Advancing grapevine phenology predictions: a comparison of regression-based statistical learning approach and conventional models in mediterranean environment

Abstract

Climate conditions can significantly affect plant dormancy and growing season, impacting crop development and its productivity. Temperature is considered the primary driver of grapevine phenology, with endodormancy and ecodormancy largely regulated by chilling and forcing accumulations, respectively. Nevertheless, additional climatic factors, such as precipitation, radiation and atmospheric circulation patterns, may also contribute and regulate variability of different phenological phases. As hotspot of climate change, Mediterranean regions are witnessing negative compounding effects by enhanced warming and changes in precipitation patterns on crop systems and linked agro-industry. Among these, viticulture is a particularly relevant economic crop in Mediterranean regions. This research aims to assess some conventional phenological modeling approaches to project several phenological stages of grapevine in Sardinia (Italy), comparing the results to regression-based stepwise Statistical Learning that can potentially account for effect of several climate variables. Combined models (i.e. Chilling Hours-Growing Degree Hours) improve prediction of budburst phenological timing (RMSEs: 7 days) in relation to more complex models (i.e., PhenoFlex). The Growing Degree Days model was more accurate than complex phenological models in predicting blooming and veraison (RMSEs of 6 and 4 days, respectively). The regression-based statistical learning approach achieved phenological prediction performances comparable to the best of the conventional models (RMSEs: 4–7 days). Regression-based statistical learning approach was able to infer contribution of several climate variables on grapevine phenology, at a larger seasonal scale. Implementing both modeling approaches showed similar spatial patterns, except for budburst phase which indicated earlier timings over mountainous regions through regression-based statistical learning contrary to conventional model, highlighting the influence of maximum and minimum temperatures from previous year.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s00484-026-03312-3.

Introduction

Grapevines (Vitis vinifera L.) have been cultivated across vast areas, from 4° to 51° north and from 6° to 45° south, adapting to a wide spectrum of climates (e.g., oceanic, warm oceanic, transition temperate, continental, cold continental, Mediterranean, subtropical, arid, and hyper-arid climates), showing a great flexibility to varied climate conditions (Van Leeuwen et al. 2024).

Globally, vineyards cover approximatively 7.1 million ha, with Spain accounting for 13%, and then respectively France (11%), China (11%), Italy (10%), and Turkey (6%), representing circa half of the world total vineyard area (International Organization of Vine and Wine (OIV), 2024).

Viticulture is basically of high importance to winemaking sector, and particularly in the Mediterranean regions (Santos et al. 2020) including Sardinia Island in Italy as well.

Changes in viticulture climate-suitability over the last decades, for viticulture in general or cultivation of specific varieties, have already been reported for many wine regions (Fraga et al. 2013; Santos et al. 2020). According to the OIV report (OIV 2024), the world vineyard area experienced an overall decreasing trend through the last two decades, particularly in 2003–2011.

Although grapevine is a highly adaptable species that can be cultivated under a wide range of climatic conditions, viticulture systems remain strongly dependent on local climate characteristics and the interactions among cultivar, environment, and management practices. Consequently, ongoing climate change may substantially alter the suitability of traditional wine-growing regions and affect the timing of grapevine phenological stages (Fraga et al. 2016a). In Mediterranean regions such as Sardinia, increasing temperatures, changes in precipitation regimes, and more frequent extreme weather events may disrupt the climatic conditions that historically supported grape production and wine quality. These changes can affect vine phenology, canopy microclimate, yield, fruit composition, and ultimately the typicity and quality of wines associated with specific terroirs (Fraga et al. 2012; Santos et al. 2020; Serra et al. 2025). Furthermore, extreme events such as heatwaves, spring frosts, droughts, and periods of excessive humidity may increase production risks and challenge the long-term sustainability of viticulture in vulnerable regions (Spellman 1999; Carbonneau 2003).

On the other hand, accurate predictions of vine phenological phases, which are significantly dependent on climate conditions, are crucial to produce high-quality grapes and wines (Malheiro et al. 2013). In fact, early warnings concerning the possible altering of vine phenological stages could support and guide more effectively field management practices, e.g., timing of pesticide application and harvesting (Tomasi et al. 2011).

Thus, tools to support better application of short and long-terms management practices, such as phenological models, emerge as important tools to address potential climate induced impacts on vine development and farming (Moriondo et al. 2015).

Given the strong influence of temperature on grapevine phenology, most conventional phenological models use chilling and forcing concepts to represent dormancy release and subsequent growth (Cesaraccio et al. 2004; Garcia de Cortazar-Atauri et al. 2009; Fila et al. 2014; Thomas et al. 2026). These models typically estimate winter chill accumulation followed by spring heat requirements, while simpler approaches such as Growing Degree Days (GDD) rely solely on temperature accumulation from a fixed starting date (Bonhomme 2000; Neumann and Matzarakis 2014). As a result, temperature is the primary predictor in most phenological models (Garcia de Cortazar-Atauri et al. 2009; Duchêne et al. 2010; Nendel 2010; Parker et al. 2011). Despite their widespread use, uncertainties remain in representing dormancy and chilling accumulation. The interaction between chilling fulfilment and heat requirements varies among cultivars and climates, and different chilling formulations can lead to substantially different phenological projections (Wang et al. 2020a). Consequently, dormancy processes remain a major source of uncertainty in phenological modelling. According to Chuine et al. (2016), the uncertainty in predicting budbreak or flowering of tree crops of numerous chill-heat accumulation process-based models could be avoided including into the parametrization process the endodormancy break data. However, the same authors stress the relatively complexity of such data acquisition (destructive and time-consuming measurements).

Surprisingly, some crop models which project phenological stages as well, e.g., STICS (Simulateur mulTIdisciplinaire pour les Cultures Standard) (Brisson et al. 2008), include relevance of precipitation only for viticultural water stress assessment (Fraga et al. 2013; Van Leeuwen et al. 2024). At least to our knowledge, conventional phenological models ignore the contribution of other meteorological variables (e.g., precipitation), in addition to temperature, to predict different growth stages. However, evidence indicates that different water conditions may change the crop phenology to some degree (Shellie et al. 2018) particularly in semi-arid and arid vineyard areas where irrigation is necessary. However, in western Europe, where the phenology models were originally tested and calibrated, irrigation is not used or is very limited in most vineyards. While some hypotheses indicate that water deficit in vine growing season may lengthen the dormancy stage, Shellie et al. (2018) found out that drought stress through delaying the onset of endodormancy, decrease the chilling accumulation and increase the readiness to resume growth in ecodormancy period, and in total can shorten the dormancy stage of grapevine. Rainfall deficit in autumn may not significantly affect leaf, flowers or fruit development.

Few studies only have showed air humidity`s effect on grapevine phenology. For example, Düring (1979), found a budburst advance of 7–14 days with relative humidity declining from 95% to 50%. On the other hand, during grape development reduced relative humidity rate, as a direct effect of increased temperatures, induces advanced maturity through higher bunch transpiration (Cabodevilla et al. 2024).

In addition to temperatures, as the principal phenology driver, other influential factors may include e.g., soil texture (Trought et al. 2008), soil temperature (Kliewer 1975), pruning dates (Dunn and Martin 2000; Gatti et al. 2016), photoperiod (Caffarra and Donnelly 2011), and rootstock (Menora et al. 2015).

Dai et al. (2019) using historical leaf unfolding observations of several tree species in Harbin, China, modeled three statistical learning algorithms compared to several phenological models. They found that statistical learning algorithms (particularly the gradient boosting decision tree model) performed better in prediction and can be largely applied in future phenological modeling along with temperature sensitivity to better understand plants’ ability to adapt for future climate change. Another study (Fasihi et al. 2025) specifically to predict grapevine phenological stages applied a semi-supervised approach integrated into eight regression-based statistical learning algorithms that automatically generate new interpretations for large number of unlabeled climatic data using a limited set of labeled phenological observations. The results of this work showed significant improvements in grapevine phenological prediction. Also, recently hybrid models, coupling regression-based statistical learning and biophysical techniques were successfully proposed for predicting blooming stage of different cherry species (Van Bree et al. 2025).

Owing to the importance of the conventional phenological models which are typically built on temperatures, in this study we aim to calibrate and test some process based phenological models (e.g., CAC (Chill Anti Chill days), GDH (Growing Degree Hours), GDD (Growing Degree Days), and PhenoFlex) to project three successive growth phases (e.g., Budburst, Full Blooming and Veraison) of grapevine (i.e., cultivar Cannonau) in Sardinia Island, Italy. Meanwhile, we built hybrid model combinations (e.g., C_GDD (Chill days_Growing Degree Days) and CH_GDH (Chill Hours_ Growing Degree Hours), to test further improvement in phenological model performances. Lack of process based phenological modeling which apply climate/environmental parameters other than temperatures, brings up statistical learning approaches such as Partial Least Square (PLS) regression and Stepwise models that can consider and parameterize statistical influence of different climatic and environmental factors (Fraga et al. 2016b; Deiana et al. 2023) to derive the most significant relationships with phenological timing variations. Hence, for the next step we use monthly meteorological timeseries to calibrate phenological modeling through PLS and Stepwise methods. We compare and validate the results of modeling performances between the conventional as well as the regression-based statistical learning models considering the obtained errors. Finally, the model performance to project the phenological phases is being tested in a spatial scale over Sardinia Island.

The objectives of this study were fourfold: (i) to calibrate and compare several conventional process-based phenological models for predicting grapevine Budburst, Full Blooming, and Veraison dates; (ii) to evaluate a regression-based statistical learning framework integrating multiple climatic variables beyond temperature; (iii) to assess whether hybrid sequential chill–heat combinations improve phenological predictions relative to conventional approaches; and (iv) to investigate the spatial applicability of the best-performing models across Sardinia Island.

The spatial implementation was designed as an illustrative regional application to evaluate whether calibrated models produce biologically coherent phenological patterns across heterogeneous climatic and topographic conditions. Therefore, spatial outputs should be interpreted as a potential indicator of model transferability and regional consistency rather than as a formal validation dataset.

Materials and methods

Data collection and preparation

The study focused on a typical Mediterranean environment, the Sardinia region, the climate is characterized by dry and hot summers and relatively mild winters. Rainfall are mainly concentrated during autumn and spring seasons and ranges between 411 and 1215 mm in the inner mountainous regions. Average annual temperature ranges from 11.6 °C to 18.0 °C. Due to its complex orographic structure, numerous bioclimates (43 isobioclimates) have been identified. Cannonau, the most widespread grapevine variety with around 7600 ha cultivated (Mercenaro et al. 2017), was used as the object of study. This grape variety is well-adapted to hot and dry conditions typical of this island. Also, “Cannonau di Sardegna”, designated as DOC (Controlled Designation of Origin) since 1972, is famous for its high-quality red wine (Petretto et al. 2021). According to ampelographic and genetic observations, the Cannonau grape has been linked to Grenache noir (France) and Garnacha tinta (Spain). The Grenache family accounts to circa 185 thousand ha worldwide (Mercenaro et al. 2017).

Observational data on grapevine phenology were gathered from 17 experimental sites in Sardinia Island, Italy (See supplementary material, SM1), available for the 1991–2020 period. Sites object of study covers several typical areas of grapevine cultivation of Sardinia (Mercenaro et al. 2017; Serra et al. 2025): northern island (Alghero, Sorso, Badesi, Luogosanto municipalities), the eastern hilly region of Ogliastra (Jerzu, Arzana, Tertenia and Lanusei municipalities) and plain areas of southern island (Oristano, Guamaggiore, Villasor). The selection of sites intended to cover as much climatic variability as possible. Both coastal (e.g. Badesi, Oristano) and inland (e.g. Guamaggiore, Luogosanto) cases were considered. Sites elevation above sea level ranged between 12 m a.s.l (Oristano) and 692 m a.s.l. (Jerzu, SM2). According to Canu et al. (2015) classification, the observed sites covered isobioclimates including: 6-Upper Thermomediterranean, Lower Dry, Euoceanic Weak; 9 - Upper Thermomediterranean, Upper Dry, Euoceanic Strong; 10 - Upper Thermomediterranean, Upper Dry, Euoceanic Weak; 17 - Lower Mesomediterranean, Upper Dry, Euoceanic Weak; 20 - Lower Mesomediterranean, Lower Subhumid, Euoceanic Weak.

The phenological data provided by the Agricultural Department at the University of Sassari, include three main phenological stages: Budburst (BBCH 09), Full Blooming (BBCH 65) and Beginning of Veraison (BBCH 81) stages (Lorenz et al. 1995). The total number of observations for each phenological stage varies by 47 for Budburst, 42 for Full Blooming and 29 for Veraison (SM3 and SM4). Averaged dates for budburst, Full blooming and Veraison stages are 95, 144, and 205 (DOY) respectively. The highest standard deviation is measured around 9 days for budburst phase, while the lowest one is 6.8 days for Veraison, confirming greater variation in budburst stages among the observations (SM4). The largest phenological period obtained between Full blooming to Veraison is around 61 days. The experimental sites are well distributed over the study area encompassing northern to southern regions.

Hourly temperatures, daily maximum, minimum, and mean temperatures, daily precipitations and specific humidity timeseries were retrieved for each site and year, in association to the available phenological observations. Because a limited number of weather stations could not cover all the experimental sites or provided incomplete long-term data, the climate timeseries were gathered from the Centre for Euro-Mediterranean Climate Change Data Delivery System (CMCC DDS) Reanalysis v5 (ERA5) downscaling at 2.2 km horizontal resolution over Italy (Raffa et al. 2021). Reanalysis products with such high resolution facilitates the parameterization and validation of climate model assessments linked to experimental on-site observations. Previous studies applied the ERA5 reanalysis temperature data series for grapevine phenological modeling and produced fairly accurate phenological projections (Oses et al. 2020; Didevarasl et al. 2023).

Temperature data series from the above-mentioned repository were extracted as NetCDF over a spatial window (8–10° E and 38.5–41.5° N) at a spatial resolution of 0.02 degree for the 1990–2020 period over Sardinia Island, with yearly vectors (from October of the previous year to September) for the coordinates of each experimental site. The Climate Data Operator (CDO; Schulzweida 2023) was used for pre-processing the climate data series and imported in R as vector lists together with phenological observations per phenological stages. Data frames for all climate vectors and phenological observations were built and used in various model fitting functions which use hourly, daily and monthly timeseries.

SM5 shows the annual distribution of daily mean temperatures calculated amongst all experimental sites and years with available phenological observations. The temperature distribution confirmed the lowest and highest median ranks in February (~ 5 °C) and July (~ 29 °C), respectively. During winter, the temperature percentiles (25th–75th and 5th–95th) were more stretched, which indicated greater temperature variations amongst the experimental sites in the cold months of year.

Phenological models

The Chill and Anti Chill days model (CAC) is a sequential model that calculate the accumulation of chill days until the chill requirements are met and complete the endo-dormancy stage (Cesaraccio et al. 2004). Afterwards, the accumulation of anti-chill days begins during the eco-dormancy stage to overcome quiescence. We ran a set of equations using the single triangle degree day computation method for different temperature conditions (Cesaraccio et al. 2004; Didevarasl et al. 2023) in order to calculate chill days (Cd) and anti-chill days (Ca) in the model. Additionally, the CAC model included both effective temperature threshold (Tc), and a chilling requirement (Cr).

Growing Degree Days model (McMaster and Wilhelm 1997) calculates heat accumulation for achieving subsequent phenological stages based on the three following parameters: base temperature (Tb), maximum base temperature (Tx.base), and heating requirement (Hr).

The GDD equation is: .

where Tx is the daily maximum temperature, Tn is the daily minimum temperature, and Tb is the base temperature. The daily minimum and maximum temperatures should be set to Tb if less than Tb, or to an upper temperature threshold (Tx.base) when greater than Tx.base, since plants cannot grow efficiently beyond these thresholds.

Growing Degree Hours model calculates the daily heat unit accumulation (GDH) from hourly temperature data, using a standard linear model or the linear model proposed by Anderson and Seeley (1992). The basic model works using a base temperature. However, Anderson and Seeley (1992) considered also optimum and critical temperatures in the model. In both model types, heat accumulation begins when temperatures are above a minimum (base temperature, Tb), and growth increases linearly with temperature. In the model proposed by Anderson and Seeley (1992), the heat unit accumulation no longer increases once the optimum temperature (Topt) is reached, meaning that GDH above it are considered constant. The critical temperature (Tcrit) is the temperature above which growth ceases (i.e. GDH = 0). The function allows the user to define/optimize Tb, Topt and Tcrit, and also uses as default the values set by Anderson et al. (1986) for fruit trees: Tb = 4 °C, Topt=25 °C and Tcrit=36 °C.

Chill days_Growing Degree Days (C_GDD): A hybrid sequential chill–heat combination adapted for this study that combines Chill Days (Cesaraccio et al. 2004) and Growing Degree Days (McMaster and Wilhelm 1997) to project Budburst timing. This combination as a sequential model calculates the accumulation of chill days until the chill requirements are met and then accumulates GDD during the eco-dormancy stage to overcome quiescence.

Chill Hours_ Growing Degree Hours (CH_GDH): A hybrid sequential chill–heat combination adapted for this study that combines Chill Hours (Weinberger 1950) and Growing Degree Hours (Anderson et al. 1986) to project Budburst timing. CH model (Weinberg 1950) sums chilling hours over winter, with one chill hour accumulated for hourly temperatures between 0 and 7.2 °C. This combination as a sequential model calculates the accumulation of chill hours until the chill requirements are met and then accumulates GDH until reaching heat requirement.

The PhenoFlex model (Luedeling et al. 2021) offers a flexible framework for modelling the dormancy breaking process. Based on the dynamic model structure and the Growing Degree Hours (GDH) model, it calculates chilling and heating accumulations using a sigmoid function. The main parameters of both models are fitted to the observation dates, and a generalized simulated annealing algorithm (GSA; Xiang et al. 2013) is applied to calibrate the model. Because the model contains a relatively large set of physiological and optimization parameters, detailed descriptions of all parameters are provided in the Supplementary Materials (SM2). In the main text, only the key chill and heat requirement parameters used in model interpretation are discussed.

The Partial Least Square (PLS) regression-based statistical learning model allows us to establish significant relationships between various climate parameters and phenological timing. Flexible capacity of PLS to model relationships between multiple predictor (independent) and response (dependent) variables, especially when dealing with high dimensionality and multicollinearity, and particularly when the number of predictor variables exceeds the number of observations or when these variables are highly correlated, encouraged us to apply PLS in this investigation. To assess the relationship between climate conditions (independent variables including Tmax, Tmin, Tmean, precipitation and air humidity at monthly scale) and phenological stages including Budburst, Full blooming and Veraison (dependent variables), we performed Partial Least Square (PLS) regression (Deiana et al. 2023). Predictor variables were internally centered and standardized using the scaleC = “standard” option implemented in the “ropls” package, ensuring that variables measured on different scales contributed equally to the extraction of latent components. For all analyses, the number of predictive PLS components was fixed at two (predI = 2), because the objective of the PLS analysis was variable screening rather than optimization of predictive performance. Climatic variables with Variable Importance in Projection (VIP) values greater than 1 were retained, and the PLS/VIP procedure was repeated iteratively until a stable subset of influential predictors was obtained. After identifying the most influential variables, a full multiple linear regression model was fitted using the retained predictors. Bidirectional stepwise variable selection was subsequently performed using the step() function in R based on the Akaike Information Criterion (AIC), resulting in the final regression equations for each phenological stage. The iterative PLS/VIP screening was completed prior to model validation. Then, the final subset of predictors identified through the iterative PLS procedure was retained, whereas the bidirectional stepwise regression was repeated within each LOOCV training fold after excluding one observation at a time. Therefore, the reported validation results should be interpreted with the understanding that the predictor-screening stage preceded cross-validation and may have resulted in slightly optimistic estimates of predictive performance compared with a fully nested cross-validation framework.

For Budburst, climatic variables were considered from June of the previous year through April of the current year, because grapevine bud differentiation and the early stages of bud development begin during the previous growing season. Climatic conditions during this period may influence bud formation, dormancy establishment, and subsequent Budburst timing (Velappan et al. 2021, 2022). For Full Blooming and Veraison, climatic variables were considered from October of the previous year through the end of June and July, respectively. Monthly climatic variables were calculated over complete calendar months rather than being truncated at the observed phenological dates. Consequently, for some observations, June and July predictors may include climatic conditions occurring after the observed Full Blooming and Veraison dates. Therefore, the regression-based statistical learning models should be interpreted primarily as retrospective statistical models for identifying climate–phenology relationships and comparing modeling approaches, rather than as operational forecasting tools. These temporal windows were selected to account for both dormancy-related climatic influences and environmental conditions associated with seasonal grapevine development.

Calibration and validation of models

To calibrate the models, we coded the error functions in R programing (R version 4.4.3 (2025-02-28) using the relevant packages (e.g., ropls, stats, GA, hyfo, chillR, raster, and other relevant packages) for most of the conventional phenological models, except for the phenoFlex and regression-based models, which have error function per se. The CAC, the C _GDD, and the CH_GDH models, which are set to project Budburst stage, start to accumulate chilling from October first and then after achieving chill fulfilment, heating units. On the other hand, the GDD and the GDH models use the Budburst dates to start heating accumulation to project Full blooming and Veraison phases. Veraison was modeled independently from climatic conditions extending from the previous dormant season rather than relative to the predicted or observed Full Blooming date. This strategy was adopted to ensure that each phenological stage was calibrated independently, thereby allowing direct comparison among phenological models without propagating prediction errors between sequential developmental stages. Nevertheless, modeling Veraison relative to Flowering, as recently proposed by De Resseguier et al. (2026), represents a promising alternative framework that merits further investigation. Meanwhile, to optimize the aforementioned essential parameters per the conventional phenological models, the developed error functions were coupled with the fitness function of the Genetic Algorithm (GA) package in R (Scrucca 2017). GA was selected for CAC, GDD, GDH, C_GDD, and CH_GDH models because these models involve a relatively limited number of parameters and lower-dimensional search spaces, allowing efficient exploration and optimization of parameter combinations. For each model, parameter ranges were defined according to biologically meaningful upper and lower thresholds, and optimization was conducted using the corresponding temperature time series and a maximum of 1000 iterations to identify the parameter set producing the lowest root mean square error (RMSE).

In contrast, the PhenoFlex model was calibrated using the Generalized Simulated Annealing (GSA) algorithm (Xiang et al. 2013), which is embedded within the original PhenoFlex calibration framework and is specifically designed for optimization of highly nonlinear and multidimensional parameter spaces. Because PhenoFlex contains a substantially larger number of interacting parameters than the other models, GSA provides a robust global-search strategy that reduces sensitivity to local minima and facilitates convergence. Nevertheless, different optimization algorithms may exhibit different convergence dynamics, and therefore some influence of optimization strategy on parameter estimation and model performance cannot be completely excluded.

For the phenology fitter in PhenoFlex, we used weather time series organized continuously from October of the previous year through September of the current phenological year (October of previous year – September) alongside the observations of three phenological stages. To prepare the data, daily maximum and minimum temperature were converted into hourly timeseries using the idealized daily temperature curve presented by Linvill (1990). The number of iterations was set to 1000, with five search steps in the algorithm, as recommended.

To run the regression-based statistical learning model, we arranged climate timeseries at monthly scale in correspondence to the three phenological stages. For budburst we arranged the time series from June of the previous year to April. The time series started in June considering beginning of bud development, where previous studies reported that environmental conditions during bud dormancy in previous summer may affect the subsequent budburst (Velappan et al. 2021, 2022). Full blooming accounted for relevant period from October of the previous year to June, and for Veraison from October of previous year to July. The PLS regression model can then specify the most relevant climate variables and months for different phenological phases and their timing. To improve model accuracy and avoid overfitting, initial PLS models included all independent variables for each phenological phases. Next variables were selected with Variable Influence on Projection (VIP) values greater than 1, to perform more advanced enhanced models (Deiana et al. 2023) according to the 4 following steps.

Stepwise regression is a method to fit regression models in which the choice of predictive variables is carried out by an automatic procedure (Efroymson 1960; Hocking 1976). In each step, the procedure adds or removes a variable from the set of explanatory variables based on predefined criteria, using forward, backward, or combined sequence of F-tests or t-tests. At the end, the stepwise regression yields in linear model equations for each phenological stage.

For validation of modeling results, and exploiting the whole number of observations, a leave-one-out cross-validation (LOOCV) was applied in R (R version 4.4.3 (2025-02-28) using the “ga” function in GA library and developed custom cod for error calculation, to test the performance of all approaches according to RMSE values (Didevarasl et al. 2023). Indeed, LOOCV repeatedly uses one observation as validation to calculate the RMSE of the model function calibrated with all the remaining observations. The mean and standard deviation across all RMSE values were used to verify the consistency rate of the performance of each modeling approach.

The results derived from the model calibration and cross-validation per phenological models have been tested using the RMSE, coefficient of correlation and standard deviation comparing observation and projection values for each phenological stage (plotting a Taylor diagram). As next step, the models with the lowest RMSEs (CH_GDH, GDD and Stepwise regression model), were implemented to present a spatial projection of the modeling performance for Sardinia. We coded a mapping function in R and using the optimized parameters and the hourly, daily (for conventional models) and monthly (for regression model) timeseries over 30 years (1991–2020) from the n ERA5 downscaled at 2.2 km over Sardinia, modeling results were derived spatially for each year.

Results

The results of phenological models’ calibration (Table 1) indicate RMSE values for each phenological phase ranging between 4 and 23 days. Most conventional models (e.g., GDH, GDD, C_GDD and CH_GDH) performed relatively well (< 8 days of RMSE) with predictions in all the three phenological stages. The GDD model achieved the lowest RMSE (4 days) for Veraison, while the C_GDD model had the highest RMSE (8 days) for Budburst phases. With RMSE of 11.5 days, the PhenoFlex model outperformed CAC model (RMSE of 15 days) for Budburst phase. However, for later phenological stages (Full blooming and Veraison phases) the PhenoFlex produced the highest errors (RMSE of 12.5 and 23 days respectively) (Table 1). On the other hand, stepwise regression modeling showed good performance with RMSE values of 7, 5.5 and 4.5 days for Budburst, Full blooming and Veraison phases respectively. The final regression models were statistically significant (P < 0.05), although not every individual predictor available in the final equations was necessarily significant after model selection(Table 2).

Table 1.

Phenological stages of grapevine (Cannonau cultivar) and the performance of conventional phenological models by calculating RMSE (Root Mean Square Error) values (in days)

	CAC	GDH	GDD	C_GDD	CH_GDH	PhenoFlex
Budburst	15			8	7	11.5
Full Blooming	-	8	6	-		12.5
Veraison	-	7	4	-		23

Table 2.

Statistics derived from the PLS-assisted stepwise regression models for phenological stages of grapevine (Cannonau cultivar), for significant (vipVn values > 1 ) climate variables for specific months (Budburst (5 variables and 11 months), Full blooming (5 variables and 9 months), and Veraison (5 variables and 10 months)). (Note: the subscript “-1“ means previous year, for example: Jun₋₁_TX means Maximum temperature of previous June)

	Predictor	Estimate	Std. Error	t	P-value	Model statistics
Budburst	Intercept Jun−1_TX Oct−1_TN Oct−1_PR Dec−1_PR	12.468353 1.520173 2.505555 − 0.023587 0.019851	24.414580 0.625317 1.092191 0.010954 0.009796	0.511 2.431 2.294–2.153 2.027	0.6122 0.0194 * 0.0269 * 0.0371 * 0.0491 *	RMSE (days) = 7 R2 = 0.4287 Adjusted R2 = 0.3743
Full Blooming	Intercept Apr_TX Dec−1_TN Apr_TN May_TN Mar_PR Jun_PR	302.06883–5.6726 -1.3284 4.3264 -4.975 -0.217 -0.15	30.57068 1.24825 0.96724 1.6998 1.12 0.06 0.07945	9.881–4.544 -1.373 2.545–4.439 -3.619 -1.9	1.16e-11 *** 6.30e-05 *** 0.17835 0.015 * 8.61e-05 *** 0.000924 *** 0.06557 .	RMSE (days) = 5.5 R2 = 0.5806 Adjusted R2 = 0.5087
Veraison	Intercept Apr_TX May_TN Jul_TN Apr_TM Jul_TM	349.783 -7.723 -4.739 -8.076 7.469 5.483	33.628 3.108 2.478 2.774 3.658 2.533	10.402–2.485 -1.912 -2.911 2.042 2.165	3.62e-10 *** 0.0206 * 0.06841. 0.00787 ** 0.05278. 0.04102 *	RMSE (days) = 4.5 R2 = 0.5525 Adjusted R2 = 0.4552

Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

However, the coefficient of determination (R²) and adjusted R² values obtained from the stepwise regression models indicate that the selected climatic predictors explain a moderate proportion (R²: 0.42–0.58 and adjusted R²: 0.37–0.5) of the variability in grapevine phenological timing. This suggests that, although the regression-based statistical learning framework achieved competitive RMSE values, a considerable fraction of phenological variability remains associated with additional factors not explicitly represented in the models, such as cultivar-specific physiological responses, local management practices, soil conditions, and other environmental interactions.

All main modeling parameters for each phenological model and phase were optimized and presented in Table 3. To assess performance in predicting Budburst dates through more conventional models, CAC, C_GDD, CH_GDH, and PhenoFlex models were applied with different upper and lower thresholds for each modeling parameters (SM6). The optimization process defined for CAC model, optimal temperature threshold (Tc) of 11.36 °C, and chilling requirement (Cr) of -104.6 chill units; for C_GDD model, heat requirement (Hr) of 433.67 GDD°, maximum base temperature (BaseTmax) of 33.9 °C, base temperature (BaseT) of 5.58 °C, and chill requirement (Cr) of -107 chill units; for CH_GDH model, temperature threshold (Tc) of 11.26 °C, base temperature (Tbase) of 6 °C, chill hour requirement (CHr) of 95 chill units, base temperature for heat accumulation (Tb) of 7.64 °C, optimal temperature (Tu) of 26.45 °C, critical temperature threshold (Tcr) of 34.78 °C, heat requirement (Hr) of 3610 GDH°; for PhenoFlex model, chill requirement (Yc) of 25.3 chill units, and heat requirement (Zc) of 114.18 heat units. Indeed, for the PhenoFlex model, all twelve model parameters were calibrated during optimization. However, only the chilling requirement (Yc) and heating requirement (Zc) are reported in Table 3 because they represent the most physiologically interpretable parameters and facilitate comparison with the other phenological models. The complete set of calibrated PhenoFlex parameters is provided in the Supplementary Materials (SM2).

Table 3.

The optimized parameters for specific phenological stages of each model (Note: for phenoFlex model we only showed chill and heat requirement among 12 parameters)

	CAC	GDH	GDD	C_GDD	CH_GDH	PhenoFlex
Budburst	Tc = 11.36; Cr = 104.6			Tc = 11.26; Cr= -107.01; BaseT = 5.59; BaseTmax = 33.9; Hr = 433.6767	Tbase = 6; CHr = 95; Tb = 7.65, Tu = 26.45; Tcr = 34.78; Hr = 3610	Yc = 25.3; Zc = 114.18
Full blooming		Tb = 9.99; Tu = 28.65; Tcr = 34.43; Hr = 9392.07	BaseT = 9.61; BaseTmax = 34.48; Hr = 376.57			Yc = 22.6; Zc = 208.4
Veraison		Tb = 9.99; Tu = 28.65; Tcr = 34.43; Hr = 27,435	BaseT = 9.61; BaseTmax = 34.48; Hr = 1309.76			Yc = 23.4; Zc = 265

For both Full blooming and Veraison phases, we implemented and validated performances of GDH, GDD and PhenoFlex models. The optimal parameters for GDH model defined base temperature (Tb) of 9.99 °C, optimal temperature (Tu) of 28.65 °C, critical temperature threshold (Tcr) of 34.43 °C, and heat requirements (Hr) of 9392 (BBCH 65) and 27,435 GDH° (BBCH 81). The optimal parameters for GDD models indicated Base temperature (BaseT) of 9.61 °C, maximum base temperature (BaseTmax) of 34.48 °C, and heat requirements (Hr) of 376.57 (BBCH 65) and 1309.75 GDD° (BBCH 81). For PhenoFlex model optimal parameters for Full blooming and Veraison phases indicated chill requirements (Yc) of 22.6 and 23.4 chill units, and heat requirements (Zc) of 208.4 and 265 heat units respectively. Thus, we obtained various thresholds of each parameter as the optimized values per each specific phenological model and phase.

The optimization process through the PLS-assisted stepwise regression defined linear modeling for each phenological phase considering the significance level of variables and months influential in the variation of the phenological timings (Table 4). The estimated coefficients corresponding to each variable/month (as the independent variables) indicate -/+ relationship with the dependent variables (phenological phases). For the Budburst stage maximum temperature of previous June (Jun-1_TX), minimum temperature of previous October (Oct-1_TN) and precipitation of previous December (Dec-1_PR) showed positive coefficients (1.52, 2.5, 0.019 respectively), whiles precipitation of previous October (Oct-1_PR) indicated negative coefficient (-0.02). These relationships explain that increasing/decreasing of Jun-1_TX, Oct-1_TN, and Dec-1_PR could correlate to postponement/advance of Budburst stage, on contrary to Oct-1_PR.

Table 4.

Phenological stages of grapevine (Cannonau cultivar), and the corresponding linear modeling equations suggested by Stepwise regression modeling to project the phenological stages. (Note: the subscript “-1“ means previous year, for example: Jun-1_TX means Maximum temperature of previous June)

Phenological stage	Linear modeling using PLS-assisted stepwise regression
Budburst	DOY = 12.468353 + 1.520173 (Jun− 1, TX) + 2.505555 (Oct− 1, TN) − 0.023587 (Oct− 1,PR ) + 0.019851 (Dec− 1, PR)
Full Blooming	DOY = 302.06883–5.67263 (Apr_TX) − 1.32845 (Dec− 1, TN) + 4.32647 (Apr_TN) − 4.97576 (May_TN) − 0.21747(Mar_PR) − 0.15103 (Jun_PR)
Veraison	DOY = 349.783–7.723 (Apr_TX) − 4.739 (May_TN) − 8.076 (Jul_TN) + 7.469 (Apr_TM) + 5.483 (Jul_TM)

In the Full blooming phase, maximum temperature of April (Apr_TX), minimum temperature of previous December (Dec-1_TN), minimum temperature of May (May_TN), precipitation of March (Mar_PR) and precipitation of June (Jun_PR) explained mostly (with negative coefficients of -5.67, -1.32, -4.97, -0.21, -0.15 respectively) advancement/postponement of Full blooming stage. But only minimum temperature of April (Apr_TN) with a positive coefficient (4.32) explained that increasing/decreasing of Apr_TN could postpone/advance the Full blooming stage.

For the Veraison phase maximum temperature of April (Apr_TX), minimum temperature of May (May_TN) and minimum temperature of July (Jul_TN) showed negative coefficients (-7.72, -4.7, -8.07 respectively) which indicate a negative relationship between the independent variables and Veraison timing (as increasing/decreasing of the mentioned variables could lead to advance/postponement of Veraison stage). Instead, mean temperatures of April and July (Apr_TM and Jul_TM) had positive relation (7.46 and 5.48 respectively) on postponing the Veraison stage.

It should be noted that the monthly climatic predictors used in the regression equations were calculated over complete calendar months. Consequently, predictors such as June precipitation (Full Blooming) and July temperatures (Veraison) may partially include climatic conditions occurring after the corresponding phenological event. Therefore, these equations should be interpreted primarily as retrospective statistical relationships rather than operational forecasting equations.

Model calibration per phenological model and phase was validated using the leave-one-out cross-validation (SM7). The validation confirmed that most models produced results similar to the calibration, with mean RMSE values for Budburst, Full blooming and Veraison ranging from 6.9 to 15.24, 5.5 to 12.3, and 4.48 to 7.6 respectively.

All the phenological models, including all process based and PLS-assisted stepwise regression models, were compared in terms of standard deviation, RMSE, and correlation coefficient in a Taylor diagram (Fig. 1) for modelling performances. The PLS-assisted stepwise and GDD models indicated better modeling performance for Full blooming and, particularly for Veraison stage. For Budburst stage CH_GDH, C_GDD and Stepwise models showed the highest correlation coefficients and standard deviation values very close to the observations.

Fig. 1.

Normalized Taylor diagram showing all phenological models’ performances in projecting the grapevine’s phenological dates (i.e., Budburst, Blooming and Veraison)

The best performing models were implemented over Sardinia to verify their potential plausibility on a spatial scale (Fig. 2). Most of the models simulated, as expected, the latest phenological dates over colder regions in mountains and earlier projections dates over lowlands with higher temperatures. In mountainous regions projected dates, are starting from 100, 160, 240 (in DOY) for Budburst, Full blooming and Veraison stages respectively. On the other hand, in lowlands the above mentioned phenological stages occur considerably earlier: 64, 110, 184 (in DOY) respectively. So, the spatial phenological patterns obviously showed the higher temperature role in faster fulfilment of heat requirements in lowlands than mountains. However, Stepwise model projected the earliest phenological dates over the mountainous regions (starting from 64 DOY) rather than lowlands (starting from 84 DOY) solely for the Budburst stage. Meanwhile, the highest differences between Stepwise and the conventional models (SM8) are at higher elevation, or at least outside areas where grapevine is cultivated and beyond the training areas used for training the modeling approaches.

Fig. 2.

Spatial implementation of the models with best performance/lowest RMSE (CH_GDH, GDD and Stepwise models) to project grapevine phenological phases/dates (in DOY), including Budburst (top row), Full blooming (middle row), and Veraison (bottom row) for grapevine (Cannonau cultivar) over Sardinia Island. Modeling were carried out using hourly, daily and monthly timeseries from the high resolution ERA5 over a 30-year historic period (1991–2020) and then a long-term mean calculated for each model and phenological phase

Discussion

This study assessed the ability of six process-based phenological models and a regression-based statistical learning approach to project three successive grapevine phenological stages across Mediterranean environments. Remarkably, several models delivered promising results despite the challenging dataset structure, characterized by short observation series and climatically contrasting sites. Study results PLS-assisted stepwise regression modeling can perform phenological projections with performances comparable to some well performed conventional phenological models in our experimental area. Meanwhile, combined models (e.g., CH_GDH) prove improvements in modeling performances to project specific phenological phases (e.g. Budburst). The successful calibration achieved under these conditions highlights the potential of such models to be transferred and applied across wider Mediterranean areas exhibiting comparable climatic regimes, although the calibrated parameterizations are specific to the cultivar considered in this study (Wolkovich et al. 2025; De Rességuier et al. 2026).

In projecting Full Blooming and Veraison phases, the GDD and GDH models outperformed the more complex PhenoFlex model. Although PhenoFlex has demonstrated good performance for modeling dormancy release and early spring phenological stages due to its detailed representation of chilling and forcing interactions (Luedeling et al. 2021), its physiological assumptions may be less suitable for later developmental stages such as Veraison. The reduced performance observed in this study may also be related to the higher complexity of the model, which requires the optimization of a relatively large number of interacting parameters. Under a limited number of observations, such parameter-rich models may exhibit increased uncertainty, convergence instability, and a greater risk of overfitting. Furthermore, later phenological stages such as Veraison are influenced not only by temperature-driven developmental processes but also by cultivar-specific physiological responses, vineyard management practices, water availability, and other environmental factors that are not explicitly represented within the PhenoFlex framework. These factors may explain why simpler temperature-based models, such as GDD and GDH, provided more robust predictions for Full Blooming and Veraison under the Mediterranean conditions of Sardinia. However, in terms of modeling functionality regarding specific phenological phases, Veraison and then Full blooming stages projections showed lower RMSE values than Budburst stage, though this situation is contrary to PhenoFlex which is generally suitable for a single phenological stage (Didevarasl et al. 2023).

Evaluating the different phenological modeling approaches, the stepwise modeling with the RMSE values ranging from 4 to 7 days for Veraison, Full blooming and Budburst respectively, indicated good performance and even slightly better of conventional models (i.e. GDD, to project Veraison and Full blooming, and CH_GDH, to project Budburst). However, the moderate values of Multiple R² and adjusted R² suggest that the selected climatic predictors explain only part of the observed variability in phenological timing. This indicates that grapevine phenology is also influenced by additional factors not explicitly represented in the models, including genotype-specific responses, vineyard management practices, soil characteristics, and other environmental interactions. Therefore, the performance of the regression-based approach should be interpreted by considering both its predictive accuracy (RMSE) and its explanatory capacity (R² values), rather than relying solely on prediction errors. A study developed by Fraga et al. (2016a) applying stepwise regression model to project grapevine phenological stages in Portuguese wine regions confirmed our results, by showing similar low modeling errors (mostly < 5 days) and outperforming classic growing degree-days phenological models. Some conventional models have been also combined to predict phenological stages, for example using the Budburst date projected by CH_GDH, as the initial point to accumulate growing degree days/GDD to fulfil heat requirement for Full blooming and then Veraison stages. This model coupling improved performance regarding the above-mentioned phases.

Unlike traditional chill and heat accumulation-based models, which primarily rely on temperature sum, regression-based modeling can consider various climate variables to project different grapevine phenological stages (Dalla Marta et al. 2010; Bonnefoy et al. 2013). Meanwhile, it has been verified that temperatures exert a strong influence over some critical periods/months on grapevine development, which could be overlooked by traditional phenological models that accumulate daily chilling units starting from harvesting time and then heat units (Malheiro et al. 2013). Hence, in order to capture relation to larger temporal scale, which possibly control phenological stages, some researchers applied monthly climate variables (e.g. temperatures) instead of heat accumulation to build significant relationship between phenology and climate variability ( Dalla Marta et al. 2010; Garcia-Mozo et al. 2010; Bock et al. 2011; Tomasi et al. 2011; Urhausen et al. 2011). It should be noted that the regression-based statistical models developed in this study used monthly climatic variables calculated over complete calendar months. Consequently, for Full Blooming and Veraison, some selected predictors (e.g., June precipitation and July temperatures) may partially include climatic information occurring after the observed phenological event. Therefore, these regression equations should primarily be interpreted as retrospective statistical models for identifying climate–phenology relationships rather than as operational forecasting tools. Future studies should investigate event-based climatic windows that incorporate only climatic information available before each phenological stage to improve the forecasting capability of the regression approach.

Our findings are confirmed by Malheiro et al. (2013), where flowering stage indicates the most significant relationship with temperature in late winter and early spring months. The most significant variables in regression models for grapevine blooming phase included maximum and minimum temperatures of April, and minimum temperature of May. Under this regression an increase in Apr_Tn implied a delay in blooming date, showing a discrepancy between diurnal and nocturnal temperature influence on grapevine development during the early blooming-flowering period. This might be related to the role of nocturnal low temperatures in determining the carbon balance between inflorescences and early leaves and photosynthetic activity of next day (Sawicki et al. 2015; Tombesi et al. 2018). Tombesi et al. (2018) observed that night temperatures from 10 °C to 15 °C downregulated the carbohydrate translocation from leaves causing a reduction of subsequent photosynthetic rate, whilst night temperatures around 25 °C were considered as optimal. This might partially explain discrepancies we observed between Apr_Tn and May_Tn. Indeed, temperatures below 15 °C are ordinary in April (SM5 and SM8). However, further studies on grapevine physiology will be extremely useful and relevant to verify this hypothesis. Other than temperatures, the linear modeling elaborated by stepwise model includes as well precipitation in March and June, minimum temperature of previous December as the most influential variables/predictors for blooming phase. According to our model, an increase in March precipitation induces an advance in blooming, suggesting a plant adaptation, shifting the phase to a most favorable period. This highlights the importance of water availability to achieve and complete successfully the flowering phase (Yang et al. 2021). However, other authors reported a lighter role of rainfall in blooming phenology, which was strictly negatively correlated to mean temperature (Martínez-Lüscher et al. 2016).

The regression model suggested maximum temperature of previous June, minimum temperature of previous October, precipitation of previous October, and precipitation of previous December as the most significant variables controlling grapevine Budburst stage. All the mentioned variables are inherent to climate conditions from previous year and confirm more complex/indirect relationship between Budburst timing in late winter-early spring and climate variables at very earlier periods. In fact, above mentioned climate predictors could have potential carry-over effects in grapevine’s Budburst timing from the previous seasons/year.

Martínez-Lüscher et al. (2016) adopting a PLS regression technique to predict Budburst of cv. ‘Riesling Italico’ grapevine in Mandicevac (Croatia), considered daily mean temperatures from June of the previous year as independent variables. Results, in line with our findings, showed that the June period, as well as October, had a relevant and positive role in the model’s performance. Authors suggested that chilling period for grapevine are consolidated between September of previous year and February. Our findings are aligned to results obtained previously (Cameron et al. 2022; Faralli et al. 2024), suggesting mean temperature increase during winter to spring could shorten budbreak to flowering length by higher advances particularly at onset of flowering stage of grapevine confirming a significant correlation between them. The carry-over effect of thermal conditions of the previous growing season we observed represents relevant practical information for producers. Indeed, these findings corroborate the hypothesis that agronomical practices aimed to mitigate summer thermal stress and regulate ripening rate, may affect the budbreak phenology of the subsequent year.

In Veraison phase significant climate predictors are available in the same year including minimum temperature of July and May, maximum temperature of April, mean temperature of July and April. For Veraison stage, temperatures merely could be the most influential variables in phenology timing of grapevine in Sardinia Island. This relationship was verified by Fraga et al. (2016b). They found that minimum, maximum and mean temperatures during March-July, and March-April were significant predictors for Veraison timing of grapevine in Portugal. Also, Cameron et al. (2021) confirmed that grapevine Veraison phase advancement significantly was linked to spring maximum temperature. Conversely, other authors highlighted the role of precipitations on veraison occurrence, mainly in terms of drought stress, and grapes quality at harvest (Wolkovich et al. 2025).

On the other hand, process-based phenological approaches resulted in various optimized thresholds for each modeling parameter and the phenological phase. For Budburst stage CAC model indicated optimized base temperature and chill requirement (Tc:11.36 and Cr: -104.6) comparable to C_GDD (Tc:11.26, Cr: -107) with differences about 0.1 °C and − 2.4 for Tc and Cr respectively, but far from CH_GDH, particularly for base temperature (Tbase: 6 and CHr: 95). Although, the latter model aligned to the classical threshold for defining chilling-effective days whose mean temperature need to be lower than 10 °C (Dokoozlian 1999). C_GDD and CH_GDH models in terms of optimized base temperature and chill requirement to calculate chill-day and -hour units showed different values (Tc:11.26 vs. Tbase:6 and Cr: -107 vs. CHr: 95). However, these two combined models both showed almost similar optimized base temperature and particularly max temperature to calculate growing degree days (BaseT: 5.589 vs. Tb:7.64 and BaseTmax: 33.9 vs. Tcr:34.78). Optimized optimal temperature through CH_GDH model was 26.4 °C in forcing summation process to project Budburst phase.

For Full blooming phase, the optimized parameters through GDH and GDD models indicated the base and maximum temperatures were very close to each other (Tb: 9.99 vs. BaseT: 9.61 and Tcr: 34.43 vs. BaseTmax: 34.48). But as expected, we obtained different optimized heat requirements (Hr: 9392 GDH vs. Hr:376.56 GDD) regarding hourly and daily units in heat accumulation process through each model. Meanwhile, GDH model suggested an optimal temperature around 28.6 °C for Full blooming phase. We applied the same optimized parameters of Full blooming for the Veraison stage except for heat requirements which are optimized considering new ranges obviously higher than those of the Full blooming (Hr: 27435 GDH vs. Hr: 1309 GDD), knowing that to meet Veraison phase crop would need more heat requirements than Full blooming phase starting from the Budburst date.

In accordance to our findings, Fila et al. (2014) showed a range of base temperatures from 2.6 to 8.4 °C, and optimal temperatures from 22.1 to 27.7 °C for Budburst and flowering stages of Chardonnay grapevine cultivar using different phenological models (i.e., chilling- forcing, Fbeta and FGDD models) in Veneto region, North-Eastern Italy, through an experimental analysis in growth-room. Our base and optimal temperatures are about 1–1.5 °C greater than the maximum values obtained by the mentioned study, though considering the mean values our results show 3–4 °C greater base and optimal temperatures. This finding could confirm the differences between local climates (Sardinia Island in southern Italy vs. North Eastern Italy) and grapevine cultivars (Cannonau vs. Chardonnay). Also, they obtained forcing/heat requirements near to our results derived from GDD model, e.g., for Budburst stage, 409.4 (FGDD7 model), Full blooming stage, 389 (FGDD7 model), and Veraison stage, 1239 (FGDD7 model).

Validation of the models` calibration results using the cross-validation method indicated very similar modeling performances among most of the conventional models and stepwise regression model for prediction of all three successive phenological phases. But only the phenoFlex model showed the highest RMSE differences between model calibration and validation with 18 days for Veraison stage and then 2 days for Budburst stage. Large difference for the Veraison phase could result from the process PhenoFlex accumulates chilling and heating over long-time spans (from October first to predict Veraison dates in summer) which might increase errors (Didevarasl et al. 2023). Meanwhile, we should keep in mind that PhenoFlex is a model that is basically qualified to project only a single phenological phase, e.g., flowering (Luedeling et al. 2021). On the other hand, other conventional models accumulated forcing units to project the Veraison stage beginning from estimated Budburst dates and using the new parametrization based on the growing degree days or hours models that avoid greater modeling errors through successive phenological phases. Also, stepwise modeling works using partial least square method and linear regression model by selecting always the most significant variables to estimate each single phenological phase and this way per se avoid producing higher modeling errors.

Although calibration RMSE provides a useful measure of model fit, lower RMSE values do not necessarily imply superior predictive performance or generalizability. Models with larger parameter spaces and greater structural flexibility may achieve lower calibration errors at the expense of increased overfitting risk. Consequently, model performance should be interpreted jointly considering calibration accuracy, cross-validation results, model complexity, parameter parsimony, and biological interpretability.

A limitation of the present validation strategy is the use of leave-one-out cross-validation (LOOCV), where observations originating from the same site may appear in both training and validation folds. Consequently, validation statistics may be somewhat optimistic relative to completely independent site-wise validation. Furthermore, the relatively limited number of observations constrained the implementation of more rigorous validation strategies. Future studies incorporating larger datasets should evaluate model transferability using independent spatial and temporal validation frameworks.

Overall, in terms of modeling performance per phenological phases our results (Fig. 1) confirmed that for Budburst stage stepwise regression model along with CH_GDH model resulted in better phenological simulation fitting. However, for Full blooming and Veraison stages stepwise and GDD models indicated the best modeling performance (RMSEs lower than those for Budburst). This finding was aligned to the results obtained by Wang et al. (2020b) that indicated better phenological modeling performances for Full blooming and veraison stages than Budburst (average RMSEs of 8.7 and 11.5 vs. 15.14 days) through multi phenological modeling of grapevine in China. Also, another investigation developed by Costa et al. (2019) obtained similar results in terms of grapevine phenological modeling confirming higher performances for flowering and veraison stages than Budburst (R2 > 0.7 vs. 0.4 < R2 < 0.7) applying GDD or Sigmoid models among the seven phenology models. Several previous investigations (Fila et al. 2014; Costa et al. 2019; Wang et al. 2020b) confirmed our results showing similar modeling performance in best (RMSEs ranging 4–9 days) through chilling-forcing models to predict different phenological phases of grapevine.

We implemented the best performed models spatially in Sardinia Island for a 30-year (1990–2020) period per three phenological stages. The spatial implementation of the selected models was interpreted as an illustrative regional application potential rather than an independent validation exercise. The goal was to examine whether models calibrated using phenological observations could reproduce biologically plausible spatial gradients across Sardinia under varying micro-climatic and topographic conditions. These spatial performances suggested the phenological modeling`s plausibility over a regional scale without model failures. All regression-based statistical learning and the selected chilling-forcing (i.e., CH_GDH) and forcing (i.e., GDD) models manifested spatially persistent phenological plausibility over the Island. Both the chilling-forcing and forcing models typically showed the earliest phenological projections in lowlands and the latest ones over mountainous regions confirming the role of temperature ranges in chill and heat accumulations to fulfill the optimized requirements, as that lower temperatures in the mountainous areas can prevent faster chilling and heating accumulations contrary to moderate temperatures in lowlands. The consistency of the model spatialization with phenological observations and thermal regime distribution in the island confirm the hypothesis of De Ressèguier et al. (2026), suggesting that models build on a local scale could be successfully extended outside the region of study as long as the temperature range is similar. Since the proposed models were built from a heterogeneous database, comprising different bioclimates representative of typical cv. Cannonau cultivation area, the model extension to the whole Sardinian territory could be considered reliable and useful as a basis of further studies aimed to precisely map phenological phases of the territory. Phenological maps of critical phases like flowering and veraison would be a fundamental tool, also in the context of climate change, that can be used to choose plant material best adapted to site characteristics or to choose the best site for a specific variety. For instance, Cannonau is characterized by relatively low anthocyanin content (Petretto et al. 2021), for this reason sensitive to high air temperatures during fruit maturation.

The stepwise model illustrated similar spatial patterns to those of the conventional models in phenological projections for Full blooming and Veraison stages. However, in spatial implementation our step wise model resulted in a controversial situation by illustrating earlier Budburst timing in mountainous areas, contrary to the traditional models which show typically earlier Budburst in lowlands and later ones over the elevations. Although the stepwise regression model demonstrated competitive predictive performance within the study area, its generalizability beyond the calibration domain remains limited. The predictor-selection procedure is inherently data-driven and therefore reflects the climatic characteristics represented in the training dataset. Likewise, the use of specific temporal windows for temperature accumulation may capture local phenological responses effectively but may not remain optimal under different environmental conditions. As a result, model relationships identified in this study should not be assumed to be universally applicable across regions, elevations, or climate regimes. This limitation becomes particularly relevant when the model is extrapolated to areas characterized by climatic combinations that were poorly represented during calibration. Moreover, including phenological observation cases with different or extreme thermal conditions could be useful to improve model performances, also in the perspective of using these models under future climatic conditions for the prediction of possible future land suitability shift (Serra et al. 2025; De Ressèguier et al. 2026).

Conclusion

Our research compared the regression-based statistical learning to conventional chilling-forcing phenological modeling approaches in simulating successive phenological phases of grapevine for Mediterranean environmental conditions, typical of Sardinia, Italy. Four chilling-forcing (i.e., CAC, C_GDD, CH_GDH, and PhenoFlex) and two forcing (GDH and GDD) phenological models were applied and evaluated, and the hybrid combinations (i.e. CH-GDH and C-GDD) improved modeling performance particularly for Budburst stage. Step wise modeling approach matched its own highest modeling performances to the best of conventional models. Meanwhile, in terms of the modeling functionality, better performances were observed for Veraison and Full blooming, than for Budburst phase. Amongst the conventional models, the most complex and data demanding model, PhenoFlex, was underperformed in projection of Veraison and Full blooming respectively. However, CAC model’s performance fell below all models in Budburst phase.

PLS-assisted stepwise regression modeling showed several advantages over the traditional models, with plausible performances. It relies on monthly time series data (does not need daily or hourly values despite the conventional models) with faster computation and data preprocessing. It accommodates and evaluates various climate or environmental variables to project several phenological sages, while the conventional models use solely temperatures. It can span backwards over period beyond previous harvesting dates to consider most far influential variables on phenology timing.

Our findings confirm that hybrid combinations of the phenological models could improve modeling performance for some specific phenological stages, e.g., Budburst. Meanwhile, some simple structured and flexible regression-based statistical learning phenological models, e.g., stepwise modeling could outperform more complex and data demanding models and process faster with least failures in spatio-temporal scale. Innovative approaches would support grapevine producers and wine sector to be well prepared to future climate through the resilient strategic management of vine cultivation, and appropriate cultural practices, considering dependable estimations and shift of different phenological timings. As future research vision, gathering more phenological observations for various grapevine cultivars and other vine-growing regions would support our approaches to produce more comprehensive validation and improve modeling performance.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

BBCH 09

Budburst

BBCH 65

Full Blooming

BBCH 81

Beginning of Veraison

DOY

Day Of Year

ERA5

ECMWF Reanalysis v5

CMCC DDS

the Centre for Euro-Mediterranean Climate Change Data Delivery System

CDO

Climate Data Operator

LOOCV

Leave-One-Out Cross-Validation

RMSE

Root Mean Square Error

Jun-1_TX

Maximum Temperature of Previous June

Oct-1_TN

Minimum Temperature of Previous October

Oct-1_PR

Precipitation of Previous October

Dec-1_PR

Precipitation of Previous December

Apr_TX

Maximum Temperature of April

Dec-1_TN

Minimum Temperature of Previous December

Apr_TN

Minimum Temperature of April

May_TN

Minimum Temperature of May

Mar_PR

Precipitation of March

Jun_PR

Precipitation of June

Aug_TN

Minimum Temperature of August

Apr_TM

Mean Temperature of April

Aug_TM

Mean Temperature of August

GA

Genetic Algorithm

GSA

Generalized Simulated Annealing Algorithm

PLS

Partial Least Square

VIP

Variable Influence on Projection

CAC

Chill Anti-Chil

GDH

Growing Degree Hours

GDD

Growing Degree Days

CH_GDH

Chill Hour_Growing Degree Hour

C_GDD

Chill_Growing Degree Days

DOY

Day of Year

Funding

Impacts on Agriculture, Forestry and Ecosystem Services (IAFES) Division at the Euro-Mediterranean Center on Climate Changes (CMCC) supported this research. This research was also funded by the Italian “Programma Operativo Nazionale” (UA2001PONRIC2021_35 (PON RI 2014/2020)), project NODES, funded by MUR—M4C2 1.5 of PNRR (grant agreement no. ECS00000036).

Data availability

Data will be made available on request.

Declarations

Competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.