Next Article in Journal
Long-Range Transport of Biomass Burning Aerosols from Southern Africa: A Case Study Using Layered Atlantic Smoke Interactions with Clouds Observations
Previous Article in Journal
Surface Ozone Trends and Health Impacts in the Yangtze River Delta Region During 2015–2019
Previous Article in Special Issue
Design of a Standards-Based Cloud Platform to Enhance the Practicality of Agrometeorological Countermeasures
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Evaluation of Phenology Models for Predicting Full Bloom Dates of ‘Niitaka’ Pear Using Orchard Image-Based Observations in South Korea

1
National Center for Agro-Meteorology, Seoul National University, Suwon 16614, Republic of Korea
2
National Institute of Horticultural & Herbal Science, Rural Development Administration (RDA), Wanju 55365, Republic of Korea
3
National Institute of Agricultural Sciences, Rural Development Administration (RDA), Wanju 55365, Republic of Korea
*
Authors to whom correspondence should be addressed.
Retired.
Atmosphere 2025, 16(9), 996; https://doi.org/10.3390/atmos16090996
Submission received: 26 May 2025 / Revised: 5 August 2025 / Accepted: 18 August 2025 / Published: 22 August 2025

Abstract

Abnormally warm winters in recent years have accelerated flowering in fruit trees, increasing their vulnerability to late frost damage. To address this challenge, this study aimed to evaluate and compare the performance of three phenology models—the development rate (DVR), modified DVR (mDVR), and Chill Days (CD) models—for predicting full bloom dates of ‘Niitaka’ pear, using image-derived phenological observations. The goal was to identify the most reliable and regionally transferable model for nationwide application in South Korea. A key strength of this study lies in the integration of real-time orchard imagery with automated weather station (AWS) data, enabling standardized and objective phenological monitoring across multiple regions. Using five years of temperature data from seven orchard sites, chill and heat unit accumulations were calculated and compared with observed full bloom dates obtained from orchard imagery and field records. Correlation analysis revealed a strong negative relationship between cumulative heat units and bloom timing, with correlation coefficients ranging from –0.88 (DVR) to –0.94 (mDVR). Among the models, the mDVR model demonstrated the highest stability in chill unit estimation (CV = 6.3%), the lowest root-mean-square error (RMSE = 2.9 days), and the highest model efficiency (EF = 0.74), indicating superior predictive performance across diverse climatic conditions. In contrast, the DVR model showed limited generalizability beyond its original calibration zone. These findings suggest that the mDVR model, when supported by image-based phenological data, provides a robust and scalable tool for forecasting full bloom dates of temperate fruit trees and enhancing grower preparedness against late frost risks under changing climate conditions.

1. Introduction

Plant phenology refers to the seasonal progression of developmental stages such as dormancy, bud break, flowering, and fruiting. For temperate deciduous fruit trees, which follow a predictable seasonal growth cycle, accurate phenological information is essential for orchard planning and climate risk management [1,2,3]. In particular, timely prediction of spring bud break and flowering is critical for preparing effective responses to spring frost damage [4,5,6].
Weather data, particularly temperature, are the most critical factor in predicting the developmental stages of fruit trees. Although various meteorological variables influence plant phenology, temperature exerts the greatest effect [7,8,9]. Seasonal temperature patterns—from late fall or winter, when flower bud dormancy begins, through spring—regulate key phenological responses such as dormancy release, bud break, and flowering [10,11,12]. For example, warmer fall and winter conditions are generally associated with delayed bud break and flowering, whereas warmer spring temperatures tend to advance these phenological stages [13,14]. Although these opposing effects may sometimes counterbalance each other without significantly altering bloom dates [15], insufficient chill accumulation during mild winters can lead to poor flowering in temperate fruit trees—often referred to as dormancy disorders. Conversely, excessive spring heat may accelerate flowering, increasing the likelihood of frost damage to flower buds [16,17,18]. Since both earlier and delayed flowering are strongly influenced by temperature, it is essential to quantify the effects of daily temperature variations on phenological development.
The agricultural sector is particularly sensitive to climate change. In recent years, abnormally warm winters have led to earlier flowering in fruit trees [19,20,21], increasing their vulnerability to late frost damage [20,22]. In response, the Agrometeorological Early Warning System (AEWS) [23] was established to predict weather risks based on crop phenology and to support timely decision making by farmers. For example, when spring frost is anticipated, the AEWS predicts bud break and flowering stages based on temperature thresholds and assesses the risk level daily. If the forecasted minimum temperature falls below a critical threshold, frost warnings and response guidelines are automatically sent to growers via mobile alerts to help protect crops. Notably, even under identical low-temperature conditions, the extent and pattern of frost damage can vary within the same orchard. This variation occurs because cold tolerance is strongly influenced by the phenological stage of each tree. Therefore, accurate estimation of both within-orchard temperature distribution and crop developmental stage is essential for reliable spring frost risk prediction [5,18].
Since the 1970s, phenology models have played a central role in simulating plant development in response to temperature using the concept of thermal time. These models quantify the accumulation of chill and heat units from autumn to spring, enabling the prediction of key phenological stages such as bud break and flowering [2,24]. Although these events may occur around the same calendar dates each year, their actual timing varies depending on interannual thermal conditions, making thermal time a more robust predictor than calendar-based estimates. A variety of models—including the Chill Hours, Utah, Dynamic, and Developmental Stage (DVS) models—have been widely applied across diverse crops and climatic zones [25,26,27,28,29,30,31]. The Utah model has shown strong performance in cooler regions, while the Dynamic model is better suited for characterizing chilling responses under warmer or subtropical climates [31]. These models continue to be refined and validated for species- and region-specific applications worldwide.
In South Korea, phenology models such as the development rate (DVR) and Chill Days (CD), calibrated under local climatic conditions, have shown prediction errors of approximately 1 to 4 days based on observed flowering dates [8,32,33]. However, most of these models were developed for specific regions, and only a limited number of studies have assessed their performance across multiple regions or evaluated multiple models within a single location [34]. While regionally distributed long-term phenology observations are essential for accurate parameter calibration, it is difficult to repeatedly observe biological events in standard trees over extended periods. Chuine et al. [34] emphasized that using phenological data collected from different locations without considering local adaptation can distort model fitting and reduce prediction accuracy. To ensure practical and nationwide use, phenology models should be optimized based on long-term regional data to reliably simulate flowering responses across diverse climatic zones [35].
Recent advances in image-based monitoring technologies have provided an effective alternative to manual phenological observations, enabling high-resolution, objective, and observer-independent monitoring of crop development. When combined with local weather data from automatic weather stations (AWS), real-time orchard imagery significantly improves the spatial and temporal quality of phenological datasets [36,37,38].
The main contribution of this study is the integration of image-based phenological observations with AWS-derived temperature data to evaluate and compare the performance of phenology models across multiple regions. This approach ensures standardized, scalable, and consistent data collection for robust model assessment. The objective of this study is to evaluate and compare the performance of three phenology models—the DVR, modified DVR (mDVR), and Chill Days (CD) models—in predicting the full bloom dates of ‘Niitaka’ pear using image-based observations, with the aim of identifying the most reliable and regionally transferable model for practical application across diverse climatic zones in South Korea.
To achieve these goals, temperature data from AWS installed at seven orchards were collected over five years. The models, which assume sequential fulfillment of chilling and heat requirements (i.e., heat units are effective only after the chilling requirement is met), were applied to simulate chill and heat accumulation leading to bloom. Model outputs were then compared with observed full bloom dates derived from real-time orchard imagery and on-site records to assess prediction accuracy.

2. Materials and Methods

2.1. Data Collection

The Fruit Tree Growth and Quality Management System (FTGQMS) [39] has been monitoring the growth status and quality of major fruit crops in South Korea—including apples, pears, peaches, grapes, and tangerines—and provides supplementary data such as pest incidence and meteorological observations. Specifically for pear trees, orchard imagery has been collected since 2012 in Icheon (Gyeonggi-do), Cheonan (Chungcheongnam-do), Yeongcheon (Gyeongsangbuk-do), Wanju (Jeollabuk-do), and Sacheon (Gyeongsangnam-do). Additional monitoring sites in Sangju (Gyeongsangbuk-do), Naju (Jeollanam-do), and Ulju (Ulsan) were added in December 2017. Each observation site is equipped with an automatic weather station (AWS) installed and maintained by the Rural Development Administration (RDA), Republic of Korea, which records temperature, humidity, solar radiation, rainfall, and wind at 10 min intervals. Most AWS units were installed after 2017 and are co-located with the phenological observation trees.
This study selected seven of the eight sites where both AWS data and orchard imagery were concurrently available (Figure 1). Daily maximum and minimum temperatures from 1 October 2017 to 31 May 2022 were used as model inputs. Missing or erroneous values were corrected using correlation-based interpolation with reference data from nearby Korea Meteorological Administration (KMA) stations. Full bloom dates were determined using orchard imagery based on the BBCH phenological scale [40], defined as the point at which 70–80% of flowers were open. In Sangju and Naju orchards, where mixed planting hindered image-based assessment, full bloom dates were instead obtained from field records.

2.2. “Niitaka” Pear Full Bloom Prediction Model

Dormancy, bud break, and flowering are critical stages in the phenological development of temperate fruit trees and are primarily influenced by temperature [2]. These stages are commonly expressed in terms of thermal time, which incorporates both temperature and time (e.g., growing degree days), allowing for more accurate prediction of key phenological events [30,41,42]. A range of phenology models has been developed to simulate these developmental stages based on accumulated chill and heat units.
In this study, we evaluated three representative phenology models used in South Korea for predicting the full bloom date of ‘Niitaka’ pear: the Development Rate (DVR) model, the Modified DVR (mDVR) model, and the Chill Days (CD) model. These models differ in their assumptions, input resolution, and threshold conditions, and were selected based on their relevance to Korean pear-growing environments. First, the DVR model, developed by the National Agricultural Technology Research Institute [43], estimates full bloom based on the cumulative development rate calculated from daily mean temperatures. It assumes that phenological development occurs only when daily mean temperature exceeds a base threshold of 5 °C. DVR values are accumulated from 1 January, and full bloom is predicted when the cumulative DVR reaches 100 (∑DVR = 100). For ‘Niitaka’ pear, the model uses the crop-specific constants A = 107.94 and B = 0.9.
The daily development rate (DVR) is calculated using the following formula, where T is the number of elapsed days:
  • DVR(T) = (1/(A × B^ T)) × 100,                                   for T > 5
Second, the mDVR model was adapted from a model originally proposed by the Japan Fruit Tree Research Institute [44], modified for application to ‘Niitaka’ pear in South Korea [33]. It estimates hourly temperatures using Sugiura’s method [45], which interpolates between the previous day’s maximum and the current day’s minimum temperature. The model simulates phenological development across two stages: dormancy and post-dormancy. Representative equations of the mDVR model used for DVR1 and DVR2 are based on Sugiura [45], but this study extends the original model by applying continuous, non-linear temperature response functions. During the dormancy phase, DVR1 represents the chill-related hourly development rate between 0 and 6 °C, with dormancy considered released when the cumulative DVR1 reaches 1.0 (∑DVR1 = 1.0). The low-temperature sensitive period, reflecting both endogenous and eco-dormancy effects, is considered complete when DVR1 reaches 2.0 (∑DVR1 = 2.0). Afterward, DVR2 reflects the heat-related development rate, and full bloom is predicted when the cumulative DVR2 reaches 0.9525 (∑DVR2 = 0.9525).
DVR1 Calculation:
  • Case 1: DVR1(T) = 0,                                           for T ≤ 0
  • Case 2: DVR1(T) = 1.333 × 10−3,                                   for 0 < T ≤ 6
  • Case 3: DVR1(T) = 2.276 × 10−3 − 1.571 × 10−4 × T,                         for 6 < T ≤ 9
  • Case 4: DVR1(T) = 3.448 × 10−3 − 2.874 × 10−4 × T,                        for 9 < T ≤ 12
  • Case 5: DVR1(T) = 0,                                        for T > 12
DVR2 Calculation:
  • Case 1: DVR2(T) = exp [35.27 − 12094/(T + 273)−1],                         for T ≤ 20
  • Case 2: DVR2(T) = exp[5.82 − 3474/(T + 273) −1],                          for T > 20
                                             (based on Sugiura [45])
Finally, the CD model, developed by Cesaraccio et al. [30], estimates phenological development using daily maximum and minimum temperatures, with weighted contributions based on a threshold temperature (Tc = 5.4 °C). Chill units (Cd) are accumulated from late fall until the chilling requirement (Cr = −86.4) is reached, indicating the release of dormancy. Thereafter, heat units (Ca) are accumulated until the heating requirement (Hr = 272) is fulfilled, at which point full bloom is predicted to occur. These parameters were calibrated by Kim [32] using field data [18].
The Cd and Ca values are computed as follows, based on the relationship between daily minimum temperature (Tn), maximum temperature (Tx), mean temperature (TM = (Tn + Tx)/2), and the threshold temperature (Tc):
Chill units (Cd) Calculation:
  • Case 1: Cd = 0,                                    for 0 ≤ Tc ≤ Tn ≤ Tx
  • Case 2: Cd = −[(TM − Tn) − (Tx − Tc)/2],                          for 0 ≤ Tn < Tc ≤ Tx
  • Case 3: Cd = −(TM − Tn),                                 for 0 ≤ Tn ≤ Tx ≤ Tc
  • Case 4: Cd = −[(Tx/(Tx − Tn)) × (Tx/2)],                           for Tn ≤ 0 ≤ Tc ≤ Tx
  • Case 5: Cd = −[(Tx/(Tx − Tn)) × (Tx/2) − (Tx − Tc)/2],                    for Tn ≤ 0 ≤ Tx ≤ Tc
Heat units (Ca) Calculation:
  • Case 1: Ca = TM − Tc,                                  for 0 ≤ Tc ≤ Tn ≤ Tx
  • Case 2: Ca = (Tx − Tc)/2,                                 for 0 ≤ Tn < Tc ≤ Tx
  • Case 3: Ca = 0,                                     for 0 ≤ Tn ≤ Tx ≤ Tc
  • Case 4: Ca = 0,                                     for Tn ≤ 0 ≤ Tc ≤ Tx
  • Case 5: Ca = (Tx − Tc)/2,                                 for Tn ≤ 0 ≤ Tx ≤ Tc
                                          (based on Cesaraccio et al. [30])

2.3. Analysis of Chill and Heat Units

Before evaluating model performance, we analyzed the interannual and regional variation in the accumulation of chill and heat units that influence the flowering physiology of ‘Niitaka’ pear. Specifically, we calculated and compared the chill units required to release endogenous dormancy and the heat units required for bud break and flowering across models, in order to assess (1) their relationship with full bloom timing and (2) the interannual stability of model predictions using the coefficient of variation (CV).
Since each model defines temperature thresholds and weighting functions differently, the estimated timing and magnitude of dormancy release vary accordingly. For chill accumulation, the period from 1 October to 15 February—commonly regarded as the dormancy phase [46,47]—was used for analysis. Chill units were calculated using the following three methods:
(1)
Hours below 7 °C, a widely used indicator of winter chilling in fruit trees [48,49];
(2)
DVR1 values from the mDVR model;
(3)
Chill values from the CD model.
Cumulative chill units were compared across regions over five dormant seasons (2017–2018 to 2021–2022).
A similar approach was applied to analyze heat accumulation. The period from 1 January to 15 April—encompassing the typical post-dormancy and flowering phase in most Korean regions—was selected. Heat units were calculated using three methods:
(1)
Hours above 5 °C, based on the DVR model;
(2)
DVR2 values from the mDVR model;
(3)
Heat values from the CD model.
Cumulative heat units were derived for each model and region for the years 2018 through 2022. Model variability was assessed by calculating the coefficient of variation (CV) across years and regions.
To facilitate cross-model and cross-region comparison, all chill and heat unit estimates were standardized using Z-score normalization. This approach allowed for the visualization and interpretation of model outputs on a common scale, highlighting deviations from the global mean and revealing consistent spatial or temporal patterns. Standardized Z-scores were subsequently used in comparative figures and correlation analyses to evaluate the climatic relevance of each model’s output.

2.4. “Niitaka” Pear Full Bloom Prediction and Model Evaluation

Predicted full bloom dates from the DVR, mDVR, and CD models were compared with observed dates from seven orchards between 2018 and 2022. These phenology models simulate sequential dormancy processes by first accumulating chill units during the dormancy phase and subsequently heat units after dormancy release. However, under abnormally warm winter conditions, chilling accumulation may be insufficient, potentially resulting in premature or inaccurate bloom predictions. To address this issue, an upper limit for endogenous dormancy was set to 15 February in the mDVR and CD models, which explicitly simulate dormancy progression [47].
The performance of each model was evaluated by considering mean bias error (MBE; Equation (1)), root-mean-square error (RMSE; Equation (2)), and Nash and Sutcliff efficiency coefficient (EF; [50,51]; Equation (3)). MBE indicates underestimation/overestimation compared to observed data. RMSE shows the prediction error of the phenology models. In general, values of MBE and RMSE closer to 0 indicate higher accuracy, while EF evaluates the predictive skill of the model relative to the observed mean; values closer to 1 indicate greater efficiency. A Taylor diagram was generated to visually compare the models in terms of standard deviation, correlation coefficient, and RMSE, thereby facilitating an integrated assessment of predictive performance [52].
M B E = 1 N i = 1 N ( P i O i ) ( < M B E < ,   I d e a l   v a l u e = 0 )
R M S E = 1 N i = 1 N P i O i 2   0 < R M S E < ,   I d e a l   v a l u e = 0
E F = 1 i = 1 N O i P i 2 i = 1 N O i O ¯ 2   ( < E F 1 ,   I d e a l   v a l u e = 1 )
where O i is the observed value, O ¯ is the mean of observed value, P i is the predicted value, and N is the total number of observations.

3. Results and Discussion

3.1. Comparison of Chill and Heat Units by Region

Prior to evaluating model performance, we first analyzed the regional and interannual variation in chill and heat unit accumulation. This preliminary analysis aimed to (1) investigate the relationship between interannual variation in chill and heat accumulation and observed full bloom dates, and (2) assess the year-to-year stability of model estimates using the coefficient of variation (CV) as an indicator of consistency.
To improve interpretability and model-to-model comparability, all chill units were standardized using Z-scores and are presented in Figure 2. Clear latitudinal trends were observed: Icheon consistently showed the highest Z-scores (e.g., Z = +1.95 in 2017–2018) for hours below 7 °C, while Sacheon exhibited extremely low values (Z < −2.5), indicating substantially reduced chilling in southern regions. Interestingly, mDVR and CD model Z-scores revealed distinct regional dynamics. For instance, the mDVR model identified Wanju as having the highest chill accumulation, while the CD model indicated Naju as relatively chill-rich and Icheon as consistently chill-deficient. This divergence is likely due to model-specific temperature weighting functions and chill effectiveness thresholds. Despite Icheon having the longest total cold exposure, its effective chill accumulation was not the highest, suggesting that duration alone does not guarantee physiological dormancy release. Among the three methods, the mDVR model showed the lowest interannual variation (CV = 6.3%), followed by the CD (CV = 12.1%) and 7 °C threshold method (CV = 13.3%) (Table 1), indicating superior stability in modeling effective chilling.
Cumulative heat unit estimates from 1 January to 15 April, standardized using Z-scores for model-wide comparison, revealed distinct interannual and regional trends (Figure 2). All three models—hours above 5 °C, mDVR, and CD—showed a general increase in heat accumulation from 2018 to 2021, followed by a notable decline in 2022. Sacheon recorded the highest relative values across all models and years (e.g., Z > +2.5 in 2019–2020), whereas Icheon consistently exhibited below-average heat accumulation (Z < −1.5). These patterns reflect a clear latitudinal gradient and align well with the earlier full bloom dates observed in southern regions (Figure 2). The relationship between spring heat accumulation and full bloom dates for ‘Niitaka’ pear was confirmed by strong negative Pearson correlation coefficients across all models: −0.882 for DVR, −0.940 for mDVR, and −0.932 for CD (p < 0.001). These results emphasize the physiological importance of heat units in regulating bloom timing. Heat unit accumulation showed greater interannual variability than chill accumulation. Among the three models, the CD model had the lowest coefficient of variation (CV = 16.4%), followed by mDVR and DVR, indicating more consistent performance under varying climatic conditions. These findings demonstrate that heat-unit-based indicators from the CD and mDVR models can serve as reliable predictors of bloom phenology across diverse climatic conditions.

3.2. Evaluation of the Full Bloom Prediction Models for “Niitaka” Pear

The predictive performance of the DVR, mDVR, and CD models for forecasting full bloom dates of ‘Niitaka’ pear was evaluated using observational data from seven regions over a five-year period (2018–2022). While the DVR model showed the highest correlation with observations (r = 0.89, r2 = 0.79), this did not necessarily indicate superior prediction performance. The mDVR and CD models also exhibited relatively strong correlations (r = 0.86, r2 = 0.74 and r = 0.88, r2 = 0.77, respectively), yet all models generally underestimated bloom timing, predicting earlier dates than observed. As shown in Figure 3, this underestimation was particularly pronounced in 2021, when most regions exhibited negative ME values, indicating that the models predicted bloom dates earlier than observations under the warm spring conditions. When aggregated across all regions, the mDVR model exhibited the lowest root-mean-square error (RMSE = 2.9 days), followed by the CD (RMSE = 4.5 days) and DVR (RMSE = 6.1 days) models. The corresponding Nash–Sutcliffe efficiency (EF) values were 0.66 for the mDVR model, 0.16 for the CD model, and −0.57 for the DVR model, confirming the superior accuracy and consistency of the mDVR approach across diverse climatic conditions (Table 2, Figure 4).
The DVR model, which estimates effective heat accumulation based on mean daily temperatures above 5 °C from 1 January, showed limited reliability, particularly under unusually warm conditions. In 2020 and 2021—years characterized by abnormally high spring temperatures—the model exhibited considerable overestimation of heat units, resulting in early predicted bloom dates and large prediction errors. Although the DVR model yielded the highest coefficient of determination (r2 = 0.79), its overall prediction performance was poor, as reflected by a root-mean-square error (RMSE) of 6.1 days and a negative Nash–Sutcliffe efficiency (EF = −0.57). Among the study sites, Sacheon showed the largest prediction error, with an RMSE of 9.3 days, underscoring the model’s limitations in southern coastal regions. In contrast, the mDVR model, which incorporates sequential chilling and heating phases, demonstrated both high accuracy and consistency. Its RMSE values ranged from 1.6 to 3.6 days across regions, and EF exceeded 0.5 in most cases, with the exception of Sangju and Yeongcheon, where EF values were 0.33 and 0.41, respectively. The CD model showed intermediate performance, with an overall RMSE of 4.5 days and an EF of 0.16. While Wanju and Sacheon exhibited relatively low RMSE values (2.6 days), the model struggled in Sangju, Yeongcheon, and Naju, where RMSE ranged from 4.9 to 6.8 days and the EF remained below 0.3. These findings highlight the superiority of the mDVR model in capturing both temporal and spatial variation in full bloom timing (Table 2, Figure 3). This conclusion is further supported by the Taylor diagrams (Figure 4), which integrate key performance statistics—correlation coefficient, normalized standard deviation, and RMSE—for each model and region. The mDVR model consistently clustered closer to the reference point across regions, indicating better agreement with observations compared to the DVR and CD models.

4. Conclusions and Future Study

In phenological modeling, obtaining objective and standardized observational data is as crucial as optimizing model parameters to enhance prediction accuracy. Traditional phenological records often suffer from inconsistencies due to variation in cultivar and tree age, site-specific environmental conditions, inadequate management practices, and observer bias, all of which compromise data reliability [8]. To address these limitations, this study employed orchard imagery and meteorological data from the Fruit Tree Growth and Quality Management System (FTGQMS). This method offers a high level of standardization, equivalent to having a single trained observer record full bloom dates across all study sites over five consecutive years, thereby ensuring consistency in the validation dataset. In this study, the latest bloom occurred on 11 April 2019, while the earliest was recorded on 4 April 2021, representing an interannual variation of about one week. Phenology models successfully reproduced these temporal patterns, and among the models evaluated, the modified development rate (mDVR) model showed the highest agreement with observed bloom dates, as indicated by superior EF values.
This performance advantage of the mDVR model became more evident when compared directly with the other two models. Both the mDVR and CD models, which incorporate sequential chilling and forcing phases, exhibited more stable performance than the DVR model across years and regions. This structural feature likely contributed to improved robustness under diverse thermal conditions. Among the three models, the mDVR model achieved the lowest root-mean-square error (RMSE = 2.9 days) and the highest Nash–Sutcliffe efficiency (EF = 0.66). The DVR model, by contrast, recorded the highest RMSE (6.1 days) and the lowest EF (−0.57), indicating poor overall predictive skill. The CD model demonstrated moderate accuracy (RMSE = 4.5 days; EF = 0.16), performing better than the DVR model but consistently less accurate than the mDVR model. The superior performance of the mDVR model may be attributed to its finer temporal resolution and regionally adapted threshold parameters, which allow it to better capture phenological responses under variable climatic conditions.
The mDVR model was identified as the most reliable phenological model for predicting regional full bloom dates of ‘Niitaka’ pear over the past five years. However, it did not produce consistently accurate predictions across all locations. For example, relatively low EF values were observed in Yeongcheon and Sangju, suggesting that the model may not fully capture region-specific phenological variability. These findings are consistent with previous studies by Hunter and Lechowicz [53] and Chuine et al. [54], which emphasized that no single model can universally represent the phenological responses of all species and environments. Therefore, multiple models should be validated and tested for each species and location. Nevertheless, generalized models such as the mDVR can still serve as useful tools in climate adaptation planning. Although these errors may be partially attributed to differential sensitivity to winter chilling or spring warming, it is also likely that non-thermal environmental factors—such as photoperiod, soil moisture, and physiological status—play a role in determining observed full bloom dates. Future research should aim to identify and quantify the influence of these additional drivers. Expanding the phenological observation dataset and incorporating additional environmental variables could enhance the robustness and accuracy of phenological models. Ultimately, the development of reliable and regionally adaptable models will support more precise forecasting of flowering times and help mitigate frost damage under future climate variability.

Author Contributions

Conceptualization, J.-H.K. and D.-J.K.; data curation, E.-J.Y.; formal analysis, D.G.K.; funding acquisition, J.-H.K. and D.-J.K.; investigation, J.-H.K., E.-J.Y., D.G.K. and J.-H.H.; methodology, J.-H.K. and D.-J.K.; project administration, J.-H.K., K.-M.S. and D.-J.K.; resources, J.-H.H.; software, D.G.K.; supervision, J.-H.K., K.-M.S. and D.-J.K.; validation, J.-H.K.; visualization, J.-H.K. and E.-J.Y.; writing—original draft, J.-H.K.; writing—review and editing, J.-H.K. and D.-J.K. All authors have read and agreed to the published version of the manuscript.

Funding

This work was carried out with the support of “Cooperative Research Program for Agriculture Science and Technology Development (Project No. RS-2024-00336321)” Rural Development Administration, Republic of Korea.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The data presented in this study are available from the corresponding author upon reasonable request. The data are not publicly available because they will be used in future studies.

Acknowledgments

The authors would like to thank Jin-il Yun for his insightful comments on the manuscript. We also thank the editor and reviewers for their constructive feedback. Finally, we gratefully acknowledge the Pear Research Institute of the National Institute of Horticultural and Herbal Science for providing the phenological data for ‘Niitaka’ pear.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Zavalloni, C.; Andresen, J.A.; Winkler, J.A.; Black, J.R.; Beedy, T.L.; Flore, J.A. The Pileus Project: Climatic Impacts on Sour Cherry Production in the Great Lakes Region in Past and Projected Future Time Frames. Acta Hortic. 2006, 707, 101–108. [Google Scholar] [CrossRef]
  2. Salazar, M.R.; Jones, J.W.; Chaves, B.; Cooman, A.; Fischer, G. Base Temperature and Simulation Model for Nodes Appearance in Cape Gooseberry (Physalis peruviana, L.). Rev. Bras. Frutic. 2008, 30, 862–867. [Google Scholar] [CrossRef]
  3. Whiting, M.D.; Salazar, M.R.; Hoogenboom, G. Development of bloom phenology models for tree fruits. Acta Hortic. 2015, 1068, 107–112. [Google Scholar] [CrossRef]
  4. Valentini, N.; Me, G.; Ferrero, R.; Spanna, F. Use of Bioclimatic Indexes to Characterize Phenological Phases of Apple Varieties in Northern Italy. Int. J. Biometeorol. 2001, 45, 191–195. [Google Scholar] [CrossRef] [PubMed]
  5. Viator, R.P.; Nuti, R.C.; Edmisten, K.L.; Wells, R. Predicting Cotton Boll Maturation Period Using Degree Days and Other Climatic Factors. Agron. J. 2005, 97, 494–499. [Google Scholar] [CrossRef]
  6. Legave, J.M.; Farrera, I.; Almeras, T.; Calleja, M. Selecting Models of Apple Flowering Time and Understanding How Global Warming Has Had an Impact on This Trait. J. Hortic. Sci. Biotechnol. 2008, 83, 76–84. [Google Scholar] [CrossRef]
  7. Oh, S.D.; Kang, S.M.; Kim, D.I.; Kim, M.S.; Kim, W.S. Fruit Tree Physiology in Relation to Temperature; Gilmogeum: Seoul, Republic of Korea, 2004. [Google Scholar]
  8. Kim, J.-H.; Lee, E.-J.; Yun, J.I. Prediction of Blooming Dates of Spring Flowers by Using Digital Temperature Forecasts and Phenology Models. Kor. J. Agric. For. Meteorol. 2013, 15, 40–49. [Google Scholar] [CrossRef]
  9. Menzel, A.; Fabian, P. Growing Season Extended in Europe. Nature 1999, 397, 659. [Google Scholar] [CrossRef]
  10. Ou, S.; Chen, C. Estimation of the Chilling Requirement and Development of a Low-Chill Model for Local Peach Trees in Taiwan. J. Chin. Soc. Hortic. Sci. 2000, 46, 337–350. [Google Scholar]
  11. Cook, B.I.; Wolkovich, E.M.; Parmesan, C. Divergent Responses to Spring and Winter Warming Drive Community Level Flowering Trends. Proc. Natl. Acad. Sci. USA 2012, 109, 9000–9005. [Google Scholar] [CrossRef]
  12. Guo, L.; Dai, J.; Wang, M.; Xu, J.; Luedeling, E. Responses of Spring Phenology in Temperate Zone Trees to Climate Warming: A Case Study of Apricot Flowering in China. Agric. For. Meteorol. 2015, 201, 1–7. [Google Scholar] [CrossRef]
  13. Atkinson, C.J.; Brennan, R.M.; Jones, H.G. Declining Chilling and Its Impact on Temperate Perennial Crops. Environ. Exp. Bot. 2013, 91, 48–62. [Google Scholar] [CrossRef]
  14. Guo, L.; Dai, J.; Ranjitkar, S.; Yu, H.; Xu, J.; Luedeling, E. Chilling and Heat Requirements for Flowering in Temperate Fruit Trees. Int. J. Biometeorol. 2014, 58, 1195–1206. [Google Scholar] [CrossRef] [PubMed]
  15. Legave, J.M.; Guédon, Y.; Malagi, G.; El Yaacoubi, A.; Bonhomme, M. Differentiated Responses of Apple Tree Floral Phenology to Global Warming in Contrasting Climatic Regions. Front. Plant Sci. 2015, 6, 1054. [Google Scholar] [CrossRef] [PubMed]
  16. Scorza, R.; Okie, W.R. PEACHES (PRUNUS). Acta Hortic. 1991, 290, 177–234. [Google Scholar] [CrossRef]
  17. Burroughs, W.J. Gardening and Climate Change. Weather 2002, 57, 151–157. [Google Scholar] [CrossRef]
  18. Kim, J.-H.; Kim, D.; Kim, S.; Yun, E.; Ju, O.; Sun Park, J.; Soon Shin, Y. Estimation of Freeze Damage Risk According to Developmental Stage of Fruit Flower Buds in Spring. Kor. J. Agric. For. Meteorol. 2019, 21, 55–64. [Google Scholar] [CrossRef]
  19. Baldocchi, D.; Wong, S. Accumulated Winter Chill Is Decreasing in the Fruit Growing Regions of California. Clim. Change 2008, 87, 153–166. [Google Scholar] [CrossRef]
  20. Luedeling, E.; Zhang, M.; Girvetz, E.H. Climatic Changes Lead to Declining Winter Chill for Fruit and Nut Trees in California during 1950–2099. PLoS ONE 2009, 4, e6166. [Google Scholar] [CrossRef]
  21. Darbyshire, R.; Webb, L.; Goodwin, I.; Barlow, E.W.R. Impact of Future Warming on Winter Chilling in Australia. Int. J. Biometeorol. 2013, 57, 355–366. [Google Scholar] [CrossRef]
  22. Lang, G.A.; Early, J.D.; Martin, G.C.; Darnell, R.L. Endo-, Para-, and Ecodormancy: Physiological Terminology and Classification for Dormancy Research. HortScience 1987, 22, 371–377. [Google Scholar] [CrossRef]
  23. Rural Development Administration. The Agrometeorological Early Warning System. Available online: https://agmet.kr/ (accessed on 17 August 2025).
  24. Heuvelink, E. Re-Interpretation of an Experiment on the Role of Assimilate Transport Resistance in Partitioning in Tomato. Ann. Bot. 1996, 78, 467–470. [Google Scholar] [CrossRef]
  25. Richardson, E.A.; Seeley, S.D.; Walker, D.R. A Model for Estimating the Completion of Rest for ‘Redhaven’ and ‘Elberta’ Peach Trees1. HortScience 1974, 9, 331–332. [Google Scholar] [CrossRef]
  26. Gilreath, P.R.; Buchanan, D.W. Rest Prediction Model for Low-Chilling ‘Sungold’ Nectarine1. J. Am. Soc. Hortic. Sci. 1981, 106, 426–429. [Google Scholar] [CrossRef]
  27. Aron, R. Availability of Chilling Temperatures in California. Agric. Meteorol. 1983, 28, 351–363. [Google Scholar] [CrossRef]
  28. Shaltout, A.D.; Unrath, C.R. Effect of Some Growth Regulators and Nutritional Compounds as Substitutes for Chilling of “Delicious” Apple Leaf and Flower Buds. J. Am. Soc. Hortic. Sci. 1983, 108, 898–901. [Google Scholar] [CrossRef]
  29. Ono, S.; Konno, T. Estimation of Flowering Date and Temperature Characteristics of Fruit Trees by DTS Method. Jpn. Agric. Res. Q. 1999, 33, 105–108. [Google Scholar]
  30. Cesaraccio, C.; Spano, D.; Snyder, R.L.; Duce, P. Chilling and Forcing Model to Predict Bud-Burst of Crop and Forest Species. Agric. For. Meteorol. 2004, 126, 1–13. [Google Scholar] [CrossRef]
  31. Dennis, F.G. Problems in Standardizing Methods for Evaluating the Chilling Requirements for the Breaking of Dormancy in Buds of Woody Plants. HortScience 2003, 38, 347–350. [Google Scholar] [CrossRef]
  32. Kim, S.-O.; Kim, J.-H.; Chung, U.-R.; Kim, S.-H.; Park, G.-H.; Yun, J.-I. Quantification of Temperature Effects on Flowering Date Determination in Niitaka Pear. Kor. J. Agric. For. Meteorol. 2009, 11, 61–71. [Google Scholar] [CrossRef]
  33. Han, J.H.; Cho, K.S.; Choi, J.J.; Hwang, H.S.; Gook Kim, C.; Kim, T.-C. Estimation of Changes in Full Bloom Date of “Niitaka” Pear Tree with Global Warming. Hortic. Sci. Technol. 2010, 28, 937–941. [Google Scholar]
  34. Chuine, I.; Cour, P.; Rousseau, D.D. Selecting Models to Predict the Timing of Flowering of Temperate Trees: Implications for Tree Phenology Modelling. Plant Cell Environ. 1999, 22, 1–13. [Google Scholar] [CrossRef]
  35. Primack, R.B.; Ibáñez, I.; Higuchi, H.; Lee, S.D.; Miller-Rushing, A.J.; Wilson, A.M.; Silander, J.A. Spatial and Interspecific Variability in Phenological Responses to Warming Temperatures. Biol. Conserv. 2009, 142, 2569–2577. [Google Scholar] [CrossRef]
  36. Wang, J.; Song, G.; Liddell, M.; Morellato, P.; Lee, C.K.F.; Yang, D.; Alberton, B.; Detto, M.; Ma, X.; Zhao, Y.; et al. An Ecologically-Constrained Deep Learning Model for Tropical Leaf Phenology Monitoring Using PlanetScope Satellites. Remote. Sens. Environ. 2023, 286, 113429. [Google Scholar] [CrossRef]
  37. Toomey, M.; Friedl, M.A.; Frolking, S.; Hufkens, K.; Klosterman, S.; Sonnentag, O.; Baldocchi, D.D.; Bernacchi, C.J.; Biraud, S.C.; Bohrer, G.; et al. Greenness Indices from Digital Cameras Predict the Timing and Seasonal Dynamics of Canopy-scale Photosynthesis. Ecol. Appl. 2015, 25, 99–115. [Google Scholar] [CrossRef]
  38. Richardson, A.D.; Braswell, B.H.; Hollinger, D.Y.; Jenkins, J.P.; Ollinger, S.V. Near-surface Remote Sensing of Spatial and Temporal Variation in Canopy Phenology. Ecol. Appl. 2009, 19, 1417–1428. [Google Scholar] [CrossRef]
  39. Rural Development Administration. The Fruit Tree Growth and Quality Management System. Available online: https://fruit.nihhs.go.kr/ (accessed on 31 October 2023).
  40. Meier, U. Growth Stages of Mono-and Dicotyledonous Plants—BBCH Monograph; Julius Kühn-Institut: Berlin, Germany, 2001. [Google Scholar]
  41. Miller, P.; Lanier, W.; Assistant, I.; Brandt, S.; Canada, A. Using Growing Degree Days to Predict Plant Stages. In Ag/Extension Communications Coordinator, Communications Services; Montana State University-Bozeman: Bozeman, MO, USA, 2001; Volume 59717, pp. 994–2721. [Google Scholar]
  42. Müller, M.; Braun, P. Development of Phenological Models over Time–a Review. Acta Hortic. 2008, 803, 111–116. [Google Scholar] [CrossRef]
  43. National Agricultural Technology Research Institute. Climatic Characteristics of Major Fruit Growing Areas; Rural Development Administration: Jeonju, Republic of Korea, 1990. [Google Scholar]
  44. Sugiura, T.; Honjo, H. The Effects of Temperature on Endodormancy Completion in Japanese Pear (Pyrus Pyrifolia Nakai) and Modeling the Relationship. J. Agric. Meteorol. 1997, 53, 285–290. [Google Scholar] [CrossRef]
  45. Sugiura, T. Prediction of Full Bloom Date of Pear Using Air Temperature. Agric. Hortic. 1999, 54, 146–149. [Google Scholar]
  46. Hauagge, R.; Cummins, J.N. Phenotypic Variation of Length of Bud Dormancy in Apple Cultivars and Related Malus Species. J. Am. Soc. Hortic. Sci. 1991, 116, 100–106. [Google Scholar] [CrossRef]
  47. Erez, A. Means to Compensate for Insufficient Chilling to Improve Bloom and Leafing. Acta Hortic. 1994, 395, 81–96. [Google Scholar] [CrossRef]
  48. Chandler, W.H.; Kimball, M.H.; Philp, G.L.; Tufts, W.P.; Weldon, G.P. Chilling Requirements for Opening of Buds on Deciduous Orchard Trees and Some Other Plants in California. Univ. Calif. Agr. Expt. Sta. Bull. 1937, 611, 63. [Google Scholar]
  49. Alburquerque, N.; García-Montiel, F.; Carrillo, A.; Burgos, L. Chilling and Heat Requirements of Sweet Cherry Cultivars and the Relationship between Altitude and the Probability of Satisfying the Chill Requirements. Environ. Exp. Bot. 2008, 64, 162–170. [Google Scholar] [CrossRef]
  50. Nash, J.E.; Sutcliffe, J.V. River Flow Forecasting through Conceptual Models Part I—A Discussion of Principles. J Hydrol. 1970, 10, 282–290. [Google Scholar] [CrossRef]
  51. Nossent, J.; Bauwens, W. Application of a Normalized Nash-Sutcliffe Efficiency to Improve the Accuracy of the Sobol’ Sensitivity Analysis of a Hydrological Model. In Proceedings of the EGU general assembly conference abstracts, Vienna, Austria, 22–27 April 2012; Volume 14, p. 237. [Google Scholar]
  52. Taylor, K.E. Summarizing Multiple Aspects of Model Performance in a Single Diagram. J. Geophys. Res. Atmos. 2001, 106, 7183–7192. [Google Scholar] [CrossRef]
  53. Hunter, A.F.; Lechowicz, M.J. Predicting the Timing of Budburst in Temperate Trees. J. Appl. Ecol. 1992, 29, 597. [Google Scholar] [CrossRef]
  54. Chuine, I.; Cour, P.; Rousseau, D.D. Fitting Models Predicting Dates of Flowering of Temperate-zone Trees Using Simulated Annealing. Plant Cell Environ. 1998, 21, 455–466. [Google Scholar] [CrossRef]
Figure 1. Geographic distribution of the study orchards in South Korea and schematic workflow for full bloom date determination and model evaluation. Red dots indicate the locations of the seven ‘Niitaka’ pear orchards monitored under the Fruit Tree Growth and Quality Management System (FTGQMS) from 2017 to 2022. Orchard imagery was used to determine full bloom dates based on the BBCH scale, and AWS meteorological data were utilized as input variables for phenology models.
Figure 1. Geographic distribution of the study orchards in South Korea and schematic workflow for full bloom date determination and model evaluation. Red dots indicate the locations of the seven ‘Niitaka’ pear orchards monitored under the Fruit Tree Growth and Quality Management System (FTGQMS) from 2017 to 2022. Orchard imagery was used to determine full bloom dates based on the BBCH scale, and AWS meteorological data were utilized as input variables for phenology models.
Atmosphere 16 00996 g001
Figure 2. Z-score standardized full bloom dates and thermal unit accumulations across seven pear orchard regions in South Korea from 2017 to 2022. The top heatmap shows observed full bloom dates (DOY), while the middle and bottom panels show chill and heat units calculated from three phenology models: the 7 °C/5 °C threshold method, the modified development rate (mDVR) model, and the Chill Days (CD) model. Red and blue indicate values higher and lower than the multi-year regional mean, respectively.
Figure 2. Z-score standardized full bloom dates and thermal unit accumulations across seven pear orchard regions in South Korea from 2017 to 2022. The top heatmap shows observed full bloom dates (DOY), while the middle and bottom panels show chill and heat units calculated from three phenology models: the 7 °C/5 °C threshold method, the modified development rate (mDVR) model, and the Chill Days (CD) model. Red and blue indicate values higher and lower than the multi-year regional mean, respectively.
Atmosphere 16 00996 g002
Figure 3. Mean prediction error (ME, in days) of full bloom date estimates for the DVR, mDVR, and CD models across seven orchard regions from 2018 to 2022. The heatmaps show the difference between predicted and observed full bloom dates by region and year. Positive values indicate late predictions, while negative values indicate early predictions.
Figure 3. Mean prediction error (ME, in days) of full bloom date estimates for the DVR, mDVR, and CD models across seven orchard regions from 2018 to 2022. The heatmaps show the difference between predicted and observed full bloom dates by region and year. Positive values indicate late predictions, while negative values indicate early predictions.
Atmosphere 16 00996 g003
Figure 4. Taylor diagrams illustrating the performance of the DVR, mDVR, and CD models in simulating full bloom dates across seven orchard sites in South Korea. Each diagram presents the correlation coefficient (angle), normalized standard deviation (radius), and root-mean-square error (RMSE, color scale). The bottom-right panel (“All the sites”) shows the overall performance based on aggregated data from all regions.
Figure 4. Taylor diagrams illustrating the performance of the DVR, mDVR, and CD models in simulating full bloom dates across seven orchard sites in South Korea. Each diagram presents the correlation coefficient (angle), normalized standard deviation (radius), and root-mean-square error (RMSE, color scale). The bottom-right panel (“All the sites”) shows the overall performance based on aggregated data from all regions.
Atmosphere 16 00996 g004
Table 1. Coefficient of variation (CV, %) of chill and heat unit estimates across seven regions and three models, based on five years of data (2017–2022). The table compares interannual variability among the 7 °C/5 °C threshold method, the modified development rate (mDVR) model, and the Chill Days (CD) model. CV was calculated as (σ/μ) × 100 for each region and model.
Table 1. Coefficient of variation (CV, %) of chill and heat unit estimates across seven regions and three models, based on five years of data (2017–2022). The table compares interannual variability among the 7 °C/5 °C threshold method, the modified development rate (mDVR) model, and the Chill Days (CD) model. CV was calculated as (σ/μ) × 100 for each region and model.
RegionChill Units (CV, %)Heat Units (CV, %)
Hours < 7 °CmDVR ModelCD ModelHours > 5 °CmDVR ModelCD Model
Icheon5.19.215.91318.517.2
Cheonan6.56.112.715.418.817.7
Sangju5.4716.39.88.78.7
Yeongcheon6.7513.711.77.97.3
Wanju7.52.67.513.813.814.1
Naju8.74.37.410.812.99.8
Sacheon135.34.811.911.610.5
Avg.13.36.312.119.519.316.4
Table 2. Model performance for full bloom date prediction of ‘Niitaka’ pear across seven orchard regions in South Korea (2018–2022). RMSE (in days) and EF (Nash–Sutcliffe efficiency) were computed using observed and predicted values across all years per region. Lower RMSE and higher EF values indicate better model performance.
Table 2. Model performance for full bloom date prediction of ‘Niitaka’ pear across seven orchard regions in South Korea (2018–2022). RMSE (in days) and EF (Nash–Sutcliffe efficiency) were computed using observed and predicted values across all years per region. Lower RMSE and higher EF values indicate better model performance.
RegionDVRmDVRCD
RMSEEFRMSEEFRMSEEF
Icheon5−0.903.30.174.2−0.34
Cheonan5.1−0.862.80.453.20.28
Sangju2.9−0.893−1.024.9−4.26
Yeongcheon4.9−4.662.4−0.426.8−10.04
Wanju7.3−2.063.60.232.60.59
Naju6.1−2.532.80.295.3−1.59
Sacheon9.3−8.651.60.712.60.25
Average6.1−0.572.90.664.50.16
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Kim, J.-H.; Yun, E.-J.; Kang, D.G.; Han, J.-H.; Shim, K.-M.; Kim, D.-J. Evaluation of Phenology Models for Predicting Full Bloom Dates of ‘Niitaka’ Pear Using Orchard Image-Based Observations in South Korea. Atmosphere 2025, 16, 996. https://doi.org/10.3390/atmos16090996

AMA Style

Kim J-H, Yun E-J, Kang DG, Han J-H, Shim K-M, Kim D-J. Evaluation of Phenology Models for Predicting Full Bloom Dates of ‘Niitaka’ Pear Using Orchard Image-Based Observations in South Korea. Atmosphere. 2025; 16(9):996. https://doi.org/10.3390/atmos16090996

Chicago/Turabian Style

Kim, Jin-Hee, Eun-Jeong Yun, Dae Gyoon Kang, Jeom-Hwa Han, Kyo-Moon Shim, and Dae-Jun Kim. 2025. "Evaluation of Phenology Models for Predicting Full Bloom Dates of ‘Niitaka’ Pear Using Orchard Image-Based Observations in South Korea" Atmosphere 16, no. 9: 996. https://doi.org/10.3390/atmos16090996

APA Style

Kim, J.-H., Yun, E.-J., Kang, D. G., Han, J.-H., Shim, K.-M., & Kim, D.-J. (2025). Evaluation of Phenology Models for Predicting Full Bloom Dates of ‘Niitaka’ Pear Using Orchard Image-Based Observations in South Korea. Atmosphere, 16(9), 996. https://doi.org/10.3390/atmos16090996

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop