Introduction

Phenology is the study of periodic events, factors that influence their timing, and relationships among such events (Lieth 1974). Individual species, communities, ecosystem function and services, biogeochemical cycles, vegetation–climate feedbacks, and agriculture all respond to phenological cycles (Richardson et al. 2013; Piao et al. 2019). Vegetation phenology is one such periodic cycle of plant growth and is driven by three basic factors: sunlight, water, and temperature (Zhao et al. 2013).

Field observations of distinct events like vegetation leaf-out or bloom have historically been the basis of vegetation phenology research and still provide important phenological data (Leopold and Jones 1947; Menzel 2002). However, applications to study phenology from remotely sensed imagery have grown rapidly since imagery from sensors such as Landsat-1 and AVHRR became widely available (Henebry and de Beurs 2013), because such satellite imagery can detect broad spectral changes associated with vegetation green-up and senescence (Reed et al. 1994; Stöckli and Vidale 2004). This field became known as “land surface phenology” because these remotely sensed phenological changes represent the aggregate activity of vegetation in a given area, rather than the phenology of individual species (Hanes et al. 2014). Land surface phenology allows research across much broader spatial and temporal scales than traditional in situ observations and is particularly suited to examining phenological shifts over time and across regions.

Climate and phenology are closely linked, so phenology serves as a biological indicator of climate change (Cleland et al. 2007). Generally, temperature is the primary factor influencing spring vegetation phenology in temperate and mid to high latitude regions where growing seasons are warm and dormant seasons are cold (Walther et al. 2002; Linderholm 2006; Wu and Liu 2013; Fu et al. 2020). Land surface phenology studies strongly indicate an advancement in the timing of spring vegetation green-up globally in the last few decades, widely credited to rising temperatures (Parmesan and Yohe 2003; Linderholm 2006; Julien and Sobrino 2009; Richardson et al. 2013; Piao et al. 2019). This earlier spring timing is also supported by climatic indices of spring onset (Schwartz et al. 2006).

Climate change effects on autumn phenology have largely not received as much attention as effects on spring phenology (Gallinat et al. 2015), but impacts to autumn dynamics may be as significant and more varied (Walther et al. 2002; Richardson et al. 2013; Garonna et al. 2014; Liu at al. 2016). Climatic controls on autumn phenology are less clear than those to spring phenology, with interactions between temperature, moisture, and photoperiod playing roles (Estrella and Menzel 2006; Way and Montgomery 2015; Ren et al. 2018a; Fu et al. 2020). Broad analyses have generally shown that climate change has delayed autumn senescence, which, along with earlier spring green-up, has consequently lengthened growing seasons (Stöckli and Vidale 2004; Linderholm 2006; Julien and Sobrino 2009; Jeong et al. 2011; Ge et al. 2015).

Phenology research has largely focused on forested ecosystems (Richardson et al. 2013), so phenology impacts in grasslands may not exhibit the spring advancement and autumn delay often found in more global analyses. For example, Ying et al. (2020) found autumn senescence delays in Inner Mongolia forest and forest steppe, but advancement in non-forested ecosystems. Similarly, autumn senescence might be occurring later throughout much of the Northern Hemisphere, but earlier in the predominantly non-forested western USA (Piao et al. 2007). Studies specific to grasslands have also supported autumn senescence advancement (Li et al. 2016; Ren et al. 2018b), and warmer temperatures have been shown in experiments to advance grassland autumn senescence (Zavaleta et al. 2003). These studies and others have also found support of spring green-up advancement in grasslands (Piao et al. 2007; Zhang et al. 2013; Li et al. 2016; Chang et al. 2017; Ren et al. 2018b), in line with more global findings. However, some have found delays in grassland spring green-up (Reed 2006; Yu et al. 2010), or advances during some time frames and delays in others (Piao et al. 2011; Wu and Liu 2013).

Though species-specific phenology models at local scales are well advanced, generalized models of phenology at large extents can be improved (Piao et al. 2019). Jolly et al. (2005) developed a generalized growing season phenology model from photoperiod, minimum temperatures, and vapor pressure deficit to address phenological timing at model sites. Xin et al. (2015) expanded this work by modeling growing season start dates in United States (US) grasslands through multiple methods, including variants of the Jolly et al. (2005) growing season index. Their models achieved reasonable accuracy in predicting growing season onset across this large study area. However, they did not model growing season end, and consequently could not address vegetation senescence or growing season length. Hufkens et al. (2016) used PhenoCam imagery to model future dynamics in western USA and Canadian grasslands, and projected earlier spring green-up and later autumn senescence, consequently lengthening growing seasons.

Changes to spring green-up, autumn senescence, and overall growing season length have important implications for grasslands, including impacts on global carbon cycling (Piao et al. 2007; Richardson et al. 2013), wildlife migration (Monteith et al. 2011), and land uses such as grazing (Ren et al. 2018b). Models of future phenological changes, particularly those that address grasslands and autumn senescence, are lacking. Planning of future climate adaptation in these ecosystems could be greatly improved by such projections of future phenological impacts. To address this need, we developed a generalized model of climatic impacts on phenology in temperate US rangelands. We then projected this model through the year 2099 under six general circulation models and RCP 4.5 and 8.5 scenarios to address how future climatic conditions will influence rangeland growing season timing.

Methods

Study area

We limited our study area to locations defined as rangelands by Reeves and Mitchell (2011) within the western USA. We further restricted the study area to a subset of rangelands, as described in Input data. To summarize the results geographically, we evaluated our results throughout major ecological provinces (McNab et al. 2007) in the study area (Fig. 1). This area includes a diversity of shrubland and grassland environments, and much of it is managed by federal agencies such as Bureau of Land Management and US Forest Service.

Fig. 1
figure 1

Maps of the study area showing all rangelands and the subset included in modeling (a), and the major ecological provinces encompassing the included rangelands (b)

Full size image

Model overview

We sampled elevation, daily climate, and daily growing season determinations from 2001 to 2014 from more than 2000 random pixels throughout the study area. We utilized these data to train and validate models of annual growing season timing. We modeled start of season (SOS) dates, end of season (EOS) dates, and growing season length (GSL) through a two-step process (Fig. 2). We first employed a binomial generalized linear model to predict the probability that each day in a year fell within the growing season. This model was fit with daily climate and elevation data, and trained on USGS eMODIS growing season determinations.

Fig. 2
figure 2

Flowchart of modeling process. Models were fit with elevation, daily gridMET climate from 2001 to 2014, and USGS eMODIS phenology from 2001 to 2014. Future projections were made using the same models, with elevation and daily CMIP5-projected climate data from 2020 to 2099

Full size image

Next, using the set of daily growing season probabilities in each year, we needed to determine which dates represented growing season start and end. To accomplish this, we fit a sinusoidal model to the growing season probabilities, and determined when the fit line surpassed and fell below given growing season probabilities. Lastly, growing season length was calculated as the difference between the determined EOS and SOS dates. Future projections were made with the same models, using elevation and future projected climate as inputs.

Input data

We used phenology products from USGS eMODIS Collection 5 Terra Western as the measure of growing season timing from 2001 to 2014 in the study area (Jenkerson et al. 2010; Brown 2016; USGS EROS 2018). These phenology data are well-established metrics of growing seasons in the study area, and have been utilized in many studies of phenology (see Howard et al. 2012; Gu et al. 2013; Liebezeit et al. 2014; Boyte et al. 2015; Meier et al. 2015; Zhou et al. 2019). We utilized SOS (start of season time) and EOS (end of season time) data from eMODIS. These metrics are derived by calculating a time series of smoothed NDVI data, calculating a 14-day delayed moving average of NDVI from the smoothed NDVI values, and determining when these lines intersect (Reed et al. 1994; Meier and Brown 2014). These intersections indicate either a sharp increase or decrease in NDVI, signaling either the onset or end of the year’s growing season (Meier and Brown 2014). This process reduces sensitivity to temporary NDVI trend changes and measurement errors from factors such as cloud contamination (Reed et al. 1994). However, snowmelt can confound vegetation green up identified by this method, so some have produced alternative phenology indices without the influence of snowmelt (Delbart et al. 2006; Gallant et al. 2018; Wang et al. 2018). These indices may somewhat improve phenology determinations, but are still somewhat novel approaches and may be most necessary in evergreen forests, not grasslands (Beck et al. 2006; Gallant et al. 2018).

From the eMODIS SOS and EOS dates, we produced a binary record of whether dates fell within the year’s growing season. For example, if the SOS was day 90 and EOS was day 300 at a given pixel in a given year, days 1–89 were not within the growing season at that pixel in that year, days 90–300 were within the growing season, and days 301–365 were not within the growing season. Dates not within the growing season were coded as 0, and dates within the growing season were coded as 1.

We associated these binary growing season data with gridMET daily climate data at 4 km resolution (Abatzoglou 2013), including maximum temperature, minimum temperature, and downward shortwave radiation. Growing degree days (GDD) were calculated daily from mean daily temperature minus a base temperature of 0ºC, and were cumulatively summed over the year to calculate accumulated growing degree days (AGDD). The base temperature of 0 ºC is commonly used in temperate grasslands (Frank and Hofmann 1989; Preister et al. 2019). While individual species have varying base temperatures, 0 ºC is suitable for temperate grasslands as a whole, and manipulations of base temperature have little effect on GDD correlations to phenology (Romano et al. 2014). Exploratory data analysis showed the AGDD value of 2500 was strongly associated with growing season end throughout our study area, so we calculated whether this threshold had been surpassed and if so by how much, and included this as a variable in modeling. Vapor pressure deficit (VPD) was estimated from the difference between saturated vapor pressure at daily maximum and minimum temperatures by the following equation from Lobell et al. 2014:

$$ {\text{VPD}} = \left( {0.6107*e^{{17.269T_{\max } / \left( {237.3\; + \;T_{\max } } \right)}} } \right) - \left( {0.6107*e^{{17.269T_{\min } / \left( {237.3\; + \;T_{\min } } \right)}} } \right). $$
(1)

Phenology in annual grasslands and deserts is highly variable, with the potential for multiple growing seasons annually and start of season occurring as early as autumn in some years (Reed et al. 1994). To assess only locations with a single annual growing season and avoid such highly variable regions, we removed locations with start of season dates earlier than January 1 or later than May 30 in any years (Fig. 1a). This filtering also eliminated areas which may have experienced major vegetation shifts or disturbances during the study period. This was desirable because we were interested only in the effects of climate, not vegetation change, on phenology timing.

Binomial growing season probability model

We used a binomial generalized linear model in R (R Core Team 2018) to fit daily climate and elevation data to daily growing season determinations. We included seven variables in the binomial model (Table 1). This binomial model predicted the probability that a given day fell within the year’s growing season for all included pixels.

Table 1 Variables included in the generalized linear model
Full size table

The model was fit with a logit link with the following equation, where variable notations match those in Table 1, β0 represents the equation’s intercept, and β1 through β6 represent variable coefficients (shown in Table S1). All climatic variables correspond to measurements from a single day, d, and µ represents the probability that the given day fell within the growing season:

$$ \ln \left( {\frac{\mu }{1 - \mu }} \right) = \beta_{0} + \beta_{1} {\text{SRAD}}_{d} *\beta_{2} {\text{VPD}}_{d} *\beta_{3} {\text{TMAX}}_{d} *\beta_{4} {\text{AGDD}}_{d} + \beta_{5} {\text{Elev}} + \beta_{6} {\text{AGDD}} - 2500_{d} $$
(2)

Of the 2000 random pixels from which we sampled climate and growing season data, we randomly selected a small subset (14% of pixels) for training the model. Data from the remaining pixels were reserved for model validation. We further restricted the input data by selecting 9 years of climate and phenology records for training the model, and reserved 5 years for model validation. The training years were manually selected to include years with varying phenological conditions (e.g., years with early, late, and typical growing seasons across the study area), training the model on a wide set of conditions. The years reserved for model validation were: 2003, 2005, 2008, 2012, and 2014.

The model was fit using only the training pixels, with climate and phenology data from only the training years. Model accuracy was evaluated using validation pixels, with climate and phenology data from validation years.

Sinusoidal growing season dates model

The binomial model output daily growing season probabilities for each year. From these daily probabilities, growing season start and end dates had to be determined. To accomplish this, we fit a sinusoidal model to the set of daily growing season probabilities. Phenological data are well fit by sinusoidal models because of their strong periodicity (see Hogg et al. 2000; Girondot et al. 2006). We applied the following second harmonic sinusoidal model (similar to one outlined in Sapiano et al. 2012), where GS Prob is growing season probability, d is the day of the year, and w is the cycle’s frequency (2π/365 days). Β0 represents the model intercept, Β1–Β4 represent variable coefficients, and ε represents error:

$$ {\text{GS Prob}}\left( d \right) = \beta_{0} + \beta_{1} \sin \left( {d*w} \right) + \beta_{2} \cos \left( {d*w} \right) + \beta_{3} \sin \left( {d*2w} \right) + \beta_{4} \cos \left( {d*2w} \right) + \varepsilon . $$
(3)

The sinusoidal model was fit at each pixel in each year. The resulting set of fit probabilities from the model was then analyzed to determine the day when the fit probability first surpassed 0.5, and the day when it next fell below 0.45. These dates were interpreted as the start of season date, and the end of season date, respectively, at a given pixel in a single year. Model validation showed these thresholds best aligned with eMODIS SOS and EOS dates throughout the study area.

Though other methods can allow more direct biophysical interpretations of factors beyond growing season start and end (Garroutte et al. 2016; Morisette et al. 2021), these interpretations were not possible in our case because our model fit modeled growing season probability rather than NDVI, EVI, or other direct vegetation indices.

Projecting future growing season dates

We projected future growing season dates by running the binomial growing season probability and sinusoidal growing season dates models with future projected climate. We utilized CMIP5 projections (Taylor et al. 2012) downscaled using the MACA method (Abatzoglou and Brown 2012). The following six global climate models (GCMs) under RCP 4.5 and 8.5 scenarios were resampled to a common 4-km resolution and utilized in modeling: BCC-CSM1.1 (m), CNRM-CM5.1, HadGEM2-ES, IPSL-CM5A-MR, MRI-CGCM3, and NorESM1-M. These models were chosen because they capture a range of possible future climatic conditions and are used in the Forest Service 2020 Resources Planning Act (RPA) Assessment (Joyce and Coulson 2020), which utilizes these projections.

For each GCM/RCP combination, we used the binomial model to predict daily growing season probabilities from 2020 to 2099, then used the sinusoidal model to determine SOS and EOS dates from these probabilities for each pixel and in each year. Growing season length (GSL) was calculated as the difference between EOS and SOS dates. We compared the future projected growing season dates to the mean growing season dates predicted by the model at each pixel during the baseline period (2001–2014) to quantify the projected growing season changes relative to current conditions. We evaluated changes in SOS date, EOS date, and GSL.

We obtained ensemble projections of change for each year by calculating the mean of projected changes from all GCMs in each year under both RCP scenarios. For example, we calculated the mean of projected changes from the six GCMs under RCP 4.5 in the year 2030 to obtain the ensemble projection of change for this year under RCP 4.5. Lastly, we calculated the mean of ensemble projections across decades (from 2030 to 2039, 2040–2049, etc.) to determine more stable projected changes without annual climatic variability.

Since our model was trained on eMODIS land surface phenology metrics, our results should be interpreted as projected changes in these phenology metrics. These products broadly capture vegetation life-cycle timing at the landscape scale (USGS EROS 2018).

Relative magnitude of growing season start and end change

We followed a procedure from Garonna et al. (2014) to compare the relative magnitude of SOS and EOS change at pixels. The following equation calculates the C-Index, where ∆SOS represents change in SOS date and ∆EOS represents change in EOS date:

$$ C - {\text{Index}} = - 1 + \frac{{2*{\text{abs}}\left( {\Delta {\text{SOS}}} \right)}}{{{\text{abs}}\left( {\Delta {\text{SOS}}} \right) + {\text{ab}}s\left( {\Delta {\text{EOS}}} \right)}}. $$
(4)

The C-Index is bound from − 1 to 1. Negative values indicate EOS change is of greater magnitude than SOS change at a location, and positive values indicate SOS change is of greater magnitude than EOS change. We calculated the C-Index at all pixels, considering the change in SOS and EOS timing in the 2090–2099 period relative to the reference period (2001–2014).

Model validation

We first evaluated the accuracy of the binomial growing season probability model to predict whether days fell within the growing season. Binomial model predictions of at least 0.50 were interpreted as within the growing season, and predictions less than 0.50 were interpreted as not within the growing season. These predictions were compared to the original eMODIS phenology data. We evaluated the model’s accuracy among training pixels in training years, validation pixels in training years, and validation pixels in validation years. These multiple lines of validation assessed whether the model was overfit to the input data and performed similarly regardless of the year and location of input data.

Next, we evaluated the accuracy of the sinusoidal growing season date model to predict growing season dates from the binomial model predictions. We compared model predictions to the eMODIS data, and evaluated root mean square error (RMSE), mean absolute error (MAE), and absolute bias of the model in predicting SOS date, EOS date, and GSL. Absolute bias was calculated as the mean of all errors (positive and negative).

Results

Binomial growing season probability model

The binomial model predicted daily growing season probability with high accuracy, and results were very similar between training and validation data. Among validation pixels in validation years, the model’s accuracy was 87.05%. Therefore, in 87.05% of instances, the model correctly predicted whether a date fell within the growing season or not (with probabilities less than 0.50 considered not within the growing season, and probabilities of at least 0.50 considered within the growing season). Prediction accuracy among training pixels in training years was 87.27%, and prediction accuracy among validation pixels in training years was 87.59%. Type III ANOVA showed all variables included in the binomial model were highly significant, besides the interaction of solar radiation, vapor pressure deficit, and accumulated growing degree days (Table S1).

Sinusoidal growing season dates model

The sinusoidal model fit daily growing season probability predictions closely. Figure 3 shows the daily growing season probabilities predicted by the binomial model, and the sinusoidal model fit to daily growing season probabilities in the year 2001 for a single pixel located at 40.979 N, 115.939 W (near Elko, Nevada). In this instance, the SOS date was predicted as day 82, while eMODIS showed the SOS date was day 74, an error of 8 days. The predicted EOS date was 318, while eMODIS showed EOS date was day 321, an error of − 3 days. Predicted GSL was 236 days, and GSL from eMODIS was 247 days, an error of − 11 days. Negative errors indicate the model predicted earlier SOS/EOS dates than eMODIS, or a shorter growing season length.

Fig. 3
figure 3

Daily growing season probabilities predicted by the binomial model at a pixel near Elko, Nevada in 2001 (points). Blue line is the sinusoidal model fit to daily growing season probabilities. Dashed vertical lines indicate the SOS and EOS dates determined by the sinusoidal model, and solid vertical lines indicate the SOS and EOS dates from eMODIS at this pixel in 2001. Green lines indicate SOS, and red lines indicate EOS

Full size image

Growing season dates were modeled with reasonable accuracy overall. Considering only validation years withheld from modeling, the mean MAE of modeled SOS dates across the study area was 19.46 days, mean RMSE was 22.85 days, and mean absolute bias was -5.24 days (Fig. S1). EOS dates were modeled with similar accuracy. Mean MAE was 21.80, mean RMSE was 27.04, and bias was 0.97 (Fig. S2). Errors modeling GSL were slightly higher. Mean MAE of GSL was 23.69, mean RMSE was 27.92, and bias was 6.21 (Fig. S3).

Future projected growing season dates

Future growing season dates projected by the models showed significant changes relative to the historical reference period (2001–2014). SOS dates were projected to advance throughout the region, and advance more dramatically in RCP 8.5 than 4.5 (Fig. 4). In results from the ensemble model, approximately half of the region was projected to experience SOS advancement of at least 5 days by the 2030–2039 period in both RCP 4.5 and 8.5 (Fig. 4a, d). By the 2090–2099 period, start of season dates were projected to advance at least 5 days in nearly the entire region, with more intense advancement under the RCP 8.5 scenario (Fig. 4c, f). In this latest time period, mean advancement in the RCP 4.5 ensemble was 9.88 days, while mean advancement in the RCP 8.5 scenario was 16.82 days. These correspond to a rate of change through 2090–2099 of 0.12 days per year in RCP 4.5, and 0.21 days per year in RCP 8.5. SOS advancement was greatest throughout southern Idaho, eastern Washington, and eastern Oregon, with SOS projected to arrive as much as 30 days earlier by the end of the century in these areas.

Fig. 4
figure 4

Mean projected start of season date change among the ensemble of all six included GCMs, among RCP 4.5 (a–c) and RCP 8.5 (d–f) scenarios

Full size image

EOS dates were also mainly projected to advance, and more dramatically in RCP 8.5 than 4.5 (Fig. 5). Overall, EOS advanced by greater magnitude than SOS, and by the 2090–2099 period most areas had EOS advancement of at least 10 days. Mean advancement in 2090–2099 was 12.47 days in RCP 4.5, and 24.17 days in RCP 8.5 (Fig. 5c, f). The respective rates of change were 0.16 days per year for RCP 4.5, and 0.30 days per year for RCP 8.5. However, isolated areas in the Rocky Mountains demonstrated EOS delays in all time periods and both RCPs, contrary to results elsewhere.

Fig. 5
figure 5

Mean projected end of season date change among the ensemble of all six included GCMs, among RCP 4.5 (a–c) and RCP 8.5 (d–f) scenarios

Full size image

GSL was not projected to change as much as either SOS or EOS (Fig. 6). GSL change was modest until the 2060–2069 period, when mean GSL contraction was 2.50 days in RCP 4.5 and 5.06 days in RCP 8.5 (Fig. 6b, e). Changes through the 2090–2099 period were similar, but with more extreme contraction in southern areas in the RCP 8.5 scenario, and with mean GSL shortening by an additional 2.29 days. Mean GSL shortening in this period was 2.59 days under RCP 4.5 (a rate of 0.03 days per year), and 7.35 days under RCP 8.5 in 2090–2099 (a rate of 0.09 days per year). In all time periods and RCP scenarios, both GSL contraction and expansion were evident in some areas, with expansion in isolated areas in the Rocky Mountains.

Fig. 6
figure 6

Mean projected growing season length change among the ensemble of all six included GCMs, among RCP 4.5 (a–c) and RCP 8.5 (d–f) scenarios

Full size image

Considering projected changes in individual ecological provinces allows for finer comparisons between regions (Fig. 7). Though isolated areas had EOS delays projected (Fig. 5), no ecological province as a whole was projected to experience a delay in EOS (Fig. 7b). The Middle Rocky Mountain Steppe and Southern Rocky Mountain Steppe had broad spreads in EOS dates due to the areas projected to experience EOS delays, but mean EOS in these provinces as a whole was still projected to advance. By the end of the century, mean EOS advancement approached 30 days in some provinces under RCP 8.5, but was closer to 10 days under RCP 4.5.

Fig. 7
figure 7

Mean projected SOS (a), EOS (b), and GSL (c) change in ecological provinces among the ensemble of all six included GCMs, for both RCP 4.5 and RCP 8.5 scenarios. Lines represent mean results by year, and shading indicates one standard deviation from the mean

Full size image

SOS changes were fairly consistent between provinces (Fig. 7a). Unlike for EOS, no areas were projected to experience later SOS. By the end of the century, SOS advancement was around 10 days for all provinces in RCP 4.5, and nearly 20 days in RCP 8.5. For both EOS and SOS change, projections for RCP 4.5 and 8.5 did not strongly differ from one another until approximately 2060.

Large GSL shifts were not projected for any ecological provinces as a whole (Fig. 7c). In all provinces, GSL was shorter in the RCP 8.5 projections than RCP 4.5. Mean GSL decreases in RCP 8.5 ranged from 5 to 10 days in all provinces, while nearly no GSL change was projected under RCP 4.5. The Middle Rocky Mountain Steppe and Southern Rocky Mountain Steppe had broad GSL change spreads, and indicated a slightly lengthened growing season by the end of the century under RCP 4.5.

Relative magnitude of growing season start and end change

In both RCP 4.5 and 8.5, the C-Index, measuring the relative magnitude of SOS and EOS change, was negative in most of the study area (Fig. 8). This indicates EOS change was of greater magnitude than SOS change in most locations. The mean C-Index throughout the study area was -0.13 in RCP 4.5 and -0.17 in RCP 8.5. However, in both RCP 4.5 and 8.5, isolated areas of the Rocky Mountains had positive C-Index values, indicating greater changes in SOS than EOS.

Fig. 8
figure 8

Mean projected C-Index, considering mean SOS and EOS change from 2090 to 2099 relative to the reference period (2001–2014)

Full size image

Discussion

Projected growing season changes

We trained a model on eMODIS phenology products, which represent the timing of vegetation growing seasons at a landscape scale, to project future changes in growing season timing and length. Our results indicated both start of season and end of season timing will occur earlier in response to future climate change in the majority of our study area (Figs. 4, 5). Overall impacts to growing season length were less pronounced, but tended toward slight contraction in most areas, with some mountainous areas experiencing growing season extension (Fig. 6).

Studies of grassland growing season shifts in response to recent climate change have typically shown advancement in spring green-up (Piao et al. 2007; Zhang et al. 2013; Li et al. 2016; Ren et al. 2018b). Impacts to autumn senescence timing are more varied, but studies in grasslands have mainly found advancement (Li et al. 2016; Ren et al. 2018a, b; Ying et al. 2020).

However, comparable studies to ours projecting future phenological shifts in grasslands are rare. A study in European grasslands found similar results, with a widespread shift toward earlier start of season timing largely outweighed by a corresponding advancement in end of season timing, resulting in growing season length contraction (Chang et al. 2017). This study found increasing spring temperatures advanced spring green-up, while increased summer aridity from high temperatures and low moisture advanced autumn senescence. These effects have also been demonstrated experimentally (Zavaleta et al. 2003). A recent study found experimental warming of cheatgrass had the same results, with earlier flowering, earlier senescence, and shortened growing season (Howell et al. 2020).

However, another study of future phenology in our study area projected earlier start of season, later end of season, and longer growing season length (Hufkens et al. 2016). This study projected reductions in productivity and vegetation cover during summer drought, which were then reversed with increased activity during autumn. Our sinusoidal model did not allow for multiple growing seasons in a year, and assumed vegetation could not enter the growing season again once growing season probability began to decrease. This may account for the discrepancy between our results and those of Hufkens et al. (2016).

Some research in forests and grasslands has shown that start of season and end of season dates are positively correlated, and thus likely to shift in the same direction, as we found (Fu et al. 2014; Keenan and Richardson 2015; Wu et al. 2016). Explanations for this relationship are numerous, and more research into this phenomenon is needed. Possibilities include limits of leaf longevity and programmed cell death (Keenan and Richardson 2015), or plant circadian rhythms (McWatters and Devlin 2011; Zu et al. 2018). Other hypotheses suggest that carbon-sink limitations force vegetation senescence when maximum carbohydrate reserves are reached, promoting earlier senescence following earlier green-up (Fu et al. 2014; Zani et al. 2020).

Though our model did not explicitly incorporate this effect, the constraint imposed by accumulated growing degree days worked similarly in our model. However, we found end of season timing was delayed while growing seasons lengthened in higher elevation grasslands in the Rocky Mountains, in contrast to other areas. This may be because these cooler, moister grasslands are temperature limited so warmer temperatures can extend their growing season, whereas moisture limitations are more significant in the drier, lower elevation grasslands through the rest of our study area (White et al. 1997).

Importance of end of season timing

Changes in end of season timing, and autumn phenology generally, have not received as much attention as start of season or spring phenology (Gallinat et al. 2015; Piao et al. 2019). Our C-Index calculation suggested end of season timing will change more than start of season timing throughout most of our study area (Fig. 8), supporting the importance of considering autumn phenology in any consideration of phenology impacts. Though end of season change may not always be of greater magnitude than start of season change, it is clearly an important aspect of phenology research that cannot be ignored. Autumn phenology impacts are generally more varied than spring phenology impacts (Walther et al. 2002; Richardson et al. 2013), and controls on autumn phenology are less clear than those on spring phenology (Ren et al. 2018a; Fu et al. 2020), so better discernment of these controls by ecosystem and region are still needed. Our results also support the finding that autumn phenology impacts are more varied than spring impacts, with both end of season advancement and delay projected in our results.

Impacts to ecosystems and ecosystem management

Livestock grazing is a dominant land use in our study area and would be impacted by phenological shifts, requiring adaptation for effective management (Rojas-Downing et al. 2017; Ren et al. 2018b). Earlier growing seasons will likely force ranchers to bring livestock to pasture earlier in the year, but contracting growing season length could reduce the amount of time livestock can graze vegetation within the growing season, potentially increasing ranchers’ demands for supplemental winter feed.

Changing climate will also have direct impacts on livestock grazing which interact with phenology. For example, heat stress induced by high temperatures and humidity can reduce cattle weight gain and health, and increase water demands (Howden et al. 2008; Polsky and von Keyserlingk 2017). Indirect effects of heat on livestock health, including reduced immune response and increased vector-borne illness spread are also possible (Nardone et al. 2010). These factors may challenge grazing management and reduce ranching profitability (Neibergs et al. 2018). On the other hand, higher winter temperatures and reduced snow cover could increase access to forage, even where growing season length contracts (Chang et al. 2017). Climate change will also impact forage quantity (Polley et al. 2013) and forage quality (Dumont et al. 2015), both crucial for livestock grazing.

Phenological mismatch and disruption of species interdependencies may result if species adjust phenological timing differently in response to climate change (Rafferty et al. 2013). Though global shifts in phenological synchrony are apparent (Kharouba et al. 2018), evidence of phenological mismatch in mutualistic relationships is rare (Bartomeus et al. 2011). Mismatch is more common in antagonistic or competitive relationships where species have not co-evolved common phenological strategies (Both et al. 2009; Renner and Zohner 2018).

Phenological plasticity affects species fitness, and the ability to adapt phenology could contribute to community compositions shifts in the future (Reed et al. 2013; Duputié et al. 2015). In the western US, for example, annual and cool-season grasses are more closely tracking climate change than perennial grasses or warm-season grasses (Munson and Long 2017). Similarly, invasive plants track climate change and adjust phenological timing more closely than native species (Wolkovich et al. 2013; Wolkovich and Cleland 2014). There could be significant implications if these phenological differences lead to plant community shifts, particularly as invasive annual cool-season grasses such as cheatgrass (Bromus tectorum) already pose significant management challenges in our study area (Knapp 1996). However, to what degree both phenotypic plasticity and genetic adaptation contribute to phenological tracking is unclear (Franks and Weis 2008), which may influence how significantly species phenology can adapt in the future.

Model error

Fitting start of season dates across the study area, our modeling method achieved a mean MAE of 19.46 days and mean RMSE of 22.85 days. End of season dates and growing season length were modeled with slightly higher error (respectively, MAE of 21.80 and RMSE of 27.04, and MAE of 23.69 and RMSE of 27.92. Xin et al. (2015) also modeled start of season timing in western U.S. grasslands using eMODIS phenology data, and their best models outperformed ours, with RMSE of 16.4 days. This indicates the potential to improve our model by incorporating either a modified growing degree day or accumulated growing season index term, as they used (Xin et al. 2015). However, Xin et al. did not address end of season timing, whereas we employed a unified model of growing season start and end, which may constrain model performance.

Errors in our model were comparable to or better than other available examples fit to satellite-derived phenology metrics. With a double logistic function fit to NDVI at two Inner Mongolia grasslands, Ren et al. (2018a) fit start of season timing with MAE from 14.7 to 15.2 days. End of season, though was fit with MAE from 32.1 to 48.1, significantly higher errors than our model. Incorporating ground-measured gross primary production measurements at five sites, Ren et al. (2018b) greatly reduced errors (RMSE of 12.9 days for start of season and 13.0 days for end of season). Therefore, incorporating remotely-sensed gross primary productivity data could improve our model, though such estimates at large spatial scales may be significantly less accurate than those from ground-based observation sites. Furthermore, future projections of such data may not be reliable, limiting the ability to project a model based on gross primary productivity into the future.

Limitations

We used a relatively simple model which projected growing season timing based on readily-available future climatic data, and did not consider factors beyond climate which could impact phenology. Carbon dioxide (CO2) concentrations were not included in our model, but CO2 concentration influences vegetative productivity in numerous ways (Polley et al. 2013), and increased concentrations may lengthen growing seasons through increased species complementarity (Reyes-Fox et al. 2014). Disturbances associated with climate change, such as wildfire, were also not included in our model, but could alter plant communities and phenology timing as well (Wang and Zhang 2020). Growing season length itself may be mediated by plant diversity, with greater diversity promoting a longer growing season (Oehri et al. 2017), but we could not include this effect.

Similarly, we did not consider phenological implications of shifts in vegetation type. Greater composition of warm-season grasses would shift growing season timing later in the year, but climate change impacts on the composition of warm-season and cool-season grasses are still unclear due to the offsetting impacts of increased CO2 concentrations and increased temperatures (Ehleringer et al. 1997; Morgan et al. 2011). Our study area has also experienced shrubland and tree encroachment in recent years (Miller and Rose 1999; Knapp et al. 2008a, b), and continued shifts from grass to these species could impact phenology timing as well.

We could not analyze changes in net primary productivity (NPP), gross primary productivity (GPP), or other carbon flux measures in our model, which have important climate change implications (Cao and Woodward 1998). Many have found a positive relationship between NPP and growing season length (Myneni et al. 1997; Zhou et al. 2001; Bao et al. 2019), but this is not true in all circumstances, with differences in climatic regimes and drought occurrence playing a role (Wu et al. 2016; Wang et al. 2017). Importantly, net ecosystem productivity (NEP), which is most important from a global carbon perspective, may not increase even when GPP and NPP do, because increased soil carbon decomposition can offset increased GPP and NPP (Piao et al. 2007).

Our model assumed a single spring green-up and autumn senescence in each year. Measures like productivity vary throughout the growing season (Piao et al. 2007; Hufkens et al. 2016), but these variations and pulses in growth are not evident in our model since we modeled only the timing of growing season start and end. Climate change is likely to induce less frequent and more intense precipitation events (Trenberth 2011), potentially altering soil moisture dynamics and subsequent vegetation growth pulses in semi-arid systems (Huxman et al. 2004; Knapp et al. 2008a, b). These pulses in growth may be particularly important in autumn, as moisture limitations increase and isolated precipitation events may sustain or renew growth. eMODIS phenology products may be improved by accounting for false green-up determinations from NDVI increases following snowmelt (Beck et al. 2006; Wang et al. 2018). Though timing of growing season start in our model could be too early due to the effects of snowmelt on NDVI, we were most concerned with trends through time which should be less susceptible to these errors, rather than growing season dates themselves.

Conclusion

Phenology exerts significant control on many ecological systems, and the timing of phenological events is strongly influenced by climate. Vegetation phenology, including the timing of spring green-up and autumn senescence, is already shifting in response to climate change, and will continue to shift in the future. Our findings suggest climate change will promote earlier spring green-up and earlier autumn senescence timing in temperate U.S. rangelands. We found autumn advancement was of greater magnitude than spring advancement in most locations, which would consequently shorten overall growing season length in most places. While many phenology studies in forested ecosystems have shown growing seasons have lengthened as spring advances and autumn is delayed, findings in grasslands have mainly shown that both spring and autumn timing are advancing. The projected growing season changes we found would have important effects on livestock grazing, species fitness, and overall ecosystem function.