Statistical Analyses of the Flowering Dates of Cherry Blossom and the Peak Dates of Maple Leaves in South Korea Using ASOS and MODIS Data

Abstract

In this paper, we aimed to examine the flowering dates of cherry blossom and the peak dates of maple leaves in South Korea, by the combination of temperature observation data from ASOS (Automated Surface Observing System) and NDVI (Normalized Difference Vegetation Index) from MODIS (Moderate Resolution Imaging Spectroradiometer). The more recent years, the faster the flowering dates and the slower the peak dates. This is because of the impacts of climate change with the increase of air temperature in South Korea. By reflecting the climate change, our statistical models could reasonably predict the plant phenology with the CC (Correlation Coefficient) of 0.870 and the MAE (Mean Absolute Error) of 3.3 days for the flowering dates of cherry blossom, and the CC of 0.805 and the MAE of 3.8 for the peak dates of maple leaves. We could suppose a linear relationship between the plant phenology DOY (day of year) and the environmental factors like temperature and NDVI, which should be inspected in more detail. We found that the flowering date of cherry blossom was closely related to the monthly mean temperature of February and March, and the peak date of maple leaves was much associated with the accumulated temperature. Amore sophisticated future work will be required to examine the plant phenology using higher-resolution satellite images and additional meteorological variables like the diurnal temperature range sensitive to plant phenology. Using meteorological grid can help produce the spatially continuous raster maps for plant phenology.

1. Introduction

Current global warming leads to the elevation of average temperature and abnormal heatwaves in some regions. Various plants sensitive to temperature were affected, and the changes in the environment and ecosystem can be severe (Kim et al., 2012). The vegetation type, distribution, and growth can vary by season. Plant phenology refers to the periodical changes of the plant activities such as germination, flowering, fresh greenness, maturation, maple leaves, and fallen leaves dependent on the season (Lee et al., 2009). Plant phenology repeats every year but is influenced by the effects of accumulated meteorology (Kim et al., 2013).

Because the changes in plant phenology can have different responses according to regions, we can acquire important information to assess the vulnerabilityto climate change through the long-term observations of the regional differences in plant phenology (Lee et al., 2009). The changes in plant phenology are a very crucial factor in the ecosystem where a variety of species have interactions. Changes in the primary producers (autotrophs) with photosynthesis can affect the dependent consumers (heterotrophs), which can bring about an unbalance in the ecosystem (Edwards and Richardson, 2004; Jo and Ahn, 2008). The plant growth cycle is also associated with soil nutrition and carbon fixation (Fridley, 2012).

Studies in plant phenology to assess the impacts of climate change have long been conducted using temperature observation data and recently employed remote sensing data (Johnson et al., 2008). The dates of flowering and open leaves are highly related to temperature. The dates of maple leaves and fallen leaves are associated with temperature and other environmental factors such as day length and moisture conditions (Lee et al., 2009; Kim, 2019). Schwartz and Reiter (2000) have defined the spring indices for germination and first bloom to examine the changes in the spring of North America using daily temperature minima and maxima. Because satellite vegetation indices can appropriately express the changes in spring temperature, many studies have been conducted to examine the spatial and temporal patterns and the cycles of plant phenology using NDVI (Normalized Difference Vegetation Index) and EVI (Enhanced Vegetation Index) (Sakamoto et al., 2005; Beck et al., 2006; Atkinson et al., 2012; Kim et al., 2013; Choi and Jung, 2014; Lee et al., 2018; Reed et al., 1994; Moulin et al., 1997; Zhang et al., 2003; Zhang et al., 2004; Fisher et al., 2006; Soudani et al., 2008; Tan et al., 2011).

To date, the studies on the plant phenology in South Korea have used temperature observation data and satellite vegetation data separately, while both data have been rarely combined for the cases studies of plant phenology. In this paper, we aimed to examine the flowering dates of cherry blossom and the peak dates of maple leaves, which are the major plant phenology events in South Korea, by the combination of temperature observation data from ASOS (Automated Surface Observing System) and NDVI data from MODIS (Moderate Resolution Imaging Spectroradiometer). The phenology data for cherry blossom were 912 cases collected from 121 ASOS stations, and the data for maple trees were 821 cases obtained from 121 ASOS stations during 2003-2020. We built MLR (Multiple Linear Regression) and RF (Random Forest) models to analyze the trend of plant phenology for the 18 years and examined the trend of the plant phenology changes.

2. Data and Methods

1) Plant Phenology Observation Data

KMA (Korea Meteorological Administration) provides the plant phenology observation data by the visual check on the vegetation around the ASOS stations, such as cosmos, azalea, forsythia, plum blossom, cherry blossom, ginkgo, and maple tree. We analyzed the flowering dates of cherry blossom and the peak dates of the maple leaves, which are the major events people are interested in. Fig. 1(a) and 1(b) show the ASOS stations for observation of cherry blossom and maple leaves, respectively. We first aggregated the 912 cases for cherry blossom and 821 cases for maple leaves during 1991-2020 in the form of DOY (date of year). The scatter plots and trend lines in Fig. 2(a) and 3(a) indicate that the flowering dates of cherry blossom are getting faster 0.2 days per year, and the peak dates of maple leaves are getting slower 0.238 days per year, respectively, although there are differences by region or latitude (Fig. 2(b) and 3(b)).

Fig. 1. Locations of the stations to observe the phenology for (a) cherry blossom and (b) maple leaves.

Fig. 2. Scatter plots of the cherry blossom flowering dates for (a) all regions and (b) each region.

Fig. 3. Scatter plot of the maple leaves peak dates for (a) all regions and (b) each region.

2) ASOS Meteorological Observation Data

ASOS, an automated system for synoptic meteorological observation, provides the variables such as temperature, rainfall, wind, pressure, humidity, radiation, cloud, and evaporation. We gathered daily temperature data and conducted various processing steps to derive temperature, cumulative temperature, GDD (Growing Degree Days), and CDD (Cooling Degree Days). In the case of cherry blossom, the cumulative temperature is the sum of air temperature from January 1 to the flowering date, and GDD denotes the sum of air temperature over 5 degrees Celsius from January 1 to the flowering date. In the case of maple leaves, the cumulative temperature is the sum of air temperature from January 1 to the peak date, and CDD denotes the sum of air temperature under 20 degrees Celsius from July 1 to the peak date (Kim et al., 2019).

3) Satellite Vegetation Index

Green vital vegetation reflects more NIR (near- infrared) rays and absorbs more red rays. So, the NDVI is expressed using the NIR and red reflectance ((NIR –RED)/(NIR + RED)). Higher NDVI means higher vitality, and lower NDVI indicates lower vitality. It ranges -1 to 1, and usually has a positive value on the ground. MOD13A3 and MOD13C₂ are the monthly MODIS NDVI products with fewer null pixels through a gap-filling process, of which spatial resolution is 1 km and 5 km, respectively. The two products had almost the same values on the pixels for the plant phenology observations. Because the MOD13A3 had some null pixels on a few stations, but the MOD13C₂ had none, we decided to use the MOD13C₂. This product has been provided since 2002, and we used the data from 2003 to 2020.

4) Linear and Non-linear Modeling

Most previous studies used only temperature or only NDVI to examine the changes in plant phenology. However, we used both temperature and NDVI because temperature is a driver of plant phenology, and NDVI can be an indicator. We used MLR and RF models to estimate the phenology DOY (dependent variable) for cherry blossom and maple leaves using temperature and NDVI (independent variables) because the relationships can be linear or can be non-linear. RF is a non-linear machine learning method to produce an ensemble prediction by summarizing the results of a number of decision trees. A bootstrap is carried out for resampling if necessary for the suitability of the sample distribution. Then, bagging (bootstrap aggregating) extracts the ensemble average from the bootstrap (Hong et al., 2017; Li et al., 2002). For accuracy evaluation, the statistical measures such as MBE (Mean Bias Error), MAE (Mean Absolute Error), nRMSE (Normalized Root Mean Square Error), and CC were used.

3. Result and Discussion

The flowering dates of cherry blossom are usually between the DOY 80 and 120 (approximately April), and the peak dates of maple leaves are between the DOY 290 and 330 (approximately November). The higher the latitude, the later the flowering dates of cherry blossom and the earlier the peak dates of maple leaves. We used the NDVI in April for cherry blossom and the NDVI in November for maple leaves. For the temperature variable, accumulated temperature and monthly mean temperature were selected because they are more related to the phenology. Table 1 shows the CC (Correlation Coefficient) between these variables and the plant phenology DOY. Cumulative temperatures had higher correlations with the phenology DOY than GDD or CDD for both cherry blossom and maple leaves. The monthly mean temperatures for cherry blossom had more significant correlations than cumulative temperatures, while the monthly mean temperatures of maple leaves had similar correlations to the cumulative temperatures.

Table 1. Correlation coefficients between phenology DOY and temperature variables

As to the monthly mean temperature, February to March temperature was used for cherry blossom, and September to October temperature was used for maple leaves. The equations for the estimation using the accumulated and the monthly mean temperature are as follows.

DOY_{CherryBlossom}= f(TEMP_FebMar, NDVI_Apr)       (1)

DOY_{CherryBlossom}= f(TEMP_Acc, NDVI_Apr)       (2)

DOY_MapleLeaves= f(TEMP_SepOct, NDVI_Nov)       (3)

DOY_MapleLeaves= f(TEMP_Acc, NDVI_Nov)       (4)

The accuracy of the MLR models for the flowering dates of cherry blossom and the peak dates for the maple leaves were presented in Table 2 and 3. For cherry blossom, Eq. 1 using the monthly mean temperature for February and March together with April NDVI showed much better accuracy than Eq. 2 using the accumulated temperature and April NDVI. The prediction error for the flowering date was about 3.3 days (MAE=3.331). As to maple leaves, Eq. 4 using the accumulated temperature and November NDVI showed a bit better accuracy than Eq. 3 using the monthly mean temperature for September and October along with November NDVI. The prediction error for the peak date was about 3.8 days (MAE=3.844).

Table 2. Accuracy statistics for the estimation of cherry blossom flowering dates by MLR models

Table 3. Accuracy statistics for the estimation of maple leaves peak dates by MLR models

Eq. 1 for cherry blossom and Eq. 4 for maple leaves were also applied to RF models to see if there is any difference between the results from the linear and nonlinear models. Eq. 1 for the flowering dates of cherry blossom had an MAE of 3.331 for MLR and an MAE of 3.335 for RF, and a CC of 0.868 for MLR and a CC of 0.870 for RF, which did not show a significant difference in the performance (Fig. 4 (a) and (b)). Eq. 4 for the peak dates of maple leaves had an MAE of 3.844 for MLR and an MAE of 3.842 for RF, and a CC=0.768 for MLR and CC=0.805 for RF, which also showed no significant difference. This indicates that the relationships between the plant phenology of cherry blossom and maple leaves and the environmental factors like temperature and NDVI seemed quite linear. However, considering the complex terrain and various tree types in the forests of South Korea, a more detailed examination using higher resolution data will be necessary.

Table 4. Accuracy statistics comparison between MLR andRFmodels fortheestimationofcherryblossom flowering dates

Table 5. Accuracy statistics comparison between MLR and RF models for the estimation of maple leaves peak dates

Fig. 4. Scatter plots of the (a) MLR for cherry blossom, (b) RF for cherry blossom, (c) MLR for maple leaves, and (d) RF for maple leaves, respectively.

The estimation results for the flowering dates of cherry blossom and the peak dates of maple leaves using the MLR models for Eq. 1 and Eq. 4, respectively, were illustrated in the form of maps in Fig. 5 to 8 and Fig. 9 to 12. The colors become dark when the phenology DOY is large and become light when the DOY is small. As closer to recent years, the flowering dates are getting faster, and the peak dates are getting slower, which were also made sure in the color distribution of the maps. The differences between estimated and observed DOYs were randomly slight, irrespectively of the years, proving that the statistical modeling was stable.

Fig. 5. Maps of the cherry blossom flowering dates: (a) MLR prediction, (b) observation, and (c) the difference, 2005- 2008.

Fig. 6. Maps of the cherry blossom flowering dates: (a) MLR prediction, (b) observation, and (c) the difference, 2009- 2012.

Fig. 7. Maps of the cherry blossom flowering dates: (a) MLR prediction, (b) observation, and (c) the difference, 2009- 2016.

Fig. 8. Maps of the cherry blossom flowering dates: (a) MLR prediction, (b) observation, and (c) the difference, 2017- 2020.

Fig. 9. Maps of the Maple leaves peak dates: (a) MLR prediction, (b) observation, and (c) the difference, 2005-2008.

Fig. 10. Maps of the Maple leaves peak dates: (a) MLR prediction, (b) observation, and (c) the difference, 2009-2012.

Fig. 11. Maps of the Maple leaves peak dates: (a) MLR prediction, (b) observation, and (c) the difference, 2013-2016.

Fig. 12. Maps of the Maple leaves peak dates: (a) MLR prediction, (b) observation, and (c) the difference, 2017-2020.

4. Conclusion

We examined a temporal trend of the flowering dates of cherry blossom and the peak dates of maple leaves in South Korea; we built statistical models and created maps to estimate the plant phenology using both temperature and NDVI. The more recent years, the faster the flowering dates and the slower the peak dates. This is because of the impacts of climate change with the increase of air temperature in South Korea. By reflecting the climate change, our statistical models could reasonably predict the plant phenology with the CC of 0.870 and the MAE of 3.3 days for the flowering dates of cherry blossom, and the CC of 0.805 and the MAE of 3.8 for the peak dates of maple leaves. We could suppose a linear relationship between the plant phenology DOY and the environmental factors like temperature and NDVI, which should be inspected in more detail. We found that the flowering date of cherry blossom was closely related to the monthly mean temperature of February and March, and the peak date of maple leaves was much associated with the accumulated temperature. A more sophisticated future work will be required to examine the plant phenology using higher-resolution satellite images and the additional meteorological variables sensitive to plant phenology. Diurnal temperature range, CDD/GDD, LST (Land Surface Temperature), and the difference between monthly NDVI may help improve the reliability of the model. Also, using meteorological grid can produce the spatially continuous raster maps for plant phenology.