3.1 Performance evaluation of the FFD estimation models
First, the three appropriate parameters for DTS, GDD, and CD models are determined to estimate FFD_{peach} and FFD_{pear} in South Korea. Considering the average FFD_{peach} (JD: 98.5) and FFD_{pear} (JD: 102.3) for 30 years, we set D_{s} at JD 18–40 for peach and JD 22–44 for pear with 2day intervals in the DTS and GDD models. To identify the most suitable parameters in DTS, we set E_{a} from 40–76 kJmol^{−1} at 4 kJmol^{−1} intervals, and calculate DTS with 120 combinations [12 (the number of D_{s}) × 10 (the number of E_{a})]. In the case of GDD, T_{b} is set from 0 to 10 °C at 1 °C intervals, giving a total of 132 combinations [12 (the number of D_{s}) × 11 (the number of T_{b})]. Then, FFD_{peach} and FFD_{pear} are estimated using each combination and the corresponding DTS and GDD. The combination with the lowest root mean square error (RMSE) between the observed and estimated FFDs is decided as the most appropriate (Snyder et al., 1999; Cesaraccio et al., 2004; Aono and Kazui, 2008; Hur et al., 2014). According to the result, DTS has the lowest RMSE for peach and pear in South Korea when E_{a} is 72 and 64 kJmol^{−1} and D_{s} is JD 34 and 44, respectively (Figure 3). GDD has the lowest RMSE when T_{b} is 0 °C and D_{s} is JD 40 for peach and JD 44 for pear (Fig. 4). The estimated D_{s} in the study is early to mid February for both trees. According to the physiological interpretation of Ono and Konno (1999), this period roughly corresponds to that of endodormancy release.
A twostep procedure is used to estimate the optimal parameters of the CD model. The parameter optimization method is basically the same as the one described in Jung et al. (2005) who suggest flowering model by extending budburst model of Cesaraccio et al. (2004). First, the optimal C_{r} and T_{b} are selected using BBD data by considering that BBD occurs when ∑Ca ≥ −1 × C_{r}. BBD is defined when over 20% of each tree's floral buds burst open. The BBD data used in this study are taken from KMA observations, which have the same observation sites and period as those of FFD. C_{r} is set from −90 to −110 Cd at −1 Cd intervals. T_{b} is set from 5 to 11 °C at 1 °C intervals by taking failure rates on budburst open into consideration. In detail, the failure rate on budburst open is over 20% when T_{b} is below 4 °C or above 12 °C. Therefore, BBD is estimated with 147 treatments [21 (the number of C_{r}) × 7 (the number of T_{b})]. Then, RMSE is calculated for each combination in the estimation of BBD for 30 years. The best combination of C_{r} and T_{b} is selected as the one that affords the smallest RMSE between the observed and estimated BBDs. According to the result, CD has the lowest RMSE in South Korea when C_{r} are −99 and −106 Cd and T_{b} is 5 °C for peach and pear, respectively (Figure 5). In the second step, the most suitable H_{r} is determined using C_{r} (−99 and −106 Cd for peach and pear, respectively) and T_{b} (5 °C) selected in step 1. When we taking into account that FFD occurs when ∑Ca ≥ H_{r}, annual H_{r} is calculated by accumulating Ca from endodormancy release (the day when ∑Cd ≤ C_{r}) till FFD observed at each year. The average H_{r} for 30 years is selected as the optimal parameter. Consequently, 183.3 and 199.4 Ca are selected as H_{r} for peach and pear, respectively. Table 3 shows three parameters for each model and species that were determined by RMSE analysis to be the most suitable study values.
Table 3. Three parameters for DTS, GDD, and CD models that were determined to be the most suitable study values by the analysis of root mean square errors (RMSE).  Peach  Pear 

DTS  D_{s} (JD)  34  44 
E_{a} (kJ/mol)  72  64 
DTS (days)  162.7  145.8 
GDD  D_{s} (JD)  40  44 
T_{b} (°C)  0  0 
GDD (GDD)  361.1  395.5 
CD  T_{b} (°C)  5  5 
C_{r} (Cd)  −99  −106 
H_{r} (Ca)  183.3  199.4 
Model performance is evaluated based not only on quantitative estimation such as temporal correlation coefficient (TCC) and RMSE but also on categorical estimation such as Hit Rate (HR) and Heidke Skill Score (HSS). Evaluation is performed at each station and averaged over 50 stations. HR and HSS are calculated using three categories based on one standard deviation (6 days) of FFD_{peach} and FFD_{pear}: below normal (< −6 day), normal (≥ −6 day and ≤6 day) and above normal (>6 day). Table 4 shows the skill scores and average of FFD_{peach} and FFD_{pear} derived from the DTS, GDD, and CD models using observed temperature for 30 years from 1981 to 2010. The observed average FFD_{peach} and FFD_{pear} are 98.5 and 102.3 days, respectively, indicating that peach and pear generally flower in early or midApril in South Korea. Average FFD_{peach} and FFD_{pear} estimated by DTS, GDD, and CD are similar to the observation with relatively slight margins of about 2 days. DTS has better skill than GDD and CD in terms of RMSE and TCC, although even GDD and CD have sufficiently low RMSE and high TCC with 99% confidence level for FFD_{peach}. As for the two categorical estimation concerns, DTS shows better estimation ability for peach in South Korea because HR and HSS are closer to 1 compared to the other models. For pear, DTS also shows lower RMSE and higher TCC, HR, and HSS. These results confirm the validity of these three methods for estimating FFD_{peach} and FFD_{pear} in South Korea. In particular, FFD derived from DTS is more similar to the observed data than that derived from GDD and CD according to various evaluations. This agrees with the claims of Aono and Kazui (2008) that DTS is more appropriate than GDD for estimating cherry FFD in Japan. Even though the CD model is a more mechanistic approach than DTS and GDD, its estimation ability for FFD_{peach} and FFD_{pear} in South Korea is lower than that of the others due to uncertainties arising from many factors such as the dormancy onset and release, and budburst. Therefore, the rather simple DTS is chosen in this study as the phenological model and is applied to the simulated temperature in order to estimate future FFD changes over South Korea.
Table 4. Average and skill scores of FFDs derived from DTS, GDD, and CD models using observation during 30 years from 1981 to 2010.  Observation  DTS  GDD  CD 


Peach  Average (JD)  98.5  99.3  99.2  100.7 
RMSE (day)  –  3.66  3.90  5.08 
TCC  –  0.72a  0.71a  0.67a 
HR  –  0.73  0.71  0.70 
HSS  –  0.58  0.54  0.52 
Pear  Average (JD)  102.3  103.0  102.9  104.5 
RMSE (day)  –  3.76  4.20  5.18 
TCC  –  0.69a  0.68a  0.59a 
HR   0.75  0.72  0.67 
HSS   0.62  0.56  0.47 
3.2 Change of FFD_{peach} and FFD_{pear}
To select a target season and to determine the temperature dependency of FFD, the relationship between temperature and the FFD of each fruit tree is examined using the 30year observation data. Figure 6 shows the average TCC between the 10day average temperature from January to April and FFD_{peach} and FFD_{pear}. This period was selected as many plants are dormant and start flowering in South Korea (Jeong et al., 2011). There is a statistically significant (p < 0.05) negative correlation between the temperatures of February to April and FFD_{peach} and FFD_{pear}. This strong temperature dependency of the two FFD values during these 3 months, which we term early spring, led us to investigate the changes in early spring temperature.
Changes in early spring temperature in association with global warming and the corresponding DTS variations of each month are estimated using simulated daily temperature. Table 5 shows the monthly average temperature and accumulated DTS derived from observation (1986–2005), Historical (1986–2005) simulation, and RCP (2071–2090) 4.5 and 8.5 simulations. In climatology, the monthly average temperature increases by 4.6 and 10.8 K from 274.9 K in February to 279.5 and 285.7 K in March and April, respectively. By 2090, the average temperatures over South Korea simulated under the RCP 4.5 (RCP 8.5) scenario are anticipated to increase by about 2.1 K (3.7 K) in February, 1.8 K (3.2 K) in March, and 1.7 K (3.0 K) in April. As a result, the temperature increase is higher under the RCP 8.5 scenario than under RCP 4.5 and higher in February than in April. This is attributed to the snow albedo feedback mentioned in Ohashi and Tanaka (2010) and Im and Ahn (2011). According to their study analysis, melted snow in high elevation causes decreased albedo and increased insolation, which means that the temperature changes in winter (December to February, DJF) can be larger than those in other seasons in South Korea. This result agrees with the analysis of Im et al. (2008), who found that winter (DJF) exhibits a larger temperature change than does summer (June to August, JJA) under the SRES B2 scenario.
Table 5. Accumulated DTS and mean temperature for three months from February to April at 50 stations over South Korea.  February  March  April 

Accumulated DTS (days)  Peach  Observation  44  80  39 
Historical  42  80  44 
RCP 4.5  52  92  21 
RCP 8.5  62  96  8 
Pear  Observation  27  73  46 
Historical  25  72  50 
RCP 4.5  31  85  32 
RCP 8.5  36  95  18 
Mean temperature (K)  Observation  274.9  279.5  285.7 
Historical  274.9  279.5  285.7 
RCP 4.5  277.0  281.3  287.4 
RCP 8.5  278.6  282.7  288.7 
In the observation and Historical simulations, accumulated monthly DTS is the highest in March. The high level of DTS accumulation in February and March implies that the daily average temperature for the period is warm enough for plants to grow after D_{s} (JD: 40 and 44 for peach and pear, respectively), which roughly corresponds to the date of endodormancy release (Ono and Konno, 1999). Moreover, the DTS accumulation in April is smaller than that in February and March, despite the higher average temperature, because sufficient DTS is accumulated during the preceding 2 months, leaving only a small amount of accumulable DTS remaining. This means that the temperature of the two preceding months has a greater effect on flowering, although the average temperature in the period is lower than that in April. The DTS accumulation derived from the RCP simulations tends to decrease in April but to increase in February and March. Accumulated DTS for peach (pear) tree is expected to increase by about 10 and 20 days (6 and 11 days) in February, and by 12 and 16 days (13 and 23 days) in March, compared to the Historical simulation under RCP 4.5 and 8.5 scenarios, respectively. On the other hand, it is expected to decrease by about 23 and 36 days (18 and 32 days) in April. Because the total DTS amount is fixed, the increased temperature and DTS accumulation in February and March lead to decreased DTS accumulation in April. In other words, as early spring temperature rises under global warming, FFD_{peach} and FFD_{pear} will become increasingly affected by February and March temperature but less affected by April temperature. This means that the floral development will be accelerated in the two preceding months, in agreement with Chung et al. (2011).
Figure 7 shows the average and standard deviation of FFD_{peach} and FFD_{pear} derived from observation and climate models. Observed current FFD_{peach} and FFD_{pear} are JD 98.1 and 102.2, respectively, on average. This indicates that these two trees mostly start flowering in midApril. FFD_{peach} moves forward by about 7.0 and 12.7 days compared to the Historical simulation by 2090 under RCP 4.5 and 8.5 scenarios, respectively, so that peach is expected to flower in late March or early April in 2090. Considering that the FFD_{peach} trend observed from 1954 to 2004 is −2.46 days °C^{−1} (Jeong et al., 2011), the average advances (−3.68 and −3.84 days °C^{−1} ) in the RCP 4.5 and 8.5 simulations are −1.22 and −1.38 days °C^{−1} higher, respectively, than the observation. As in the case of peach, FFD_{pear} is expected to advance by 6.1 and 10.7 days in 2090 with a negative trend towards temperature of −3.21 and −3.24 days °C^{−1} under RCP 4.5 and 8.5 scenarios, respectively. Assuming steady advances, the average annual FFD advances in the RCP 4.5 and 8.5 simulations will be 0.08 and 0.15 days year^{−1} for peach and 0.07 and 0.13 days year^{−1} for pear, respectively. This result shows that the increase in February and March temperature will accelerate the growing speed of peach and pear trees and advance the flowering date.
Kim et al. (2013) determined that the average FFD of forsythia, azalea, and cherry blossom will advance by 25 days from 2071 to 2100 under RCP 8.5 scenario, which is more than 10 days faster than the advance estimated in this study. Even with due regard to the possible discrepancy in various factors such as climate data, phenological model, and tree type, this is a large difference. Considering that those plants flower earlier than peach and pear, the difference can be attributed to the insistence of Roetzer et al. (2000) that earlyflowering species are more variable in flowering time than lateblooming species.
The observed FFD_{peach} and FFD_{pear} have an average standard deviation of 6.2 and 6.3 days, respectively. All simulations, including MME results, have lower variations than that of observation. This is a general characteristic of climate prediction models that underestimate the fluctuations of variables such as temperature (Ines and Hansen, 2006; Hur et al., 2014). Moreover, the change in the standard deviations of FFDs did not exhibit any relationship with global warming in this study.
The altitude dependency of FFD_{peach} and FFD_{pear} changes is also investigated (Figure 8). FFD_{peach} and FFD_{pear} at high altitude are delayed compared to those at low altitude due to lower daily average temperature. The temperature change rate with the altitude of observation is −1.17, −1.16, −1.13, −1.11 °C/100 m in Historical, RCP 4.5, and RCP 8.5 simulations, respectively, indicating that the high altitude temperature increases slightly more than the low altitude temperature under global warming. According to Student's ttest, the variation of the temperature change rate with altitude under RCP 4.5 and 8.5 scenarios is statistically significant at 82 and 65% confidence levels, respectively. This characteristic of temperature change with altitude is in agreement with the results of Im and Ahn (2011), who attributed it to the snowalbedo feedback mechanism. This characteristic infers that the FFD delay with increasing altitude in RCP 4.5 and 8.5 simulations will be reduced compared to the Historical simulation. However, unlike our inference, the slopes of FFD_{peach} in the two RCP simulations are steeper than that in the Historical simulation in the peach case. While the FFD_{peach} change with altitude occurs more steeply in the future projection, the slope of FFD_{pear} with elevation shows a decrease of 0.10 and 0.11 day/100 m in RCP 4.5 and 8.5 simulations, respectively, compared to the Historical simulation (Figure 8). Hur et al. (2014) presented similar results and gave two explanations. Firstly, all stations used for the analysis are located under 280 m elevation, which explains the limit in clearly explaining the altitude dependency of FFD change. Secondly, flowering is affected not only by temperature but also by other environmental variables such as daylength, moisture, and solar radiation. Therefore, FFD change will not show a linear correlation with temperature variation (Diekmann, 1996; Tyler, 2001; Yeang, 2007).
The spatial distribution of FFD_{peach} derived from observation and simulations is shown in Figure 9. Observed FFD_{peach} on average is JD 98.1, which reflects the topographical signal. FFD_{peach} is earlier at low altitude and in flatland than at high altitude and in mountainous regions. The Historical simulation successfully simulates the spatial pattern and general characteristics of the observation in qualitative terms but gives an estimated FFD_{peach} that is 2.4 days later than observation in quantitative terms. The JD 90 line is located around 33.5°N in the observation and Historical simulation at 126.5°E (the location of Jeju island off the southern tip of the Korean peninsula), indicating that peach flowers in March over the region. The areas with values lower than JD 90 account for 3.8 and 1.8% of the land area of South Korea in the observation and Historical simulation, respectively. FFD_{peach} under RCP 4.5 and 8.5 scenarios is uniformly advanced over all stations, irrespective of the altitude, compared to that of the Historical simulation (Figure 8), while maintaining a topographical effect on FFD. The JD 90 line moves northward by 1°N and 2.5°N to 34.5°N and 36.0°N at 126.5°N under RCP 4.5 and 8.5 scenarios, respectively, 85 years later. This indicates that the two RCP simulations have a northward moving speed of approximately 0.01 and 0.03°N year^{−1}, respectively. Therefore, areas with values lower than JD 90 are increased to about 15.6 and 45.5% in RCP 4.5 and 8.5 simulations, respectively.
The spatial distribution of FFD_{pear} derived from observation and simulations is shown in Figure 10. Observed FFD_{pear} on average is JD 102.2, which well reflects the topographical signal, as in the peach case. FFD_{pear} in the Historical simulation is JD 104.0, which is later on average than in the observation, indicating that the climate models underestimate the flowering time in general. In qualitative terms, however, the model captures the spatial pattern of observation in that FFD_{pear} appears relatively early in the southern and eastern coasts. FFD_{pear} derived from RCP 4.5 and 8.5 simulations advances by 6.1 and 10.7 days, respectively, in all stations in 2090. Quantitative analysis reveals that the JD 90 line, which is not shown in the observation and Historical simulation, appears at 33.5°N and 34.5°N at 126.5°E under RCP 4.5 and 8.5 scenarios, respectively. The areas with values lower than JD 90 are increased from 0 to about 1.8 and 12.8% of the land area of South Korea in RCP 4.5 and 8.5 simulations, respectively. This implies that pear is expected to start flowering in late March by the end of this century, compared to midApril on average these days.