve pool based on a target reserve pool size that depends on how much C is required to replace growth of leaves and fine roots annually. When the reserve pool becomes too large, resources are moved back into the labile pool.

The start of the growing season is determined by heat accumulation through growing degree days (GDD) since last dormancy accounting for PFT-specific chilling requirements and soil moisture above a PFT-specific threshold (Eqs. – and Table ). Further, GDD can only start accumulating if snow cover thickness is less than 2 cm and the soil is not frozen . Outside the growing season all growth fluxes are set to zero but the labile pools remain to provide resources for maintenance respiration of persisting tissue such as fine roots in cold grasslands (fine root persistence described in Sect. ).

Plant growth in QUINCY is a balance of source (photosynthesis and nutrient uptake) and sink processes (tissue production, respiration, storage) and C and N taken up by plants enter the labile pool and are distributed from there towards the three sinks. For each tissue type, maintenance respiration is determined by N content and tissue temperature and has priority over new growth. The growth of new tissue is determined by tissue stoichiometry and allometry as well as regulated by meristem activity which is reduced by low soil moisture and temperature.

Photosynthesis in QUINCY's multi-layer canopy is calculated separately for sunlit and shaded leaves and is directly affected by soil N availability, since leaf N content, which determines photosynthetic capacity, depends on soil available N. Leaf N is vertically distributed in the canopy and decreases exponentially from the top to the bottom-most layer. Further, photosynthesis is regulated by sink limitation when growth is limited by water, temperature or a lack of nutrients. In this case, labile C accumulates beyond its target size, which then down-regulates photosynthetic activity. Stomatal conductance is calculated following the “Medlyn” formulation and it is limited by soil moisture.

Fine roots of plants are distributed over the entire soil column, but their density generally exponentially decreases with depth. Plants adjust their root distribution dynamically over time to the active layer in the case of permafrost . If water and/or nutrients become limiting, carbon allocation to roots increases . Plant nutrient uptake linearly depends on fine root biomass in each soil layer. Plant water uptake is driven by transpiration which is regulated by water stress. Water stress is calculated based on soil water potential and root fraction for each soil layer, so plant water uptake depends on the root profile and transpiration.

Leaves and fine roots turn over into litter based on the conditions described in Sect.  and sapwood turnover follows leaf turnover. Upon entering the soil, litter is divided into slow- (structural) and fast- (metabolic) decomposing pools. Soil dynamics are simulated through five organic matter pools (slow and fast turning soil organic matter and metabolic, structural and woody litter), which are calculated using first-order kinetics and the formation of soil organic matter follows the CENTURY approach .

The soil in QUINCY is divided into 15 layers of varying layer thickness that increase with depth and total soil depth is 9.5 m (see Table  for specific soil layer depths). Soluble, inorganic nutrients in the soil are available for plant and soil microbial uptake and they compete for these resources. Physical soil properties, such as water content at saturation and field capacity as well as parameters for the water retention curve come from . Soil temperature and moisture are calculated per layer of soil and depend on physical soil properties as well as heat diffusion and water transport and energy exchange with the atmosphere . QUINCY also features snow and freeze-thaw processes that influence soil heat and water exchanges as well as inhibiting fine root growth and nutrient uptake in frozen soil layers.

The end of the growing season equations (Eqs. –) as well as the default leaf turnover model (Eq. ) and their parameters described in this section are unchanged from . During the growing season, leaves are shed at a low, constant rate (1/τleaf, see Table ) which serves as an empirical proxy for herbivory. For the grass PFTs, the growing season ends if either the average air temperature (Tair) over the previous 7 d (τphen) is lower than the PFT-specific temperature threshold (Tairsen, the superscript “sen”, read “senescence”, refers to variables linked to end-of-season processes) or soil moisture stress on stomatal conductance or photosynthesis (βsoilgs) over τphen is lower than a fixed threshold (βsoilsen) or by a carbon starvation condition, where the carbon balance of the plant over τphen is negative, i.e. C required for maintenance (Rmaint,labile) is higher than labile C assimilated through photosynthesis (GPP_τlabile). A minimum leaf age requirement (ageminleaf) for ending the growing season prevents the death of young leaves in the early growing season, where temperatures can fluctuate around 0 °C (Eqs. –). Once the end of the growing season is reached, all leaves and sapwood (stems) are shed at a rate of LAItargetLAI until LAI reaches zero at which point leaf turnover (fturnleaf) is set to 1. The shedding of leaves is constrained by fshed,max, a parameter that ensures that leaves are gradually shed rather than all at once. The default leaf turnover model is shown in Eq. () and all associated variables and parameters are defined in Table . In summary the end of the growing season is triggered when the mean leaf age in days (agemeanleaf) exceeds ageminleaf: 1agemeanleaf>ageminleaf and one of the following conditions is met: 2Tair<Tairsen3orβsoilgs<βsoilsen4orGPPτlabile<Rmaint,labile. Once the end of the season is reached, grasses shed above-ground biomass, including both leaf and stem tissue, at a rate calculated by: 5fturnleaf=max1τleaf×dt,fshed,max×LAItargetLAI×dt where LAI_target is a model-wide variable determined by root-to-leaf biomass ratio, fine root biomass and constrained to below the parameter LAI_target,max and dt denotes one time step.

During the growing season, fine roots also turn over continuously (Eq. ) and this is assumed to be mostly driven by predation on fine roots . At the end of the growing season, fine roots broadly follow an annual or perennial strategy depending on the climatic conditions at the time of senescence . They are either shed or retained over the dormancy period depending on the type of grassland (Eqs. –). In cold grasslands, i.e., where low temperatures initiated the end of the growing season, fine roots are maintained throughout the dormancy period and in seasonally dry grasslands, i.e., where water stress triggers the end of the growing season, fine roots are shed and then regrown once the growing season starts again. If the end of the growing season is initiated by neither cold nor drought but through a negative carbon balance, roots are also shed (Eqs. –). Storage pools (labile and reserve) are maintained throughout the unfavorable season.

Due to the default models' basic representation of end of season dynamics and building on 's species-specific drought response model in Australian herbaceous systems, we have developed a new representation of leaf turnover to make QUINCY more dynamic (i.e. decoupled from fixed thresholds) and process based. We have decoupled leaf turnover from the growing season and represent it as a continuous function of air temperature, soil water availability and day length. This allows plant growth to be constrained by multiple environmental factors simultaneously and allows biomass turnover on a more continuous basis, as a more accurate representation of multi-species systems. Additionally, the model formulation also includes co-limitation of multiple factors, either concurrent or at different times of year and importantly allows for leaf abscission during the growing season if conditions become unfavorable.

Turnover in the dynamic model is dictated by three components: temperature, moisture and day length which are shown in Fig. . The temperature component (ftuntemp, see Fig. a) which has been adapted from QUINCY's temperature control function on the meristem (see Eq. ) is calculated as: 6fturntemp=e-λtemp×Tairktemp where λtemp and ktemp are scaling factors (see Table ), and Tair is the air temperature in °C averaged over τphen.

The soil moisture stress-induced leaf turnover component of the dynamic model (fturnmoist, Eq. ) was adapted from the soil moisture constraint function on stomatal conductance (βsoilgs, see Table ) in QUINCY. With this formulation, leaf turnover increases linearly from its minimum value 0 when soil water potential in the rooting zone (Ψsoil) is 0 MPa to its maximum value 1 at the permanent wilting point (Ψleafmin, see Table  and Fig. b). Ψsoil is defined as the water potential within soil layers where roots are present, weighted by the fraction of roots present in the individual layers. We have opted for a linear relationship between Ψsoil and fturnmoist due to the non-linear relationship between Ψsoil and soil water content (SWC; Θsoil): 7fturnmoist=min1,ΨsoilΨleafmin where Ψsoil denotes the soil water potential in the rooting zone and the parameter Ψleafmin is the minimum leaf water potential (see Table 1) from . The calculations for Ψsoil, which is derived from pedotransfer functions of , and Θsoil can be found in the Supplementary Material (Sect. 6.3) of .

Having a day length component (fturnlight) in the dynamic leaf turnover model is important especially for high-latitude grasslands for two reasons. First, autumn phenology in those regions happens rapidly which can be attributed to the short growth period (∼3 months of snow-free season) and second, light may become a limiting resource for plants toward the end of the growth period . We utilize the difference in day length throughout the year to capture the rapid autumn phenology at higher latitudes whilst still retaining a low impact of day length at lower latitudes where phenological events take longer and light may not be a limiting factor. We choose this approach over other more direct measures such as photosynthetic photon flux density (PPFD), because such metrics are strongly influenced by short-term meteorological variability and we wanted to avoid that our model responds to conditions such as prolonged cloud-coverage by shedding leaves at the wrong time of the year. Moreover, when averaged over longer time scales, PPFD exhibits a seasonal pattern that closely aligns with the annual cycle of daylight duration, therefore providing no significant advantages over the simpler day length approach.

Day length (Dn, h), defined as the time between sunrise and sunset, follows the calculations from and is calculated from site latitude in degrees (ϕdeg) and solar declination (δdeg, Eq. ) for a given day of the year (n). ϕdeg and δdeg are converted to radians (Eqs.  and ) to calculate the sunset hour angle (ωs, Eq. ), which is then used to calculate Dn (Eq. ). 8δdeg=23.45°×sin⁡2π365×(n-81)9ϕrad=ϕdeg×π18010δrad=δdeg×π18011cos⁡ωs=-tan⁡ϕrad×tan⁡δrad12Dn=max0,min24,24π×arccos⁡cos⁡ωs where -90°≤ϕdeg≤90°, ϕrad and δrad are the site latitude and declination in radians, respectively and Dn∈[0,24] h.

To calculate fturnlight, we use the rate of change in photoperiod, defined as the difference between the current day length (Dn) and the length of the previous day (Dn-1) to shed increasingly more leaves. Because day length is a daily variable, fturnlight is calculated once per day and remains constant across all half-hourly time steps (dt) within that 24 h period. If day length is increasing (i.e. Dn>Dn-1) or more than 15 h of day light are available on the current day, fturnlight is zero (Eq. ). After the summer solstice, once day length starts to decrease (i.e. Dn<Dn-1) and less than 15 h of day light are available, we use the difference between Dn-1 and Dn to calculate the rate of change over Dn to shed increasingly more leaves until the maximum rate is reached at the September (March on Southern hemisphere) equinox (Eq. ). We introduced the 15 h threshold as we assume that above 15 h of day light, light is not limiting. From the equinox the shedding rate decreases again until the winter solstice or when day length reaches zero (polar night, see Eq. ). The use of solstices and equinoxes follows directly from the geometry of Earth's orbit. The solstices mark the turning points of this cycle where the Sun's declination changes the most slowly at the summer and winter solstices, so daily changes in day length approach zero. At the equinoxes the Sun crosses the celestial equator, where the rate of change of declination is maximal, and therefore the rate of change of day length (i.e. Dn-Dn-1) is also highest. In summary, the three cases of the light component fturnlight are calculated as:

if Dn>Dn-1, or Dn>15 h: 13fturnlight=0 if Dn<Dn-1 and day length is <15 h: 14fturnlight=Dn-1-DnDn if polar night (i.e. Dn=0 h): 15fturnlight=1 where Dn is the current day length and Dn-1 is the length of the previous day.

Using this approach, we are able to increase the rate at which leaves are shed in high latitudes, while lower latitudes remain virtually unaffected as the rate of change of day length at lower latitudes is much lower (see Fig. c).

The full dynamic leaf turnover (fturnleaf) calculation is shown in Eq. (). We use a multiplicative interaction to represent the compounding effects of temperature (fturntemp), soil moisture (fturnmoist) and day length (fturnlight) components so that under conditions where more than one factor is limiting, leaf turnover is higher than under conditions where only one of the factors is limiting. The maximum turnover rate per timestep is capped through a parameter fshed,max, which prevents shedding of all leaves at once. 16fturnleaf=1-∏i=1n1-fi×fshed,max×dt where fi are the individual environmental stress-induced turnover functions (fturntemp, fturnmoist, fturnlight, Eqs. –) and dt is the length of one timestep.

The parameter fshed,max, which is also part of the default model and prevents grasses from shedding all leaves immediately when the growing season ends, is set to 5 % (0.05) by default. We have kept this parameter in the dynamic model but ran a parameter sensitivity test in QUINCY to optimize fshed,max as 5 % imposes a too slow shedding of leaves under unfavorable conditions. The results of these tests, which we have conducted at the four main sites, reveal that 10 % or a shedding of all leaves over 10 d at the highest shedding rate (fshed,max=0.1) is a more suitable value for our model in terms of daily GPP fit to observational data.

Finally, we have added an additional condition (dsnowsen) to the growing season in the dynamic model so that an accumulated snow depth of >2 cm stops the growing season. This does not directly impact the dynamic turnover model as it is not controlled by the growing season, but it stops the growth of new tissue under unsuitable conditions and when the turnover rate is already high due to low air temperature.

In the seasonally dry grassland, ES-LMa, we are able to reduce the delay of the onset of senescence (Fig. b) and the overall seasonal dynamics of modeled GPP (Fig. a) fit the EC data better in our dynamic model compared to the default model (adj. r2=0.77 and 0.88, and RMSE 1.25 and 0.76 g C m⁻² d⁻¹, respectively, see Table ). The dynamic turnover improves daily GPP fit in spring and immediately before the dry season, as well as after the dry summer period at the start of the new growing season. This leads to an increase in annual GPP (from default: 481.80±338.143 to dynamic: 613.06±403.52 g C m⁻² yr⁻¹, Fig.  and Table ) meaning that the dynamic model's annual GPP is closer to the observed annual GPP (EC) of 729.43±479.62 g C m⁻² yr⁻¹, although the default model is not significantly different from the EC data to begin with. Further, there is no detectable significant difference in annual GPP between the default and dynamic model (Table ). However, overall long-term ecosystem C storage in the dynamic model is higher than in the default model (Fig. ). The same improvements as in the seasonal GPP dynamics can also be seen in the modeled LAI compared with GCC. We are able to reduce RMSE from 0.17 in the default model to 0.15 in the dynamic model and increase the adj. r2 from 0.10 to 0.39 (Fig. b). Although, the seasonal mismatch of LAI and GCC is still quite pronounced in the dynamic model, this is expected due to the comparison of two different metrics.

For the montane, seasonally cool grassland, CGE, where cold winter temperatures are the main determinant of leaf senescence, the default model was unable to capture the onset of winter and showed a delay of nearly 60 d (Fig. f). With the direct response to temperature in our dynamic model we are able to shed leaves quicker at the end of the season around November and therefore achieve an overall better fit of predicted LAI to observed GCC (adj. r2 default model: 0.31, dynamic model: 0.36) and we are able to reduce RMSE from 0.27 to 0.23. Leaves also start to grow earlier in spring compared to the default model as shown by an earlier increase in LAI, which brings the start of the season closer to the observed start of the growing season as shown by the GCC data. Long-term C storage however remains unaffected by the dynamic model (Fig.  and Table ).

At the cold and seasonally light-limited site, US-Tol, the default model was unable to reproduce GPP dynamics, especially in the summer, where it underestimated the magnitude of daily and consequently annual GPP (Fig. ). Similarly to ES-LMa, the dynamic model is able to reduce the error in daily GPP here (Fig. c). The RMSE of daily GPP was reduced from 0.60 g C m⁻² d⁻¹ in the default model to 0.46 g C m⁻² d⁻¹ with the dynamic model (Table ) and annual GPP increased from 191.10±8.49 in the default model to 204.30±12.72 g C m⁻² yr⁻¹, compared to the EC flux annual GPP estimate of 224.57±22.72 g C m⁻² yr⁻¹ (Table ). Although the mean differences exceed the inter-annual standard deviations, the ANOVA indicates that annual GPP of neither model is significantly different from each other nor from the EC flux estimate (p>0.05, Fig.  and Table ), likely reflecting the small sample size (nyears=3). The increase, especially in summer daily GPP, also leads to a better model fit of the dynamic model (adj. r2: 0.76 (default) and 0.86 (dynamic)), but LAI still shows an overall poor fit in the dynamic model (adj. r2: 0.0002 (default) and 0.0028 (dynamic)) even though we are able to replicate the steeper drop of GCC better and shift the date of minimum LAI from mid January to early November (Fig. d). Just as at the seasonally cold site CGE, long term C storage remains unaffected with negligible decreases in SOC stock (Fig.  and Table ).

At IE-Dri, a site that is neither driven by strong temperature nor moisture fluctuations or extremes, with a more continuous growing season and turnover, the dynamic model performs very well (Fig. e), as the capacity for continuous turnover is a key new factor in the model. The main drivers of leaf turnover at IE-Dri are soil moisture and day length, though their overall contributions remain small (Fig. ). With the dynamic model, we can capture the seasonal GPP fluctuations and its amplitude accurately and we reduce the RMSE of daily GPP from 1.56 in the default model to 1.20 in the dynamic model and increase the fit from 0.83 to 0.86 (adj. r2) compared to daily EC flux GPP (Table ). This also leads to a significant increase in annual GPP from the default model with 1519.36±75.43 to 1780.66±54.45 g C m⁻² yr⁻¹ in the dynamic model (p=0.03, Fig.  and Table ) and annual GPP, which in the default model differed significantly (p=0.01, Table ) from the observed EC annual GPP (1859.25±132.35 g C m⁻² yr⁻¹), shows no significant difference in the dynamic model (p=0.59). This also influences long-term C storage, where the dynamic model shows a significantly higher C stock in soil and vegetation (Fig.  and Table ).

Overall, we are able to improve the response of grasses to seasonal moisture, cold and light limitation across all four main sites. We are able to reduce error and improve model fit in predicting seasonal trends of GPP and LAI (Fig.  and Table ), which lead to improves annual GPP estimates (Fig.  and Table ) resulting in significant differences in long term soil C storage (Fig.  and Table ).

We tested the performance of our dynamic model at 56 sites across climate zones. We tested the whole season fit using mean annual GPP (Fig. a and b and Table ), our models' capability to predict the onset of senescence using end of season date for the same sites (Fig. c and d and Table ), as well as end of season fit, i.e., the senescence period between the end of season date to the first following minimum GPP, before the next growing seasons starts (Fig.  and Table ).

Over all sites, both the default and dynamic models have similar model fits (Table ), however, the dynamic model performs better in several key areas. The default model consistently underestimates annual GPP at tropical savanna and hot summer Mediterranean sites (Aw: 665.00 (default) vs. 1358.00 (EC), n site years = 17; Csa: 592.00 (default) vs. 967.00 g C m⁻² yr⁻¹ (EC), n site years = 21), while overestimating GPP in both hot and cold semi-arid sites (BSh: 731.00 (default) vs. 661.00 (EC), n site years = 20; BSk: 522.00 (default) vs. 210.00 g C m⁻² yr⁻¹ (EC), n site years = 36). The dynamic model improves model fit at tropical savanna sites (RMSE default: 698.86 and dynamic: 681.06), although the GPP remains low (dynamic: 685.00 g C m⁻² yr⁻¹), and the dynamic model shows a significant improvement at Mediterranean sites (RMSE default: 482.72 and dynamic: 315.81). Model GPP at subarctic and tundra sites is also improved by the dynamic turnover (ET: 358.00 (EC), 157.00 (default), 171.00 g C m⁻² yr⁻¹ (dynamic) and RMSE: 220.77 (default) vs. 202.24 (dynamic, n site years = 16); Dwc: 318.00 (EC), 358.00 (default), 393.00 g C m⁻² yr⁻¹ (dynamic), n site years = 2).

The model performance in predicting EOS dates (Fig. c and d and Table ) follows a pattern similar to the annual GPP results, except that Csa sites perform worse in the dynamic model (RMSE: 26.65 (default) and 52.41 d (dynamic)) and BSh (hot semi-arid) and BWh (hot desert climate) show reduced errors. Overall, the model shows improved RMSE of EOS dates at 12 out of 37 sites. The dynamic model performs better again at Aw, Dwc and ET sites compared to the default model. At tropical savanna (Aw) sites the dynamic model overestimates the EOS date compared to EC flux data where the end of the season is on DOY 150. The dynamic model EOS is 166 and the default model is 144. However, the dynamic model shows a lower RMSE of 45.50 d compared with the default model where RMSE is 56.89 d (n sites = 3). At hot semi-arid sites (BSh) the dynamic model shows a significant improvement over the default model with a near perfect match of EOS DOY (EC: 165, default: 188, dynamic: 164) and RMSE reduces from 53.97 to 45.78 d (n sites = 4). Hot desert sites (BWh) on the other hand, show a reduced RMSE from 74.81 (default) to 71.26 d (dynamic, n sites = 3) but similarly to tropical savanna sites, place the EOS date later in the year (EC: 174, default: 169, dynamic: 209). At subarctic sites, the EOS is a week earlier in the dynamic model, moving it closer to the EC EOS (EC: 268, default: 299, dynamic: 292) and has a reduced RMSE (default: 31.00, dynamic: 24.00). At the tundra site the dynamic model shows a similar trend but only a fractional improvement of one day (EOS EC: 259, default: 284, dynamic: 283) which results in a marginally improved RMSE (default: 25.00, dynamic: 24.00).

The final analysis in which we separated leaf senescence from the rest of the year to quantify site level end of season fit of daily GPP (Fig.  and Table ) reveals that the dynamic model overall performs better across the 32 sites in terms of error (RMSE default: 1.39 g C m⁻² d⁻¹, dynamic: 1.13 g C m⁻² d⁻¹). The dynamic model shows a better fit at 17 of the sites (lower RMSE) which span across a wide range of climates, from cold-winter, semi-arid (BSk: n sites = 2), over oceanic (Cfb: n=2), humid-continental (Dfb: n=6; Dwb: n=1) and humid subtropical (Cfa: n=3) to dry-winter (Dwc: n=1), subarctic (Dfc: n=1) and tundra climates (ET; n=1). At hot semi-arid (BSh: n=2), but also some oceanic (Cfb: n=4) and some humid subtropical (Cfa: n=2), hot-summer Mediterranean (Csa: n=2) and continental (Dfb, Dfc and Dwb; n=1 each) sites the dynamic model perform less well. Generally, the fit of senescence periods is better at sites where the dynamic model struggled to estimate annual GPP and EOS date. The extracted periods represent the overlap between the individual EOS periods of EC data and the two models, so it is decoupled from the actual EOS dates, since it is difficult to evaluate the significance of calculating singular dates, which change based on the method used, it may be more important to look at the entire senescence period.

Traditional leaf phenology research often focused on mid-latitude temperate systems, where a temperature threshold for both start of season and end of season may be a good approximation, but even in those ecosystems, the modelling focus has been largely on start of season phenology, with end of season processes being much more poorly understood. While progress has been made in stand-alone phenology models , implementation in LSMs has been slow and mainly focused on trees and shrubs . Many models, including LPJ-GUESS and CLM5 , have integrated similar multi-factor phenology models, yet they often rely on discrete environmental triggers. For instance, CLM5 simulates leaf senescence of grasslands through accumulated stress based on moisture, temperature, and day length thresholds. Once these predefined thresholds are surpassed, senescence is triggered. Similarly, LPJ-GUESS utilizes survival fractions of leaves based on climatic conditions and sheds all leaves on the statistically coldest day of the year, therefore introducing a hard-coded threshold for grasses. These discrete, trigger-based mechanisms often fail to capture how ecosystems respond to environmental stress.

Grasslands may be one of the most moisture-sensitive ecosystems on the planet and drought frequency and intensity is expected to increase with climate change . Yet, models are not able to reproduce observed drought responses of grasslands , even though the timing of these events has implications on C fluxes and the capability of shifting grasslands from a C sink to a source . To address these limitations, our new dynamic model moves away from binary triggers, allowing us to capture the seasonal drought response in a dehesa, savanna-like, ecosystem, ES-LMa, more accurately and show that controlling senescence as a continuous process rather than a classical threshold end of season can lead to better whole-season model performance. Arid and semi-arid grasslands are frequently composed of species with a variety of levels of drought resistance and a threshold approach can be unsuitable, therefore having a dynamic response to changes in soil moisture conditions can be more ecologically relevant. We show that improving end of season phenology impacts the whole year by alleviating constraints placed on early season growth in threshold models (here: after the dry summer at ES-LMa), which leads to higher early season growth and a better fit with the EC data. Further, more accurate leaf senescence can reduce carbon costs in plants as maintaining leaves requires carbon allocation to respiration. Importantly, we show that improving leaf turnover does not only affect seasonal dynamics, but impacts long term C storage, as well (Fig. ), where our dynamic model shows significantly higher C storage in the entire ecosystem.

Photosynthetic activity is primarily limited by light and even more so in high latitude ecosystems. At present, light limitation may not be very relevant at high latitudes as snowfall occurs before day length becomes very limiting, but under future climate which will shift the snow-free period later into the year , light limitation may become increasingly important . This also applies to spring phenology, where earlier snow melt and higher spring temperatures may be able to extend the growing season, however plants may not be able to take advantage of this extended growing season as the short photoperiod very early in the year becomes limiting . Our model could predict high latitude grasslands more accurately and give evidence for the importance of day length control in leaf turnover . Our dynamic model is able to reproduce the end of the season better at US-Tol and to a certain extend improves the peak GPP in the summer at this site (Fig. c and Table ). However, peak GPP in the dynamic model is still underestimated compared to observations, similarly to other tundra sites (Table ). This is likely due to the model's shortcomings in representing nutrient availability in high-latitudes, an issue common across LSMs .

Continental climates are characterized by seasonal temperature variations and temperature is one of the main drivers for autumn phenology in grasslands . Higher temperatures in autumn have been shown to delay the onset of leaf senescence in herbaceous plants and climate change is predicted to prolong the growing season later into the year , so it is important for models to accurately predict autumn phenology and dynamically respond to changes in temperature. Our dynamic model is able to predict the onset of senescence at CGE better than the default threshold model and is able to reproduce the observed “brown-down” in the GCC better (Fig. f). Even at more extreme cold sites such as the tundra site US-Tol, we were able to improve the end of the season and shift the minimum LAI forward by 60 d and closer to the observed minimum GCC. Yet, LAI lags behind the GCC data at both sites in regard to timing of its peak and end of season, but this is expected and consistent with other observations . Since GCC is the green fraction of an image taken of the canopy from above and LAI is a measure of leaf area per ground area, the optical greenness of the upper canopy may begin to decline while significant leaf biomass remains, causing a lag in the data compared to actual leaf biomass. This discrepancy is further compounded by snow cover as GCC is highly sensitive to the high albedo of snow, which obscures leaves that may still be physically present and accounted for in the LAI data. However, since GPP at US-Tol also shows a lagged response in autumn phenology, we can assume that the mismatch in GCC and LAI can not solely be contributed to them being different metrics. This lag in GPP is very likely not caused by issues with the dynamic leaf turnover representation alone, but rather leaf growth in QUINCY, as the growing season at US-Tol on average ends only at 280 (see Fig. ) at which point GPP is already near zero in the EC flux data. Decoupling leaf growth from the growing season, which is also based on a threshold approach may solve this issue. Further, CGE is a managed grassland subject to regular mowing and we did not include a mowing scheme into our model. This leads to sharp drops in greenness after each mowing event, which is lacking in our model . This likely at least partly contributed to this mismatch of GCC and LAI apart from them being two different metrics. Finally, our model is also able to improve whole season fit of GPP even at sites that do not experience (severe) cold nor dry periods throughout the year, such as we have shown at the site IE-Dri. The gradual shedding of leaves which is required to replicate these moderate conditions, is only achievable with our flexible approach and the default threshold model which essentially applies the same shedding rate in all climates, cannot reproduce this. As a result we are able to significantly improve annual GPP, match observed GPP closely and show that improving end of season dynamics strongly impacts annual GPP predictions as well as long-term C storage (Fig. ), which contributes to reducing uncertainties present in global carbon budgets .

Our model does not prescribe a priori which of the three triggers is limiting and we show that the model is widely applicable at sites across the globe. It provides a first step towards more flexible, physiology-driven leaf turnover but highlights the need for a more adaptable drought model that is able to incorporate different drought responses such as presented in . The model is able to capture end of season phenology more accurately across a wide range of climates (see Fig.  and Table ). A poorer model fit at some of the sites with the dynamic model is likely caused by knock-on effects from changes in growing season length on new growth at the start of season. The default turnover model may get things right for the wrong reason and may have been able to “mask” delayed start of season as it frequently allowed for very long growing seasons which do not match site observations from EC data, and therefore in terms of carbon budgets (e.g., annual GPP). Furthermore, the dynamic model allows for leaf turnover during the entire growing season which may further impact early season growth by inducing leaf senescence of young leaves if conditions become harsher. This is a more appropriate response and is able to capture events such as late frost in spring that damages leaves, but the default model which effectively removed leaf turnover during the growing season was not able to capture this and therefore placed less pressure on the already underestimated early season growth in QUINCY.

Our model performs well in terms of annual GPP at sites with strong seasonality of precipitation (Aw: tropical savanna and Csa: hot summer Mediterranean) and also temperature (winter-dry subarctic and tundra, Table ). This demonstrates that our model successfully captures leaf senescence under pronounced seasonal extremes in both moisture and temperature. However, at a number of seasonally cold sites the dynamic model (temperate/continental) shows a worse fit. This may suggest that the model's single temperature response function may not sufficiently account for local thermal adaptation , leading to reduced sensitivity in temperate regions with less severe winters. Plants in seasonally cold sites may also lack photoperiodic signaling which triggers senescence through gene expression whereas our model primarily accounts for light limitation. The subarctic (Dwc) site is CN-Dan located at 30.49° N, the subarctic climate of the site is therefore a result of altitude rather than latitude. This suggests that day length turnover in the model has a near zero impact at CN-Dan and the same goes for the second tundra (ET) site, CN-HaM, at 37.37° N. Since the model performs better at these two sites and the sites are located at mid-latitude like the temperate/continental sites that perform worse in the model with the only difference being altitude, it could point toward a lacking temperature sensitivity to cold-hardiness, not a lack of day length control for the worse model fit at continental/temperate sites. Further, models often lack proper representation of constraints on photosynthesis at mid-latitudes which can influence annual GPP, such as the suppressing of photosynthetic activity by low temperature which may contribute to the poor model fit at these sites .

Autumn phenology and its timing is a key part of ecosystem functioning and has been shown to play an important role in predicting net ecosystem productivity . The dynamic model improves the accuracy of the timing of leaf senescence at subarctic (Dwc) and tundra sites (ET), but not at temperate sites (C climates). Thus, further showing that the model as is parameterized now may be more appropriate for harsher climates and less so for temperate sites. Although, the number of sites with ET and Dwc climate is very low (between one and two sites per class and metric) which reduces confidence in these results, those sites as well as the tundra site US-Tol consistently perform better in terms of errors across all three metrics that we have tested (annual GPP, EOS date and EOS period). Beside the Dwc and ET sites, the model performs better at tropical savanna sites (Aw), and unlike annual GPP, at hot semi-arid and desert sites (BSh and BWh). However, at hot summer Mediterranean sites (Csa), where the end of season is in the summer months caused by a lack of precipitation, the model performs less well. It should be noted that ES-LMa (Csa) is not part of this analysis. At Csa sites, the estimated EOS date in the dynamic model is day 192 (default: 219). This is earlier than the estimated EOS date of the EC data (DOY 225). This indicates that our model, may be overly moisture-sensitive in ecosystems that are defined by dry summers. Yet at tropical savannas, ecosystems that are also characterized by strong seasonal rain patterns and at hot semi-arid and desert sites the model has improved. This highlights that grasslands have vastly different responses to dry conditions and may suggest that a singular drought-response mechanism may not be sufficient and implementing different drought sensitivities, similar to could improve model performance.

The mismatch in moisture sensitivity is also reflected when extracting the senescence period from the whole year, where our model performs well at sites that do not have a dry season (12 of 18 sites with climate Cfa, Cfb, Dfb, Dfc; indicated by second letter “f”) and it performs less well at sites that are either summer- (second letter “s”) or winter-dry (second letter “w”), with the exception of two cold, semi-arid sites (BSk), one hemi-boreal site (CN-Du2) and one winter-dry subarctic site (Dwc, site CN-Dan). However, finds that temperature rather than precipitation is the main driver for autumn phenology in Chinese herbaceous species, which could explain why the model performed better at CN-Du2 and CN-Dan even though seasonally dry sites generally show a weaker fit in our model. Furthermore, at US-Whs, one of the two summer-dry BSk sites located at a relatively high elevation (1380 m a.s.l.), leaf turnover appears to be driven primarily by temperature rather than moisture, which may explain the better model performance there. The generally better fit at mesic sites, as well subarctic/tundra sites could indicate that our model may struggle to reproduce species-specific water-stress adaptations.

Overall the model allows for interaction between triggers and co-limitation, which works well in terms of end of season fit at subarctic sites and temperate/continental sites in central Europe, systems which experience both temperature and light-limitation in both reality and the model. However, other sites in similar climates perform less well, showing the existing uncertainty and high site-specificity of which factor is more important in a co-limitation ecosystem . We show that the model is suitable for many herbaceous ecosystems with differing climatic conditions across the globe but its varying performance across similar ecosystems further shows that species- and site-specific adaptations to environmental conditions bring a high uncertainty into modelling grassland leaf turnover and may call for more adaptable responses. Still, our model provides a first step away from threshold-controls in LSMs and towards more process-based models.

Under future conditions where sites may move into more seasonally dry regimes, the dynamic turnover representation may be more adaptable than the previous threshold approach and it consistently performs better at simulating tropical savanna systems. Further, day length provides a hard threshold even as the climate gets warmer and many studies show that plants use it as signaling to induce leaf senescence . With the incorporation of day length, we can simulate future subarctic and tundra ecosystems better than the previous threshold model which often predicted very long growing seasons well into the light-limited winter, such as we have shown at US-Tol.

All herbaceous plants share the same leaf turnover in QUINCY and there is no current representation of different life strategies. To improve our current model, the triggers for leaf senescence in herbaceous plants across different climates need further attention, even though efforts over the last decade have been made to close this gap , but most studies only focus on single sites or species so that there is limited understanding of how day length, temperature and precipitation jointly determine leaf senescence across the globe. Even if this gap is slowly being filled, it will be very difficult to capture differing responses to seasonality caused by differences on a species-level with ecosystem scale data. Our continuous model is a first step towards addressing this issue, however, we also show that this needs further attention in the form of site-specific responses based on climate under the assumption that plants are adapted to the environment they grow in . Further, limited data availability for some of the sites introduced uncertainty into our results and reduced statistical power. A limited number of continuous data spanning multiple years combined with the frequency of extreme events over the last decades complicates the development of a general model for leaf turnover. Finally, many grasslands are intensively managed systems, which we do not consider in this study. Management of grasslands, through grazing, harvesting or fertiliser use, has a variety of implications for ecosystem functioning and since we do not consider them in our model, but compare it to EC data from these managed systems, our results were likely impacted by this as well.

The difference in performance between site-level evaluations (Table ) and climate-level aggregations (Table ) likely reflects the inherent challenge of representing diverse ecological strategies with a single parameterization. Potential future directions include the incorporation of a drought response parameter based on site level climatological and soil hydrological conditions, to represent local adaptation, so that sites with little soil moisture and temperature fluctuations would show a stronger drought response than seasonally cold or dry and (semi-) arid systems which are more adapted to these conditions. Currently fshedmax ensures shedding of leaves at maximum stress over 10 days. This fixed parameter could be replaced with a flexible moisture and temperature sensitive variable, similar to the models of and , to allow for different shedding rates under different climatic conditions. Yet, different plant responses to stress (tolerance, avoidance and escape) as described in will make the development of a general global response very difficult. Next to moisture and temperature it may be worth investigating the day length component of our model, as it currently virtually only impacts sites at high latitudes to simulate light starvation. However, plants use day length conditions of the environment to track seasons and induce senescence as a response to decreasing day length as a way to avoid harsher conditions of winter and found that more than a third of the variation of leaf senescence can be explained by hours of daylight in herbaceous species and up to 60 % for graminoids in China. An updated day length model may therefore expand into lower latitudes as well, though this may compromise the accurate representation of Mediterranean ecosystems where a lot of growth happens after the dry summer as the first rain occurs in autumn, which would then coincide with the day length model increasing its leaf shedding rate.