Abstract
Background and goals Climate change is a growing concern for winegrape producers worldwide. Increasing temperatures may accelerate grapevine development, which can change the timing of key phenology stages and affect final crop quantity and quality. Many researchers have explored the relationships between grapevine physiology and temperature-based phenology models. However, these models perform inconsistently from one region to another, especially when models calibrated in one context are applied where pedoclimatic factors and vineyard management methods differ.
Methods and key findings Temperature-based phenological models were calibrated using observed grapevine phenology dates from 2012 to 2021 for 18 commercial Pinot noir vineyards across Oregon’s Willamette Valley. Model performance during cross-validation was assessed using root mean squared error, model efficiency, and % accuracy. Resulting models were used to project phenology dates from 2020 to 2100 under four climate change scenarios. By the end of century, mean bloom, veraison, and harvest dates were projected to occur 1 to 2 wk (low-emissions scenarios) and 3 to 4 wk (high-emissions scenarios) earlier than current averages. Budbreak dates are projected to advance ~10 days for most emissions scenarios.
Conclusions and significance These results represent valuable information that Willamette Valley winegrape producers can use to anticipate challenges and make informed decisions to address potential long-term climate change impacts.
Introduction
There is growing concern in the wine industry as global air temperatures are projected to increase over future decades (IPCC 2021). Higher growing season temperatures are particularly concerning due to the relationship between increasing air temperature and more rapid grapevine (Vitis vinifera L.) growth stages (Menzel et al. 2006). The timing of key phenology stages such as budbreak, bloom, veraison, and harvest are important to producers as they affect harvest yields, crop quality, and vineyard and winery management (van Leeuwen and Seguin 2006, Santos et al. 2020, Zhu et al. 2020, Cameron et al. 2022). Vine damage risk assessment for events like spring frosts, poor fruit set, or poor ripening conditions are directly linked to the dates of key phenology stages and their corresponding meteorological conditions, including cold-related tissue damage/necrosis (Poni et al. 2022), rain and wind causing poor fruit set (Zhu et al. 2020), and severe heat causing sunburn or changes to berry chemistry (Camps and Ramos 2012, Rienth et al. 2016). As climate change continues to affect winegrape production and vineyard management decisions, the ability to model grapevine phenology under projected climate scenarios will enable producers to assess future opportunities and challenges, such as adapting current practices, migrating vineyard developments to new regions, or planting varieties with traits more suitable to the changing climate (Parker et al. 2013, 2020).
Plant phenology models typically identify the date of key growth stages based on daily air temperature data. Previous works describe the relationship between daytime air temperature and photosynthesis (Ferrini et al. 1995, Greer and Weedon 2012, Zufferey et al. 2000). Modeling grapevine phenology using temperature data was first explored using growing degree day (GDD) models such as spring warming (de Reaumur 1735). Recent GDD approaches often used 10°C as the base temperature (Winkler et al. 1974) or used the Huglin Index (Huglin 1978). These early modeling approaches are accurate, simple, and easy to apply. More sophisticated modeling approaches like the Wang (Wang 1960), Wang-Engel (WE; Wang and Engel 1998), and best sigmoidal (SIG; Chuine et al. 1999) models incorporate nonlinear responses to grapevine phenology based on optimal and suboptimal temperature ranges. These nonlinear models replicate the nonlinear response of grapevine photosynthesis depending on optimal/suboptimal temperature ranges (Ferrini et al. 1995, Greer and Weedon 2012, Zufferey et al. 2000). More recent advancements involve modeling grapevine dormancy over the autumn/winter period to more accurately project budbreak in spring. Models such as Richardson (Richardson et al. 1974), smoothed-Utah (SU; Richardson et al. 1974, Bonhomme et al. 2010), and UNIchill (Chuine 2000) use chilling units to characterize when tree fruit transitions from endodormancy to ecodormancy. The ecodormancy phase is characterized by daily heat units (referred to as ‘forcing units’). This sequential application of chilling units and forcing units to modeling efforts may more accurately characterize grapevine phenology in the context of climate change (Fila et al. 2014), since warm winters with fewer chilling units may delay grapevine endodormancy release. Additionally, nonlinear forcing unit accumulation will likely better characterize the acceleration of the ecodormancy phases due to warmer springs and excessive heat during summer causing slowing phenology. However, depending on the criteria adopted in defining the calibration data set, these complex modeling frameworks are more susceptible to overfitting the training data set (Fila et al. 2014) during parameter calibration, making them less accurate when applied to environmental contexts outside the domain of applicability defined by the calibration data set, such as other vineyards or regions.
Despite improvements to grapevine phenology models over recent decades and increased availability of phenology data sets for calibration (Yiou et al. 2012, Parker et al. 2013, García de Cortázar-Atauri et al. 2017), key challenges remain for accurate and robust phenology modeling. Models applied to future weather conditions assume stationarity of the relationship (Witt et al. 1998), but this assumption depends on the variability of environmental conditions in the calibration data set, which may affect projection accuracy if those conditions differ. Second, the genetic diversity within V. vinifera phenotypes (Wolkovich et al. 2017) requires that model parametrization be undertaken separately for each variety, since varieties may possess different forcing unit requirements before reaching the same phenology stage (Parker et al. 2013). Producer decision-making also plays an important role, as certain winegrape growing regions have well-established traditions or regulatory frameworks that drive vineyard location and management practices, like site selection, training system (Trought et al. 2017), or leaf pulling (Parker et al. 2014), and that may affect phenology. Reconciling the physiological and production diversity for grapevines and vineyards into a universal and effective phenology model is impossible, but efforts should be made to establish a framework that is customizable to other varieties, locations, and production scenarios.
Regional specificities, particularly climate and soils, heavily influence grapevine phenology. Arid winegrape growing regions with soils that have low available water-holding capacity and limited water supply tend to have smaller vines with lower leaf area-to-fruit mass (Pou et al. 2011, Parker et al. 2014, Pérez-Álvarez et al. 2021), due to less available water to sustain photosynthesis and active vegetative development (Lereboullet et al. 2013). Cool and humid regions with soils that have high fertility and high available water-holding capacity typically face challenges such as excessive vigor, pest pressures, or insufficient growing season length and air temperatures for adequate ripening (Shaw 1999, Cortiñas Rodríguez et al. 2020).
Oregon’s Willamette Valley American Viticultural Area is characterized as a cool climate, with 1181°C GDD from 1 April to 31 Oct and annual rainfall totals near 1270 mm, yet little-to-no rainfall (~70 mm) is received during summer (June through August), according to 1991 to 2020 climate normals (PRISM Climate Group, https://prism.oregonstate.edu/). These environmental conditions and soils with good water-holding capacity enable most vineyards in the Willamette Valley to grow vigorous canopies despite dry summers. Nearly 70% of vineyards in the region are planted to Pinot noir (IPRE 2023), a cool-climate variety known for small clusters and low yield potential (Castagnoli and Vasconcelos 2006). However, crop reduction is practiced by winegrape producers annually to hasten ripening (Uzes and Skinkis 2016) and avoid losses in fruit quality and quantity in cool, wet autumns. These pedoclimatic conditions and management challenges differentiate the Willamette Valley from other Mediterranean grapegrowing regions and are worth exploring in the context of climate change, due to greater water and land resource scarcity in neighboring winegrape growing regions that are currently greatly affected by drought, extreme heat, and wildfires (California and Washington). Winegrape producers within these regions are interested in Oregon for future viticultural land investments.
Due to the long lifespan of vineyards, forecasts of grapevine performance are necessary to maintain a long-term, sustainable, and robust winegrape industry. A phenology model framework for Willamette Valley Pinot noir was developed to address the unique properties and limiting factors of the region. The aim was to calibrate phenology models that possess adequate complexity to capture the nonlinear relationship between grapevine growth and phenology, as well as to guard against overfitting through robust cross-validation (CV) procedures and adequate model selection criteria, and to apply the resulting modeling framework to four climate change scenarios. This research intends to provide producers with resources to inform their planning and decision-making for the future of Oregon viticulture. Equipped with these climate change projections on Pinot noir phenology, producers may adapt vineyard management strategies, plan new vineyard developments, or pursue new varieties or growing areas.
Materials and Methods
Process-based phenology models that input average daily temperature data and output projected phenology dates in Julian day of year (DOY) were used. Recorded phenology dates from commercial vineyards across the Willamette Valley were used for ground-truthing to calibrate model parameters, thereby improving the models’ predictive capabilities in this pedoclimatic context. Weather data were then applied from modeled climate change scenarios to explore how Pinot noir phenology in the Willamette Valley might change between 2016 and 2100.
Observed field data
Field observations of grapevine (V. vinifera L. Pinot noir) phenology stages were recorded over a 10-yr period at a maximum of 18 commercial vineyard sites located within Oregon’s Willamette Valley (Figure 1). Dates were recorded for the time point at which vines reached 50% budbreak, 50% bloom, and 50% veraison, as defined by the modified Eichhorn-Lorenz (E-L) scale (Coombe 1995), from 2012 to 2021 (Table 1). Data were obtained from vines within the same 0.4 to 1.2 ha area of each vineyard each year to maintain consistency and reduce confounding factors. Commercial harvest dates and berry total soluble solids (TSS; Brix) were also recorded for each vineyard. TSS were obtained by collecting six or more fruit samples across the vineyard observational area at harvest, with each sample consisting of 20 clusters. Fruit was brought to the Oregon State University Viticulture Lab, destemmed manually, and berries were pressed to juice. TSS were measured using a refractometer (Model 300051, Sper Scientific).
Map of Oregon’s Willamette Valley showing vineyard locations where phenology field data collection occurred during 2012 to 2021. A total of 18 commercial vineyard sites were used in the project, but all of them were not necessarily included in all 10 years.
Number of observations across commercial Pinot noir vineyards in Oregon’s Willamette Valley by year for each phenological stage modeled. Budbreak, bloom, veraison, and harvest correspond to 50% of each phenology stage as defined by the modified Eichhorn-Lorenz scale (Coombe 1995).
Phenology modeling
Budbreak, bloom, and veraison dates were modeled as described (Morales-Castilla et al. 2020), which is well-suited for characterizing grapevine phenology under climate change scenarios. Budbreak modeling used the sequential application of the SU chilling units function (Richardson et al. 1974, Bonhomme et al. 2010), followed by the WE forcing units function (Wang and Engel 1998). Chilling units, defined by the SU function, were accumulated daily starting on 1 Aug of the previous year (t0 SU = 1 Aug year n-1). Grapevine buds exit endodormancy once the condition ∑ Daily Chilling units ≥ C*budbreak is met, where C represents the threshold for chilling units required to complete endodormancy. Elevated temperatures during late summer do not contribute to chilling units and may result in negative contributions during extreme heat (Table 2). Once the chilling requirement (C*budbreak) was satisfied, the WE function was applied to accumulate forcing units daily until budbreak, as defined by the condition ∑ Daily Forcing units ≥ F*budbreak. Here, F represents the threshold for forcing units required to initiate budbreak. The WE model accounts for the nonlinear relationship between grapevine growth and temperature, with temperatures outside the optimal range providing reduced or no contribution to ∑ Daily Forcing units (Table 2).
Grapevine phenology models used. Equations are applied daily using mean daily temperatures (Td) from a given starting date (t0) and continue until the corresponding chilling/forcing unit requirement (C*, F*) has been reached.
To project bloom dates, the accumulation of daily forcing units was tracked using the WE function starting from the projected budbreak date (t0 bloom) until the F*bloom was reached (∑ Daily Forcing units ≥ F*bloom). Similarly, the veraison date was projected to occur once the F*veraison from bloom onward (t0 veraison = projected bloom date) was reached (∑ Daily Forcing units ≥ F*veraison). The nonlinearities observed in grapevine growth under elevated temperatures (Ferrini et al. 1995, Zufferey et al. 2000, Zaka et al. 2017) are better represented in the projected bloom and veraison dates using the WE function as compared to standard GDD functions.
The harvest date model was calibrated independently, as it is estimated separately from the three prior phenological stages (budbreak, bloom, and veraison), which are modeled sequentially. In this process, each stage’s output serves as the starting point for the next, capturing phenological progression. The harvest date model, however, was based solely on sugar accumulation, was not connected to earlier phenological events, and was adapted from a grapevine sugar ripeness (GSR) model (Parker et al. 2020). The GSR model does not project a specific phenology stage, but rather uses daily GDD calculations to model the time to reach certain sugar concentrations (170, 180, 190, 200, 210, and 220 g/L). Since field observational data from the vineyard data set did not include time-course measurements of sugar accumulation prior to harvest, harvest was approximated at sugar concentrations of 230 g/L or 23 Brix (about average for TSS at harvest across all within-vineyard field observations). F*GSR values from 170 to 220 g/L developed for Pinot noir (Parker et al. 2020) were used to extrapolate the F*GSR value for Pinot noir at 230 g/L (see Supplemental Figure 1). The GSR function calculates forcing units starting at t0GSR using daily temperature until ∑ Daily Forcing units ≥ F*GSR. As such, harvest date modeling estimates when ~23 Brix has been reached. Unlike WE modeling of forcing units, the GSR modeling structure does not capture nonlinearities using optimal and suboptimal temperature ranges, but relies solely on a base temperature (Tbase = 0), below which no forcing units are accumulated.
Weather data and climate projections
Temperature data were sourced from the PRISM Climate Group (https://prism.oregonstate.edu/). Daily mean, maximum (max), and minimum (min) temperatures were obtained in 4 km × 4 km grids for the span of recorded field data (2011 to 2021). PRISM grids corresponding to each commercial vineyard location were used to project phenology and compare against observed field data for model calibration.
Daily mean, min, and max temperatures for climate change scenarios were obtained from NASA NEX-GDDPCMIP6 (Thrasher et al. 2022). The NEX-GDDP-CMIP6 data set was prepared by the Climate Analytics Group and NASA Ames Research Center using the NASA Earth Exchange and was distributed by the NASA Center for Climate Simulation. This data set is a reprocessing of CMIP6 climate change projections produced by the IPCC AR6 report (IPCC 2021). Daily data were available at 0.25°lon × 0.25°lat spatial resolutions from 2015 to 2100. Four shared socioeconomic pathways (SSP) were evaluated as defined by the IPCC: SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5. The 1-2.6 pathway projects drastically reduced global greenhouse gas (GHG) emissions in the near future, while the 5-8.5 pathway projects that GHG emissions continue to increase through the end of the century. Pathways 2-4.5 and 3-7.0 are considered intermediate, with drastically reduced GHG emissions by the middle or late 21st century, respectively.
Model calibration
The budbreak, bloom, and veraison models were calibrated using a parameter grid-search by testing different combinations of parameters and evaluating best model performance (Bergstra and Bengio 2012). Grid-searches were conducted on the variety-specific parameters within each model (Table 2), which are associated with species-level traits and do not vary from one variety or region to another (Parent and Tardieu 2012, Chuine et al. 2013) (Supplemental Table 1). A parameter grid-search was conducted for the optimal chilling unit requirement (C*budbreak) within the SU modeling of budbreak date, as well as the optimal forcing unit requirement (F*budbreak) for the WE component of budbreak modeling. Once optimal parameters for budbreak were established, grid-search was applied to the required WE forcing units for bloom modeling (F*bloom), followed by the WE forcing units for veraison (F*veraison). Since WE forcing unit accumulation for bloom begins only once budbreak has been reached (just as forcing unit accumulation for veraison begins only once bloom has been reached), model parametrization must be applied sequentially.
The harvest date model was calibrated independently, since the harvest date is estimated separately from the previous three phenological stages. Although optimal t0 GSR was determined to be 1 April (DOY 91) (Parker et al. 2020) for all target sugar concentrations and across all varieties following their sensitivity analysis, optimal t0 GSR values and overall model performance varied depending on the target sugar concentration. As a result, we chose to calibrate both t0 GSR and F*GSR to focus our modeling framework on the specifics of Willamette Valley Pinot noir. The optimal start date (t0 GSR) and forcing unit requirement (F*GSR) for ~23 Brix accumulation were established via grid-search. This adaptation to the GSR model is hereafter referred to as the modified GSR model.
Model validation and selection criteria
To avoid modeling shortcomings associated with overfitting (Janssen and Heuberger 1995, Sohil et al. 2022), CV procedures were applied and three model performance metrics were used to assess model performance (Table 3): root mean squared error (RMSE), Nash-Sutcliffe model efficiency (EF) (Nash and Sutcliffe 1970), and % accuracy. The % accuracy method is defined as the number of projected dates within +/− 3 days of the observed phenological date. K-fold CV was conducted for each iteration of the parameter search (e.g., unique combination of parameters), where each fold (K) is a unique train/test split. Each fold used eight of the 10 data years for the training data set and the remaining two data years as the test data set (~80/20 train/test split). Each unique permutation of which two data years comprised the test data set were explored, resulting in a total of 45 unique folds for each iteration of the parameter grid-search (e.g., test years 2012 and 2013, then years 2012 and 2014, etc.). Model performance metrics were calculated for each train/test fold of all parameter combinations for both train and test projections, as well as the variance of performance metrics across all train/test folds. The train and test folds were developed so that data from the test set did not leak into the training set (data-leakage), thus preventing overfitting in the CV step. By using the averages and standard deviations of the test data set model performance metrics, parameters were selected that resulted in the strongest and most consistent model performance, while limiting overfitting the models to the test data sets.
Modeling performance metrics used to evaluate Pinot noir phenology models.
Figures and statistical analysis were produced using R statistical software (R Core Team 2020). Modeling, calibration, and CV were conducted using the Python programming language (Python Software Foundation, python.org).
Results
Model calibration and performance evaluation
After calibrating key parameters in the SU, WE, and GSR functions by grid search, the best-performing set of parameter values were trained according to the mean and standard deviation of performance metrics across all 45 fold test data sets (Table 4). Using default parameters for Pinot noir modeling (Morales-Castilla et al. 2020, Parker et al. 2020) (Table 5), budbreak, veraison, and harvest dates each had negative EF values, indicating that projections on the test data set performed worse than predicting the mean of the test data set (Table 4). RMSE values ranged from 5.24 to 8.43 days using default parameters and % accuracy ranged from 29.1% to 52.3%. However, with calibrated parameters, RMSE, EF, and % accuracy values improved for all phenology stages, with the exception of % accuracy for bloom. Even after parameter calibration, budbreak EF remained negative, indicating that the mean of budbreak dates in the test data set served as a better predictor than the calibrated budbreak model. Improvements to the test data set performance metrics for bloom, veraison, and harvest indicate that the retained calibrated parameters were better suited to consistently predict phenology dates when applied to Willamette Valley Pinot noir.
Mean of Pinot noir phenology model performance metrics across all 45 fold validation data sets before and after parameter calibration. RMSE, root mean squared error; EF, model efficiency.
Pinot noir phenology model parameters before and after calibration procedures. Default values correspond to parameter values used in Morales-Castilla et al. (2020) and Parker et al. (2020) for Pinot noir phenology modeling. Calibrated values correspond to model parameter values retained following our calibration procedures. GSR, grapevine sugar ripeness model.
Qualitative model improvements varied for each phenology stage after parameter calibration. For budbreak modeling, calibration decreased C*budbreak and increased F*budbreak, indicating that Willamette Valley Pinot noir exits endodormancy with fewer chilling units and requires more forcing units before arriving at the budbreak date, compared to default C*budbreak and F*budbreak for Pinot noir. After adjustments to C*budbreak and F*budbreak (Supplemental Figure 2), the linear regression slope of the calibrated model more closely resembled the optimum 1:1 slope, indicating a better fit of the calibrated model (Figure 2A). Next, in the bloom calibration, the F*bloom decreased from 21.9 to 20.3 forcing units (Supplemental Figure 3), meaning that fewer forcing units were required to reach the projected bloom date for Willamette Valley Pinot noir (Figure 2B). Improvements to budbreak predictions also affect bloom predictions, since the improved DOY budbreak date becomes the new t0 bloom. As such, F*bloom was calibrated according to the new t0 bloom. Similarly, improved bloom projections affect veraison projections through the new t0 veraison. Next, the veraison model had a required increase of F*veraison from 59.5 to 63.1 (Supplemental Figure 4), indicating that additional forcing units are required to reach the projected veraison date from the projected bloom date (Figure 2C). Finally, harvest date results were taken independently from other modeling improvements, since the GSR model does not depend on other phenology stages and does not use a measure of grapevine growth to designate this stage (i.e., it uses sugar levels of juice/must). The decrease in t0GSR and increase in F*GSR indicate the model performed better with forcing unit accumulation beginning earlier in the year and requiring a larger sum of forcing units than the default GSR model (Supplemental Figure 5). This is likely explained by the Willamette Valley’s cooler climate and shorter growing season influencing Pinot noir ripening, compared to training data used in GSR model development (Figure 2D).
Comparison of model outputs between default parameter values and calibrated parameter values for budbreak (A), bloom (B), veraison (C), and harvest dates (D). The diagonal line represents a perfect 1:1 fit between observed and modeled phenology dates. DOY, Julian day of year.
Climate change projections
Using projected climate data, calibrated models were applied to explore how Pinot noir phenology may evolve in the Willamette Valley between now and the end of the 21st century. When comparing near and distant future (2050s and 2090s mean phenology, respectively) to the reference period (2020s mean phenology), Pinot noir phenology in the Willamette Valley is projected to advance earlier in the growing season for all SSP scenarios (Table 6). Projected phenology dates during the reference period do not vary significantly across different scenarios (Figure 3). These results are expected, considering that the various SSPs imply slow and long-lasting trends of changes to the global climate system, which take decades to observe the differentiation between socioeconomic pathways. What little distinction exists between these four scenarios during the reference period is likely explained by the inherent uncertainty associated with each climate projection. Likewise, the retrospective projections of phenology dates during the historical period (1990 to 1999) using PRISM weather data applied to our calibrated models show the same pattern (Figure 3). It should be noted that phenology projections during the reference period indicate phenology stages have already begun to advance compared to the historical period. Hastening of phenology dates is projected to continue from the reference period to the near future. During the 2050s, mean budbreak is projected to range from 1 to 2 April, bloom from 1 to 5 June, veraison from 6 to 10 Aug, and harvest from 6 to 20 Sept. Despite all four stages projected to earlier dates, the variation across scenarios remains small, indicating that even by mid-century, the discrepancy between socioeconomic pathways has not yet manifested itself in projected phenology dates. This is to be expected, given that global air temperatures only begin to diverge significantly by scenario after the mid-century.
Aggregate mean day of year (DOY) or change in DOY of Willamette Valley Pinot noir phenology projections according to four different shared socioeconomic pathways (SSP). Near and distant future delta values correspond to the difference between near and distant future phenology dates compared to the reference period. Negative values indicate advancing phenology.
Aggregate of Pinot noir phenology projections across the Willamette Valley American Viticultural Area from 1982 to 2100 using calibrated models. Shared socioeconomic pathways 1-2.6 (A), 2-4.5 (B), 3-7.0 (C), and 5-8.5 (D) are presented with all four major phenology stages: budbreak, bloom, veraison and harvest. The dashed line corresponds to a 5-year rolling mean, while solid lines correspond to linear regressions from 1982 through 2100. Dates above lines within the figure correspond to the mean of phenology projections within the historical (1990 to 1999), reference (2020 to 2029), near future (2050 to 2059), and distant future (2090 to 2099) decades (source 1982 to 2019 weather data from PRISM, https://prism.oregonstate.edu/). DOY, Julian day of year.
Finally, we can observe projected phenology dates in the distant future period. During the 2090s, mean budbreak is projected to range from 26 March to 6 April, bloom from 19 May to 3 June, veraison from 24 July to 8 Aug, and harvest from 30 Aug to 19 Sept. By the end of the century, the influence of different socioeconomic pathways on phenology projections is visible, with mean harvest projections spanning as much as 20 days. The SSPs 3-70 and 5-85 indicate significant departures from the reference and historical periods for most phenology stages by the 2090s, while SSPs 1-26 and 2-45 are not affected as severely. These findings are in line with global temperature projections according to each socioeconomic pathway. However, regardless of SSP or period, results show that all phenology stages hasten over the coming century (Supplemental Figures 6 to 17).
Additionally, the largest range of values and strongest differentiation between SSP scenarios are shown for harvest dates. This may be due to the strictly linear framework of harvest date modeling, as opposed to bloom and veraison modeling which incorporates nonlinear responses based on daily temperature. It is also important to remember that the harvest date is not a physiological stage, but rather an approximation of when 23 Brix is reached. As such, advancement of the harvest date is likely more sensitive to net temperature increases. Conversely, the smaller range of delta DOY values for budbreak modeling is likely caused by the conflicting influence of increasing temperatures on chilling and forcing unit accumulation. Although forcing unit accumulation during ecodormancy may be more favorable due to warmer springs, meeting chilling unit requirements to lift endodormancy may be delayed due to warming winters. This may explain why advancing budbreak is more limited than all other modeled phenology stages.
Discussion
A framework was developed to calibrate grapevine phenology models to Willamette Valley Pinot noir. This was designed around a sequential modeling approach for budbreak, bloom, and veraison, which others (Parker et al. 2011, Fila et al. 2014, Leolini et al. 2020) argued may more accurately capture projected budbreak dates (and subsequent phenology dates) than models that accumulate forcing units from a fixed budbreak date (Amerine and Winkler 1944). In addition, the nonlinear nature of daily chilling/forcing unit accumulation is likely better suited to characterize effects of warmer winters or summer heat extremes on vegetative development in the context of climate change (Fila et al. 2014, Molitor et al. 2014). Contrary to previous works, the model calibration shown here relied on phenology data from a single region, as opposed to aggregating phenology data sets across multiple regions and pedoclimatic contexts. By calibrating to one region, the models captured Pinot noir phenology for the Willamette Valley, which is characterized by soils with high water-holding capacity and a cool climate with abundant annual rainfall yet dry summer months. Additionally, robust CV strategies were used to ensure that parameter calibration procedures did not overfit the data sets. These joint efforts were critical to ensure projected changes to phenology were accurate and robust across multiple climate change scenarios.
Results show that model calibration improved predictive accuracy and robustness according to model performance metrics. The magnitude of model performance improvement varied between each phenology stage. For example, initial veraison and harvest date models using default parameters had worse performance metrics than those for budbreak and bloom, and therefore greater potential for improvement. However, the modeling structure (WE for veraison versus modified GSR for harvest) and the irregular number of observations between phenology stages play an important role in improving model performance through calibration. RMSE values from previous modeling studies typically ranged between 4 and 10 (Fila et al. 2014, Parker et al. 2011, 2013, 2020), and EF values from 0.8 to below 0, indicating that the modeling framework falls within expected performance benchmarks. Some test data set EF and RMSE performance values were somewhat poor (negative EF for budbreak and RMSE > 7 for veraison and harvest), which were partially attributed to not removing outlier observations. The approximation of all harvest dates to 23 Brix and relative difficulty of capturing precisely 50% veraison at the field scale could also explain why harvest and veraison performed worse than budbreak and bloom. We also chose to report test data set metrics rather than training data set or full data set metrics. Regarding projected changes to phenology in four climate change scenarios, our findings show phenology may advance from 2 to 34 days earlier in the growing season by the end of the century, depending on the stage and SSP scenario. These findings are in line with previous works that projected incremental advancements (~0 to 40 days) for all phenology stages throughout the century (Fila et al. 2014, García de Cortázar-Atauri et al. 2017). More significant phenology advancements for SSP scenarios corresponding to larger end of century warming are also expected.
Understanding the impact of model calibration on predictive accuracy and reliability requires evaluating the extent that each parameter adjustment contributed to better characterizing phenology for the region. Adjustments to C*budbreak, F*budbreak, F*bloom, and F*veraison parameters were generally small (−3.2%, +20.8%, −7.3%, and +6.1%, respectively); however, each adjustment affected all subsequent projections. For budbreak predictions, we hypothesized that the decrease of C*budbreak and increase of F*budbreak speaks to regional specificities of Oregon’s Willamette Valley pedoclimatic characteristics and their effects on endodormancy release/ecodormancy phase. Although the test data set RMSE values for budbreak improved by nearly one day, EF remained negative after calibration, indicating further improvements were needed. In other cases, such as veraison and harvest, test data set EF values changed from negative to positive values, indicating that calibration had relevance to the specific growing season dynamics of the Willamette Valley, since these model projections now outperformed the mean of the test data set.
Harvest dates were initially calibrated to the GSR model’s F*GSR for Pinot noir at a target of 230 g/L sugar concentration exclusively and maintained t0GSR = DOY 91, as defined (Parker et al. 2020). However, adding t0GSR to the calibration noticeably improved performance. Since Parker et al. (2020) observed higher sensitivity of EF to target sugar concentrations than to t0, and our observational field data did not contain time-course sugar accumulation, pursuing calibration on both t0GSR and F*GSR was justified to pursue the greatest degree of regional specificity for Willamette Valley Pinot noir. Resulting application of the modified GSR model successfully led to improved model performance metrics as compared to the default GSR parameters (t0 = 91, F* = 2988) and the calibrated F*GSR with t0GSR fixed at DOY 91, indicating a more accurate and reliable characterization of Willamette Valley Pinot noir.
Robust CV procedures are essential to determine the applicability and reliability of phenology models. Our CV strategy of testing all possible permutations, of which two data years were used as test data sets, captured year-to-year effects. This approach prevented data leakage from train to test data sets, unlike other studies that relied on a single arbitrary test/train split, regardless of the years in which the data points originated. The resulting performance metrics are strong indicators of the modeling framework’s robustness in external settings. However, the irregular number of observations per year in our phenology data set led to varying numbers of observations in validation folds, depending on which two data years were selected as test years. Although this approach prevented data leakage during CV, it led to highly variable performance metrics, depending on the size and distribution of the data points contained in each validation fold.
Further efforts in phenology modeling will likely lead to improved model performance. As with any model calibration, high quality ground-truthing data is essential to solve for the most optimal set of parameters. This study, like many others, has irreducible error associated with field observations that cannot be fully remediated (intrafield variability and human error). For future work, increased number of phenology observations across diverse Willamette Valley vineyards would further improve model parameter calibration to the region.
Higher quality ground-truthing data will also allow better calibration using temperature data, especially within each region. This study focused primarily on calibration using average daily temperatures, but incorporating cardinal and optimal temperatures into the calibration process could further refine and improve model performance. For GSR modeling, specifically, we did not have access to time-course sugar accumulation data and could not attempt GSR modeling at multiple sugar levels or guarantee that 23 Brix was reached at each predicted harvest date. Further work on Pinot noir sugar accumulation is warranted to investigate late-season berry sugar accumulation under Willamette Valley pedoclimatic conditions and how it may deviate from default GSR modeling. Regarding the choice of the overarching modeling framework, it may also be worth exploring other temperature-based process models, including those that consider cardinal and optimal temperatures in calibration (Richardson [Richardson et al. 1974], UniChill [Chuine et al. 1999], UniForc [Chuine 2000], BRIN [García de Cortázar-Atauri et al. 2009], and Best SIG [Schultz 2016]), which could outperform previous modeling strategies.
To capture vineyard-specific environmental conditions for model calibration, we used the PRISM data set at a resolution of 4 km × 4 km, available from 1981, to calibrate the phenology model parameters across various vineyard locations. The primary challenge was acquiring spatially explicit data that could accurately represent individual vineyard environments. Given the limited temporal extent of PRISM data, for future projections, we incorporated the NEX-GDDP-CMIP6 data set, which has a 0.25°lon × 0.25°lat resolution and a longer data range. This combination allowed application of temperature changes specific to each grid, avoiding uniform linear increases and instead linking projected temperature changes to our model. Although this approach is limited by the absence of finer-resolution data over a longer period, it provides a basis to address spatial differences in climate projections. Future studies could enhance vineyard-specific climate projections by incorporating finer-scale data, potentially through high-resolution downscaling or satellite-based climate data (see Supplemental Figure 18 for a comparison of historical PRISM and NEX-GDDP-CMIP6 minimum and maximum temperatures).
For climate change projections toward the end of the century, advances to phenology may entail significant economic impacts for growers. Earlier budbreak dates shift the timeframe of the developing buds’ vulnerability to spring frost. Even as winters are projected to become warmer on average, spring frost risks may persist, since the window of bud vulnerability to frost will occur during a cooler part of the year. Advancing bloom dates could be detrimental to yield, as rain, wind, or suboptimal temperatures can negatively affect flowering and/or fruit set (Schultz 2016, Zhu et al. 2020). The Willamette Valley’s cool Mediterranean climate historically has little to no rain during the summer months. However, bloom dates are now projected to shift earlier in June and potentially to as early as late May, when cooler temperatures and rainfall are more likely to occur. Poor bloom conditions would also affect the following year’s yields, since floral initiation determines the number of inflorescences per shoot for the following season. Advances to timing of veraison pose concerns about the conditions under which berry ripening occurs. Ripening berries will likely be more frequently subject to the higher temperatures of August, rather than ripening during the typically cooler September temperatures. Severe heat events occurring during the ripening phase may exacerbate challenges with accelerated sugar accumulation due to berry dehydration, and may negatively affect phenolic compounds and color, since anthocyanin accumulation is reduced under extreme heat conditions (Sadras et al. 2013). Vineyards around the world have begun experiencing these dynamics, now referred to as the decoupling of sugar accumulation and phenolic maturity (Previtali et al. 2022). Advancing harvest dates imply that grapes will likely be picked at higher temperatures, which may have downstream effects for winemaking due to increased enzymatic/bacterial threats to musts and wines. Logistical challenges such as maintaining sufficient labor under high harvesting temperatures or labor availability conflicts when other crops in the region are being harvested (e.g., summer blueberries) may also arise as a result of earlier harvest dates. Night harvesting and machine harvesting under cooler conditions may be required more often. The arrival of new insect threats or the acceleration of pest generations (e.g., vine mealybug) throughout the growing season may pose novel threats to the region, although further work is needed to evaluate these risks.
The financial implications of these projections could be significant. It is important to remember that these model outputs rely on a key assumption that all variables other than weather remain unchanged as compared to our training data (2012 to 2021 phenology observations). As with similar modeling strategies, our modeling assumes stationarity, meaning that the relationship between weather and grapevine phenology will remain unchanged over time. However, as shifts in grapevine phenology are forecasted, growers may choose to offset these changes by adapting different vineyard design and management practices. Changes such as selecting different cultivars, rootstocks, planting densities, trellis/training systems, irrigation, nutrition, canopy management, and vineyard floor management may influence vineyard characteristics and performance. Growers should look to other winegrape growing regions with analogous varieties and pedoclimatic conditions for inspiration regarding adaptations to warmer climates and earlier phenology stages.
Although we chose to model budbreak, bloom, and veraison using process-oriented models, increased accessibility of data science tools such as machine learning present avenues to explore further modeling. Similarly, availability of high-resolution climate change projection data has increased significantly over recent decades and creates opportunities for high-resolution spatial analysis of climate change projections and their effects on phenology at a regional scale. Despite these avenues for improvements moving forward, we posit that the calibration and modeling framework outlined in this study is currently applicable to any region with available weather data and observational phenology data sets for ground-truthing for many varieties.
Conclusion
The model presented here shows that Willamette Valley Pinot noir budbreak, bloom, veraison, and harvest are projected to advance earlier in the growing season by the end of the century under all SSP scenarios. This indicates that elevated air temperatures will affect grapevine phenology differently than current phenology dates. Advancing phenology dates will be incremental throughout the century and few distinctions are observed between scenarios until mid-century. Impacts to phenology become differentiated by SSP scenario after mid-century, and phenology advances are projected to be more significant for SSP scenarios with higher temperatures. Despite these projected changes to Pinot noir phenology in the Willamette Valley, growers possess a variety of tools and adaptation strategies to offset unwanted phenology advancements and preserve yields and quality, particularly as they plan future vineyard renovations and developments.
Models of grapevine phenology are an important tool for growers to better understand the nature of the relationships between air temperature and grapevine growth. These relationships can be difficult to describe explicitly, given the large intraspecific variability among grapevine cultivars, and the nonlinear effects of temperature on photosynthesis. In addition to these challenges, regional vineyard traits such as pedoclimatic conditions and management practices likely influence grapevine responses to air temperature. In this work, we provide a method for robust calibration of phenology models specific to vineyards at the regional/variety scale, to further improve the accuracy and reliability of phenology models in a given context. We believe the method outlined in this paper is readily applicable to other regions and cultivars, so long as adequate phenology and temperature data sets are available.
Accurate and robust phenology models are essential for growers to adequately characterize the potential effects of climate change on their vineyards. In our case, phenology models will help inform Willamette Valley growers of the potential magnitude of phenology advancement in Pinot noir resulting from climate change. Given the importance of phenology on final yield and quality parameters, as well as logistical challenges of day-to-day vineyard operations, these results provide critical information for growers to prepare adaptation and mitigation strategies to build durable and economically thriving vineyards for the future of Willamette Valley viticulture.
Supplemental Data
The following supplemental materials are available for this article in the Supplemental tab above:
Supplemental Table 1 Fixed model parameters for each phenology stage and model. Smooth-Utah: Tm1, the decrease in cold efficiency for bud endodormancy; Topt, the optimal mean daily temperature for chilling unit accumulation; Tn2, the temperature with half the cold efficiency of Topt; min, the negative impact of high temperatures (Richardson et al. 1974, Bonhomme et al. 2010). Wang-Engel: Tmin and Tmax, threshold values at which forcing units are no longer accumulated; Topt, the optimal temperatures for forcing unit accumulation at each respective phenology stage (Wang and Engel 1998).
Supplemental Figure 1 F*GSR for the default parameters of 230 g/L target sugar concentration of the grapevine sugar ripeness (GSR) model according to the linear regression model of GSR F* values from target concentrations 170 g/L through 220 g/L. F*GSR, the threshold for forcing units required to reach GSR.
Supplemental Figure 2 Budbreak modeling parameter exploration for C*budbreak and F*budbreak. Test data set mean model efficiency (EF; A), root mean squared error (RMSE; C), and mean % accuracy (E) calculated across all folds for each parameter combination and their corresponding standard deviations (B, D, F). Black and red points correspond to default and calibrated performance metrics, respectively. C*budbreak, optimal chilling unit requirement to complete endodormancy; F*, threshold forcing unit requirement to initiate budbreak.
Supplemental Figure 3 Bloom modeling parameter exploration for F*bloom. Test data set mean model efficiency (EF; A), root mean squared error (RMSE; C), and mean % accuracy (E) calculated across all folds for each parameter combination and their corresponding standard deviations (B, D, F). Black and red points correspond to default and calibrated performance metrics, respectively. F*bloom, threshold for forcing units required to reach bloom.
Supplemental Figure 4 Pinot noir veraison modeling parameter exploration for F*veraison. Test data set mean model efficiency (EF; A), root mean squared error (RMSE; C), and mean % accuracy (E) calculated across all folds for each parameter combination and their corresponding standard deviations (B, D, F). Black and red points correspond to default and calibrated performance metrics, respectively. F*veraison, threshold for forcing units required to reach veraison.
Supplemental Figure 5 Pinot noir harvest modeling parameter exploration for F*GSR and t0GSR. Test data set mean model efficiency (EF; A), root mean squared error (RMSE; C), and mean % accuracy (E) calculated across all folds for each parameter combination and their corresponding standard deviations (B, D, F). Black and red points correspond to default and calibrated performance metrics, respectively. F*GSR, threshold for forcing units required to reach grapevine sugar ripeness (GSR); t0 GSR, optimal GSR model start date.
Supplemental Figure 6 Mean of projected budbreak (A), bloom (B), veraison (C), and harvest (D) dates for Willamette Valley Pinot noir over the reference years 2020 to 2029. Projections correspond to SSP1-2.6 climate scenario. DOY, Julian day of year; SSP, shared socioeconomic pathways.
Supplemental Figure 7 Mean of projected budbreak (A), bloom (B), veraison (C), and harvest (D) dates for Willamette Valley Pinot noir over the reference years 2020 to 2029. Projections correspond to SSP2-4.5 climate scenario. DOY, Julian day of year; SSP, shared socioeconomic pathways.
Supplemental Figure 8 Mean of projected budbreak (A), bloom (B), veraison (C), and harvest (D) dates for Willamette Valley Pinot noir over the reference period 2020 to 2029. Projections correspond to SSP3-7.0 climate scenario. DOY, Julian day of year; SSP, shared socioeconomic pathways.
Supplemental Figure 9 Mean of projected budbreak (A), bloom (B), veraison (C), and harvest (D) dates for Willamette Valley Pinot noir over the reference years 2020 to 2029. Projections correspond to SSP5-8.5 climate scenario. DOY, Julian day of year; SSP, shared socioeconomic pathways.
Supplemental Figure 10 Delta values of projected budbreak (A), bloom (B), veraison (C), and harvest (D) dates for Willamette Valley Pinot noir between the reference period (2020 to 2029) and the near future (2050 to 2059). Projections correspond to SSP1-2.6 climate scenario. DOY, Julian day of year; SSP, shared socioeconomic pathways.
Supplemental Figure 11 Delta values of projected budbreak (A), bloom (B), veraison (C), and harvest (D) dates for Willamette Valley Pinot noir between the reference period (2020 to 2029) and the near future (2050 to 2059). Projections correspond to SSP2-4.5 climate scenario. DOY, Julian day of year; SSP, shared socioeconomic pathways.
Supplemental Figure 12 Delta values of projected budbreak (A), bloom (B), veraison (C), and harvest (D) dates for Willamette Valley Pinot noir between the reference period (2020 to 2029) and the near future (2050 to 2059). Projections correspond to SSP3-7.0 climate scenario. DOY, Julian day of year; SSP, shared socioeconomic pathways.
Supplemental Figure 13 Delta values of projected budbreak (A), bloom (B), veraison (C), and harvest (D) dates for Willamette Valley Pinot noir between the reference period (2020 to 2029) and the near future (2050 to 2059). Projections correspond to SSP5-8.5 climate scenario. DOY, Julian day of year; SSP, shared socioeconomic pathways.
Supplemental Figure 14 Delta values of projected budbreak (A), bloom (B), veraison (C), and harvest (D) dates for Willamette Valley Pinot noir between the reference period (2020 to 2029) and the distant future (2090 to 2099). Projections correspond to SSP1-2.6 climate scenario. DOY, Julian day of year; SSP, shared socioeconomic pathways.
Supplemental Figure 15 Delta values of projected budbreak (A), bloom (B), veraison (C), and harvest (D) dates for Willamette Valley Pinot noir between the reference period (2020 to 2029) and the distant future (2090 to 2099). Projections correspond to SSP2-4.5 climate scenario. DOY, Julian day of year; SSP, shared socioeconomic pathways.
Supplemental Figure 16 Delta values of projected budbreak (A), bloom (B), veraison (C), and harvest (D) dates for Willamette Valley Pinot noir between the reference period (2020 to 2029) and the distant future (2090 to 2099). Projections correspond to SSP3-7.0 climate scenario. DOY, Julian day of year; SSP, shared socioeconomic pathways.
Supplemental Figure 17 Delta values of projected budbreak (A), bloom (B), veraison (C), and harvest (D) dates for Willamette Valley Pinot noir between the reference period (2020 to 2029) and the distant future (2090 to 2099). Projections correspond to SSP5-8.5 climate scenario. DOY, Julian day of year; SSP, shared socioeconomic pathways.
Supplemental Figure 18 Comparison of minimum and maximum temperatures between NEX-GDDP and PRISM from 1981 to 2014. Each boxplot represents the temperature distribution for a given year, with whiskers extending to 1.5 times the interquartile range. The figure suggests minimal differences between the two data sets over time.
Footnotes
Funding was provided by the Northwest Center for Small Fruits Research. The authors thank the commercial vineyard and winery companies and their vineyard management staff who were involved in generating data through the OSU Statewide Crop Load Project that were used in this manuscript. We appreciate the efforts of the following Skinkis Lab staff in assisting in data coordination and acquisition over 10 years: Annie Chozinski, Cody Copp, Amelia Doyle, Michael Kennedy, and Kelli Whisenhunt.
Delelee LM, Woodill AJ and Skinkis PA. 2025. Climate change projections indicate shifts in phenology for Willamette Valley Pinot noir. Am J Enol Vitic 76:0760003. DOI: 10.5344/ajev.2024.24033
By downloading and/or receiving this article, you agree to the Disclaimer of Warranties and Liability. If you do not agree to the Disclaimers, do not download and/or accept this article.
The data underlying this study are available on request from the corresponding author.
- Received June 2024.
- Accepted November 2024.
- Published online February 2025
This is an open access article distributed under the CC BY 4.0 license.
References
Vol 76 Issue 1
Jump to section
Related Articles
Cited By...
More from this TOC section
Similar Articles
