Korean Journal of Remote Sensing (대한원격탐사학회지)

Korean Society of Remote Sensing (대한원격탐사학회)
권호 표지
DOI QR코드

DOI QR Code

Improvement of Thunderstorm Detection Method Using GK2A/AMI, RADAR, Lightning, and Numerical Model Data

  • Yu, Ha-Yeong (Department of Atmospheric Science, Kongju National University) ;
  • Suh, Myoung-Seok (Department of Atmospheric Science, Kongju National University) ;
  • Ryu, Seoung-Oh (Department of Atmospheric Science, Kongju National University)
  • Received : 2021.02.08
  • Accepted : 2021.02.17
  • Published : 2021.02.26
Download PDF
⟨ Previous Next ⟩

Abstract

To detect thunderstorms occurring in Korea, National Meteorological Satellite Center (NMSC) also introduced the rapid-development thunderstorm (RDT) algorithm developed by EUMETSAT. At NMCS, the H-RDT (HR) based on the Himawari-8 satellite and the K-RDT (KR) which combines the GK2A convection initiation output with the RDT were developed. In this study, we optimized the KR (KU) to improve the detection level of thunderstorms occurring in Korea. For this, we used all available data, such as GK2A/AMI, RADAR, lightning, and numerical model data from the recent two years (2019-2020). The machine learning of logistic regression and stepwise variable selection was used to optimize the KU algorithms. For considering the developing stages and duration time of thunderstorms, and data availability of GK2A/AMI, a total of 72 types of detection algorithms were developed. The level of detection of the KR, HR, and KU was evaluated qualitatively and quantitatively using lightning and RADAR data. Visual inspection using the lightning and RADAR data showed that all three algorithms detect thunderstorms that occurred in Korea well. However, the level of detection differs according to the lightning frequency and day/night, and the higher the frequency of lightning, the higher the detection level is. And the level of detection is generally higher at night than day. The quantitative verification of KU using lightning (RADAR) data showed that POD and FAR are 0.70 (0.34) and 0.57 (0.04), respectively. The verification results showed that the detection level of KU is slightly better than that of KR and HR.

Keywords

1. Introduction

Thunderstorm refers to a well-developed deep convection cloud and accompanied by strong rain, gusts, lightning, and thunder (Lutgens et al., 1995). In Korea, three sides are surrounded by the sea, and high mountain ranges are located in the north-south direction in the eastern region of the inland, and small mountain ranges derived from this are formed east-west, so the meteorological phenomena vary greatly depending on the region (Eom and Suh, 2009). In particular, localized torrential rain, heavy snow, and convective precipitation are affected by such topographical influences, making forecasting difficult. Also, in recent years, the number of days of extreme rainfall, the intensity of rainfall, and the intensity of lightning have increased due to the effect of global warming, leading to an increase in economic and social damage (Price and Rind, 1994; Kim and Suh, 2009; Eom and Suh, 2010). Therefore, the development of methodologies for detecting and predicting such thunderstorms and cloud cells with the possibility of a thunderstorm is an urgent task. Meteorological scientists are making a lot of effort to understand and predict the mechanism of occurrence of thunderstorms, such as feature analysis using satellite, RADAR, model development using the supercomputer, data assimilations, and ensemble techniques (Eom and Suh, 2010).

As the detection and prediction of thunderstorms become more important, various studies have been conducted to develop and improve techniques for detecting thunderstorms. According to the previous studies of lightning, there is a correlation between the frequency of lightning and rainfall intensity (Zipser, 1994; Sheridan et al., 1997; Petrova et al., 2009). An analysis of the relationship between lightning and thunderstorms in Florida shows that the maximum frequency of lightning strikes precedes the maximum rainfall (Piepgrass et al., 1982). Also, studies on the prediction of rainfall using lightning data suggest that lightning data can be used to predict severe weather (Tapia et al., 1998; Grecu et al., 2000).

As quantitative lightning observations began in Korea, various characteristic analysis studies were conducted. Eom and Suh (2009) statistically analyzed lightning in South Korea from 2002-2006 and presented the characteristics of lightning in South Korea. Oh et al. (2010) analyzed the relationship between lightning and rainfall in South Korea during the summer of 2006-2007. Also, the values of stability indices for lightning and studies on the predictability of lightning using these were conducted (Eom and Suh, 2010). In a study using satellite data, Lee (2010) showed the possibility of thunderstorm detection using satellite images (infrared (IR) and water vapor channels) by analyzing the characteristics of satellite brightness temperatures on the top of the cloud with lightning. However, since most of these previous studies have focused on case analysis, it is difficult to be considered as general characteristics of thunderstorms. This is attributed to the lack of lightning and rainfall observation data in Korea. Therefore, a longer period of data is needed for the systematic analysis of characteristics using data.

Thunderstorms often occur in environments where the atmosphere is very humid and thermodynamically unstable (Eom and Suh, 2009). Therefore, when thunderstorms develop, the altitude (temperature) of the cloud top rises (decreases) rapidly. (Lee, 2010; Mecikalski et al., 2010; Seong and Suh, 2014; Han et al., 2015). However, it is difficult to detect thunderstorms only with satellite data equipped with optical sensors because the characteristics of the cloud top of a thunderstorm are very similar in the mature stage where lightning is frequent and the extinction stage where lightning rarely occurs. Besides, early detection of thunderstorms is important, but since the cloud top temperature is not low and the rainfall intensity is not strong at the beginning of the occurrence, it is not easy to detect even with a RADAR.

To monitor and predict the lifetime of thunderstorms, Météo-France developed a thunderstorm detection technology (Rapid Development Thunderstorm: RDT) using geostationary satellite data (Météo-France, 2013a). RDT provides various information on thunderstorm characteristics, spatial continuity, and thunderstorms that the forecaster is interested in through an object-oriented approach. Several countries have introduced RDT technology and are conducting studies to optimize RDT for their country (Météo-France, 2013a, 2013b; Gijben et al., 2017; Müller et al., 2019; Lee et al., 2020). Recently, the Korea Meteorological Administration has also introduced RDT technology to the National Meteorological Satellite Center (NMSC) to improve the level of thunderstorm detection. Lee et al. (2020) introduced CI (convective initiation) outputs of GK2A/ AMI (GEO-KOMPSAT-2A/Advanced Meteorological Imager) to improve the level of thunderstorm detection in RDT and adjusted some variables to suit Korea’s climate environment (K-RDT). However, as GK2A data are provided from July 2019, optimization has not been successfully made for thunderstorms and GK2A/AMI in Korea, and as a result, there is a problem of low accuracy.

Therefore, this study aims to improve the level of thunderstorm detection of the K-RDT currently being operated by NMSC based on the GK2A/AMI and other ancillary data. Section 2 introduces the input data including GK2A data, a schematic K-RDT algorithm, and its improvement process. And section 3 presents the qualitative and quantitative validation of thunderstorm detection results. Discussion and summary are presented in Section 4.

2. Data and Method

1) Data

In this study, the GK2A/AMI data for about 80 cases of thunderstorms and non-thunderstorms over the Korean peninsula were used. Additionally, temperature, relative humidity, and atmospheric instability indexes (K Index: KI, Showalter Stability Index: SSI, Lifted Index: LI, etc.) from the Unified Model (UM) N1280, along with lightning and RADAR data provided by the Korea Meteorological Administration (KMA) were used. The GK2A/AMI, which has been in service since July 2019, has 16 channels and observes every 10 minutes (NMSC, 2019). Of the 16 channels, the spatial resolution of the infrared channels is 2 km, and the resolution of the visible channel (0.64 μm) is 500 m. The channel number, center wavelength, resolution for each channel, and the channels used in this study are presentedin Table 1. The spatial resolution and temporal frequencyof the UM N1280 are 10 km and 3 h, respectively. Considering the differences in the spatial and temporal resolution of the numerical model data and the GK2A/AMI data, bi-linear interpolation and selection of the nearest numerical model data were used.

Table 1. Specification of GK2A/AMI data used in this study

OGCSBN_2021_v37n1_41_t0001.png 이미지

Lightning and RADAR data provided by the KMA were used as input data of K-RDT and to quantitatively validate the detection level of the thunderstorm detection algorithm improved in this study. Currently, the KMA observes lightning at 21 observatories in South Korea and provides information on the time, location, polarity, and intensity of the lightning. The observation area is over 95% of the land and adjacent seas over the Korean Peninsula. Since lightning data is point data, temporal and spatial colocation is needed to use it with satellite data. In this study, as shown in Fig. 1, temporal and spatial colocation between lightning and satellite detection cloud cells was performed under the assumption that lightning strikes may occur within a maximum radius of 35 km and between -15 min. and +5 min. based on the cloud cell detection time (Météo-France, 2013b; Météo-France, 2018). The RADAR data used in this study is Hybrid Surface Rainfall (HSR) provided by KMA. The RADAR data has a horizontal resolution of 500 m and a time resolution of 5 min. To use RADAR with GK2A/AMI data, the spatial resolution difference was corrected by a weighted average method that is inversely proportional to the distance (Fig. 2). The total number of input variables derived from GK2A/AMI, numerical models and RADAR is 313.

OGCSBN_2021_v37n1_41_f0001.png 이미지

Fig. 1. The method of temporal and spatial collocation of cloud cell and lightning data.

OGCSBN_2021_v37n1_41_f0002.png 이미지

Fig. 2. The method of spatial collocation of GK2A/AMI data and RADAR/HSR data.

In South Korea, the characteristics of thunderstorms vary according to the season, stage of development, geographic location, and time of day (Eom and Suh, 2009; Myeong and Suh, 2010). And machine learning, which is recently used in dangerous weather detection, is highly dependent on the quality and quantity of input data. Therefore, to accurately detect thunderstorms by applying machine learning methods to satellite data, various thunderstorm cases are needed. However, in this study, about 80 cases of thunderstorms and non- thunderstorm over the Korean peninsula from July 2019 to July 2020 were selected due to the limitation of the study period and the limitations of GK2A/AMI data. So, the total number of data sets used for training was 11520 sets (80 days × 24 h/day × 6 sets/h).

2) Method

(1) The K-RDT_KNU Algorithm

Fig. 3 is a flow chart briefly showing the thunderstorm detection algorithm (RDT) improved in this study. RDT is a very complex system that comprehensively performs thunderstorm-capable cloud detection, thunderstorm development stage setting, duration calculation, thunderstorm status determination, and future thunderstorm location prediction. In this study, we focus on expanding the input data including GK2A/AMI data and RADAR data and subdividing the thunderstorm classification from 36 classes to 72 classes to optimize the RDT for the Korean climate environment. GK2A/AMI, numerical model, RADAR, and lightning data are used as input data, and the lifetime of thunderstorms is monitored through four major steps. In the first step, the cloud cell is detected using the brightness temperature of the 10.4 μm IR channel, and the second step tracks the cloud cell by analyzing the overlap of the cloud cell and monitors the lifetime of the thunderstorms. In step 3, it is determined whether the cloud cell is a thunderstorm or not, and finally, in step 4, the direction of thunderstorm movement is predicted. In particular, in step 3 of discrimination, the logistic regression equations, which were previously retrieved empirically through machine learning, is used.

OGCSBN_2021_v37n1_41_f0003.png 이미지

Fig. 3. Flow chart of K-RDT_KNU for the detection of thunderstorms.

(2) Optimizing Regression Eqs. Using Machine Learning Techniques

The RDT algorithm was developed to detect and track thunderstorms in European environments. Therefore, the discrimination regression coefficients of RDT used to distinguish thunderstorms may not be suitable for Korea (Lee et al., 2020). In this study, various statistical techniques and machine learning techniques were used to derive the logistic regression equations suitable for the climate environment of Korea. Fig. 4 shows the flow chart of calculating the regression formula for the discrimination step of K-RDT_KNU, which is largely composed of four steps. The first step is to prepare input data for thunderstorm detection. The second step is to conduct the quality control of each data and collocate among different data sets in space and time (preprocessing step). Then, the cloud cell is classified as a thunderstorm or not based on the lightning occurrence. Additionally, for each cloud cell classification, thunderstorm and non- thunderstorm cloud cells were sampled 1:1. In step 3, variables with strong multicollinearity among the 313 input variables are excluded using the variance inflation factor (VIF) method and t-test, and statistically independent variables are selected. The stepwise method is used to derive a logistic regression equation for determining whether the cloud cell is a thunderstorm or not from the previously selected independent variables. Finally, the optimal discrimination regression equations for the 72 cloud cell types were derived using training and test data.

OGCSBN_2021_v37n1_41_f0004.png 이미지

Fig. 4. Flow chart for the selection of optimized logistic regression equation according to the 72 thunderstorm classifications in K-RDT_KNU.

Fig. 5 is a schematic diagram of how cloud cells are classified. Cloud cells are divided into 6 categories (development stage: 0-5) based on the maximum temperature of cloud tops, and the cloud cells are subdivided into 6 depths (sustained time: 0-5) based on the duration time. We also add a day/night separation, considering that the different availability of satellite channels and the different characteristics of thunderstorm activity between day and night. Therefore, regression equations for thunderstorm detection were developed for a total of 72 categories (6 × 6 × 2). In the following, category, depth, and day/night (Solar Zenith Angle: SZA) are aggregated and denoted as CaDeS.

OGCSBN_2021_v37n1_41_f0005.png 이미지

Fig. 5. The method of cloud cell classification based on the cloud top temperature (Category) of cloud cells, duration time (Depth) of cloud cells, and time (SZA) of day.

Since the general regression model assumes independence among variables, it is necessary to check whether there is multicollinearity among variables, and if so, resolve them. In general, if the VIF value is 5 or higher, multicollinearity exists, and if the value is 10 or higher, multicollinearity can be regarded as serious. Therefore, the multicollinearity among the discrimination parameters was partially resolved by repeating removal until the discrimination parameter with a VIF value of 10 or more is removed. Additionally, through the t-test, variables that did not show statistically significant differences (p_value ≥ 0.05) according to the presence or absence of lightning were removed.

Machine learning is performed for each cloud cell classification to find an optimized discrimination equation. To prevent overfitting in the process of deriving the regression equation for each cloud type, all data were divided into training (80%) and evaluation (20%). Machine learning of stepwise variable selection and logistic regression was performed using the training cases, and the results were evaluated by applying the learned results to the test cases. At this time, the probability threshold used for logistic regression was optimized through a sensitivity test with a 0.05 interval from 0 to 1. Finally, optimal discrimination parameters, regression coefficients, and probability thresholds for each cloud cell classification (72) were selected through Receiver Operating Characteristics (ROC) analysis of Threat Score (TS) and False Alarm Ratio (FAR). In ROC analysis, discrimination parameters, regression coefficients, and probability thresholds with large TS and small FAR were finally selected.

Quantitative assessment of thunderstorm detection level of K-RDT_KNU was performed based on the contingency table as shown in Table 2 (Wilks, 2011; Han et al., 2019). The formula for the evaluation elements using the contingency table is as shown in Eqs. (1)-(5). The higher the value of the probability of detection (POD), TS and equitable threat score (ETS) is, the lower the value of FAR is, the better the detection level. Quantitative evaluation of thunderstorm detection using lightning data is performed based on the lighting frequency (Table 3) as shown in Météo-France (2013b). Also, a quantitative evaluation of thunderstorm detection using RADAR data was performed based on the maximum RADAR reflectivity (Table 4). In general, a RADAR reflectivity of less than 30 dBZ indicates weak precipitation, 30 dBZ-40 dBZ is moderate precipitation, 40 dBZ-50 dBZ is strong precipitation, and above 50 dBZ is defined as a high probability of hail (Morné et al., 2017). The RADAR reflectivity of 35 dBZ to 45 dBZ is mainly used as a threshold defining thunderstorms (Météo-France, 2013b; Morné et al., 2017).

\(\mathrm{POD}=\frac{\mathrm{GD}}{(\mathrm{GD}+\mathrm{MI})}\)       (1)

\(\mathrm{FAR}=\frac{\mathrm{FA}}{(\mathrm{GD}+\mathrm{FA})}\)       (2)

\(\mathrm{TS}=\frac{\mathrm{GD}}{(\mathrm{GD}+\mathrm{MI}+\mathrm{FA})}\)       (3)

\(\mathrm{ETS}=\frac{(\mathrm{GD}-\mathrm{Href})}{(\mathrm{GD}-\mathrm{Href}+\mathrm{MI}+\mathrm{FA})}\)       (4)

\(\mathrm{Href}=\frac{(\mathrm{GD}+\mathrm{MI})(\mathrm{GD}+\mathrm{FA})}{(\mathrm{GD}+\mathrm{MI}+\mathrm{FA}+\mathrm{CN})}\)       (5)

Table 2. The 2 × 2 contingency table

OGCSBN_2021_v37n1_41_t0002.png 이미지

Table 3. Classification of the frequency of thunderstorm activity using lightning data

OGCSBN_2021_v37n1_41_t0003.png 이미지

Table 4. Classification of the intensity of thunderstorm activity using RADAR data

OGCSBN_2021_v37n1_41_t0004.png 이미지

3. Results and discussion

1) Verification results

To compare the level of thunderstorm detection of K-RDT (KR), Himawari-8 Rapid Development Thunderstorm (H-RDT: HR), and K-RDT_KNU (KU), we conducted qualitative and quantitative verification using stability indexes, RADAR, and lightning data. As verification cases, a total of 4 cases (November 10, 2020, May 18, 2020, July 9, 2020, July 23, 2020) were selected based on the frequency of lightning. Although the criteria of day and night are different depending on the season and geographic location, in this study, day and night were classified by applying the same conditions (day: 00:00 ~ 08:50 UTC; night: 09:00 ~ 23:50 UTC) to all cases.

Fig. 6 shows the spatial distribution of (a) surface weather map at 00:00 UTC, (b) daily precipitation, stability indexes of KLAPS Forecast System (KLFS) at 05:00 UTC ((c) KI, (d) SSI, (e) LI), (f) Lightning and (g) RADAR/HSR for the daytime case of May 18, 2020. In this case, the temperature difference between the upper and lower layers increased due to the cyclone in the west sea and the strong cyclonic vortex located in the south of Mongolia, resulting in gusts and thunderstorms in thunderstorms in Seoul, Gyeonggi, and Gangwon provinces. A large amount of precipitation occurred in the north-central region of South Korea due to the influence of low pressure approaching from the west sea (b). And KI has more than 40 in Seoul, Gyeonggi, and Gangwon provinces, and the probability of thunderstorms is more than 90%. Also, SSI and LI show the possibility of strong thunderstorms (<-3) and general thunderstorms (<-3), respectively (Jeon, 2003). Lightning strikes frequently in front of cold fronts accompanied by low pressure in the west sea, and strong precipitation is also seen in RADAR images.

OGCSBN_2021_v37n1_41_f0006.png 이미지

Fig. 6. Spatial distribution of (a) surface weather map at 00:00 UTC, surface, (b) daily precipitation (mm/day), stability indexes of KLFS at 05:00 UTC, (c) KI, (d) SSI, (e) LI, (f) Lightning, and (g) RADAR/HSR on May 18, 2020.

Fig. 7 shows the detection results by three thunderstorm detection algorithms (KR, HR, KU) at 05:00 UTC on May 18, 2020, along with satellite infrared images. All of the three thunderstorm detection algorithms are good at detecting thunderstorm that occur strongly in the west sea. Overall, the detection was similar to that of KR and KU, while HR is relatively under-detecting. In particular, there is a large difference in the detection level of the three algorithms in the region from the central part of the Korean Peninsula to the East Sea.

OGCSBN_2021_v37n1_41_f0007.png 이미지

Fig. 7. Spatial distribution of thunderstorm detection results (a) KR, (b) HR, and (c) KU at 05:00 UTC (daytime) on May 18, 2020. The colors and green lines indicate the development stage of the thunderstorm and the trajectory of detected lightning cells, respectively.

Fig. 8 shows the spatial distribution of various meteorological variables for the night case of Oct. 11, 2019, as shown in Fig. 6. In this case, a large amount of precipitation occurred in the western region of South Korea due to the influence of low pressure approaching 40 km/h from Eastern China. At this time, the KI over the western region is 30 or more, showing a probability of a thunderstorm of more than 60%. And both SSI and LI are less than 0, so there is a possibility of a thunderstorm (Jeon, 2003). Since the instability of the atmospheric environment was not strong, the lightning was not frequent and strong.

OGCSBN_2021_v37n1_41_f0008.png 이미지

Fig. 8. Same as Fig. 6 except for 11:00 UTC (nighttime) on Oct. 10, 2019.

Fig. 9 shows the thunderstorm detection results for the night case of Oct. 11, 2019, as shown in Fig. 7. In this case, all three algorithms are well-detecting lightning strikes from north to south in the western region of South Korea. In this case, KR is detecting more thunderstorms than HR and KU in Busan, the northern part of the Korean peninsula, and the east coast.

OGCSBN_2021_v37n1_41_f0009.png 이미지

Fig. 9. Same as Fig. 7 except for 11:00 UTC (nighttime) on Oct. 10, 2019.

Table 5 shows the quantitative verification results of KR, HR, and KU by time (total/day/night) and frequency of thunderstorm activity using lightning data. From various evaluation measures such as POD, TS, and ETS, indicated that the thunderstorm detection level of KU is superior to that of KR and similar to that of HR, although it differs depending on the intensity of the thunderstorm and day/night. Overall, in KR (HR), POD is low (medium) and FAR is high (low), whereas KU has high POD and low FAR. Also, the thunderstorm detection level of the three detection algorithms is proportional to the lightning frequency, and the higher the lightning frequency, the better the detection. In particular, in all three detection methods, when the frequency of lightning strikes is low, the detection level is very low, so improvement methods for consistent detection regardless of lightning frequency are needed. Although there are differences depend on the frequency of lightning and detection methods, the level of thunderstorm detection is generally higher at night than day.

Table 5. Summary of validation results of KR, HR, and KU using lightning data according to the lighting frequency. The best detection level in each evaluation element in the table is indicated in bold

OGCSBN_2021_v37n1_41_t0005.png 이미지

Table 6 shows the quantitative verification results of KR, HR, and KU by time (all/day/night) and frequency of thunderstorm activity using RADAR data. All three algorithms show differences in detection levels according to the intensity of RADAR reflectivity, as in the quantitative verification results using lightning data, and the higher the intensity of RADAR reflectivity, the higher the detection level. Also, both POD and FAR are lowered than in the verification using lightning data regardless of the day and night, the frequency of lightning strikes, especially in the FAR. When the RADAR reflectivity is low, the low POD seems to be related to that lightning rarely occurs when the rainfall intensity is weak. Although differences are depending on the intensity of RADAR reflectivity, day and night, overall, KU outperforms the other two algorithms.

Table 6. Summary of validation results of KR, HR, and KU using RADAR data according to the intensity of reflectivity. The best detection level in each evaluation element in the table is indicated in bold

OGCSBN_2021_v37n1_41_t0006.png 이미지

2) Discussion

The level of thunderstorm detection using lightning and RADAR data is affected by the time interval and search radius. Table 7 shows the thunderstorm detection levels according to the time interval and search radius in the verification process of thunderstorm detection using lightning data. In this study, the search time was set from -15 min. to +5 min. based on the thunderstorm detection time, and the search radius was set to 35 km as in the Météo-France (2013b). So, the sensitivity tests were conducted at intervals of 5 min. and 5 km. Overall, the POD is inversely proportional to the time interval and search radius, but FAR is the opposite. However, interestingly, TS and ETS are not sensitive to time intervals and search radius. Since the thunderstorm detection level is sensitive to the time interval and search radius, it is necessary to optimize the verification criteria by utilizing more cases.

Table 7. Sensitivity test result for search time and radius in verification using lightning data

OGCSBN_2021_v37n1_41_t0007.png 이미지

4. Conclusions

In this study, RDT, the thunderstorm detection method of EUMETSAT, was optimized using 80 days (11520 sets) of GK2A/AMI data, RADAR data, lightning data, and numerical model data, for detecting thunderstorms occurring in the Korean peninsula. The detection levels of three algorithms (KR, HR, and KU) were verified using lightning and RADAR data for the selected four cases of thunderstorms that occurred in 2019-2020. To select the parameters that contribute to the thunderstorm detection regression equations, the machine learning method of logistic regression and stepwise selection was used. In this process, considering the developing stages and duration time of thunderstorms, and data availability of GK2A/AMI, a total of 72 types of detection algorithms (category: 6 × depth: 6 × day/night: 2 = 72) were developed (KU).

Visual inspection using the verification cases showed that all three thunderstorm detection algorithms detect thunderstorms that occurred on the Korean Peninsula relatively well. The level of detection differs according to the frequency of thunderstorm activity, and the higher the frequency of thunderstorm activity, the higher the detection level is. And the level of thunderstorm detection is generally higher at night than during the day, although differences are depending on the frequency of lightning and detection methods. The quantitative verification using lightning data showed that POD and FAR of KU (KR, HR) are 0.70 (0.53, 0.63) and 0.57 (0.78, 0.55), respectively. And the quantitative verification using RADAR data showed that TS and ETS of KU (KR, HR) are is 0.34 (0.31, 0.19) and 0.18 (0.12, 0.07), respectively. It can be seen that the thunderstorm detection level of KU is partially improved compared to the detection level of KR and HR currently operated by NMSC.

In this study, the cases used to calculate the discrimination regression equation was about 80 days (11520 sets), therefore, the average number of training and test data sets for 72 types is 128 and 32, respectively. Due to the insufficient number of data used and the influence of the characteristics of thunderstorm occurrence, there was a problem that training and testing were not performed in some types of thunderstorms. So a stable regression equation could not be derived in some cloud cell types due to the lack of sample numbers

Therefore, if machine learning is performed using data for at least more than two years in the future, it is expected that the level of thunderstorm detection will be significantly improved. To improve the level of thunderstorm detection using GK2A and auxiliary data, it is necessary to improve the detection level, especially in low lightning frequency, and the difference in detection level during the day and night. Also, since the verification results are different according to the temporal and spatial collocation method, it is necessary to optimize the verification method through characteristic analysis of verification data and thunderstorms using more diverse verification cases.

Acknowledgements

This research was supported by “The Technical Development on Weather Forecast Support and Fusion Service using Meteorological Satellites” of the NMSC/KMA and KMI2020-01414 project of the Korea Meteorological Industry and Technology Institute.

References

  1. Eom, H.S. and M.S. Suh, 2009. Statistical characteristics of recent lightning occurred over South Korea, Journal of the Korean Earth Science Society, 30(2): 210-222 (in Korean with English abstract). https://doi.org/10.5467/JKESS.2009.30.2.210
  2. Eom, H.S. and M.S. Suh, 2010. Analysis of Stability Indexes for Lightning by Using Upper Air Observation Dataover South Korea, Atmosphere, 20(4): 467-482 (in Korean with English abstract).
  3. Gijben, M. and E. De Coning, 2017. Using satellite and lightning data to track rapidly developing thunderstorms in data sparse regions, Atmosphere, 8(4): 67. https://doi.org/10.3390/atmos8040067
  4. Grecu, M., E.N. Anagnostou, and R.F. Adler, 2000. Assessment of the use of lightning information in satellite infrared rainfall estimation, Journal of Hydrometeorology, 1(3): 211-221. https://doi.org/10.1175/1525-7541(2000)001<0211:AOTUOL>2.0.CO;2
  5. Han, H., S. Lee, J. Im, M. Kim, M. -I. Lee, M.-H. Ahn, and S.-R. Chung, 2015. Detection of convective initiation using Meteorological Imager onboard Communication, Ocean, and Meteorological satellite based on machine learning approaches, Remote Sensing, 7(7): 9184-9204. https://doi.org/10.3390/rs70709184
  6. Han, J.H., M.S. Suh, and S.H. Kim, 2019. Development of day fog detection algorithm based on the optical and textural characteristics using Himawari-8 data, Korean Journal of Remote Sensing, 35(1): 117-136 (in Korean with English abstract). https://doi.org/10.7780/KJRS.2019.35.1.8
  7. Jeon, B.I., 2003.Characteristics of Atmospheric Stability Index of Airmass thunderstorm day at Busan, Journal of Wetlands Research, 5(1): 29-40 (in Korean with English abstract).
  8. Kim, C.S. and M.S. Suh, 2009. Change-point and Change Pattern of Precipitation Characteristics using Bayesian Method over South Korea from 1954 to 2007, Atmosphere, 19(2): 199-211 (in Korean with English abstract).
  9. Lee, J. G., K.H. Min, H. Park, Y. Kim, C.Y. Chung, and E.C. Chang, 2020. Improvement of the Rapid-Development Thunderstorm (RDT) Algorithm for Use with the GK2A Satellite, Asia-Pacific Journal of Atmospheric Sciences, 56(2): 307-319. https://doi.org/10.1007/s13143-020-00182-6
  10. Lee, Y.J., 2010. Characteristics of Rainfall and Brightness Temperature of Satellite on Lightning Events during Summer, Kongju National University, Gongju, KOR(in Korean with English abstract).
  11. Lutgens, F.K., E.J. Tarbuck, and D. Tusa, 1995. The atmosphere (Vol. 462), NJ: Prentice-Hall, Englewood Cliffs, USA, pp. 279-311.
  12. Mecikalski, J.R., W.M. MacKenzie Jr, M. Koenig, and S. Muller, 2010. Cloud-top properties of growing cumulus prior to convective initiation as measured by Meteosat Second Generation. Part I: Infrared fields, Journal of Applied Meteorology and Climatology, 49(3): 521-534. https://doi.org/10.1175/2009JAMC2344.1
  13. Meteo-France, 2013a. Algorithm Theoretical Basis Document for "Rapid Development Thunderstorms" (RDT-PGE11v3.0), https://www.nwcsaf.org/AemetWebContents/ScientificDocumentation/Documentation/MSG/SAFNWC-CDOP2-MFT-SCI-ATBD-11_v3.0.pdf, Accessed on Feb. 17, 2021.
  14. Meteo-France, 2013b. Validation Report for "Rapid Development Thunderstorms" (RDTPGE11 v3.0), https://www.nwcsaf.org/AemetWebContents/ScientificDocumentation/Documentation/MSG/SAF-NWC-CDOP2-MFT-SCI-VR11_v3.0.pdf, Accessed on Feb. 17, 2021.
  15. Meteo-France, 2018.Validation report of the Convection Product Processors of the NWC/GEO, https://www.nwcsaf.org/Downloads/GEO/2018/Documents/Scientific_Docs/NWC-CDOP3-GEOMF-PI-SCI-VR-Convection_v1.0.pdf, Accessedon Feb. 17, 2021.
  16. Muller, R., S. Haussler, M. Jerg, and D. Heizenreder, 2019. A Novel Approach for the Detection of Developing Thunderstorm Cells, Remote Sensing, 11(4): 443. https://doi.org/10.3390/rs11040443
  17. Myoung, J.S. and M.S. Suh, 2010. Characteristics by the Cloud-to-Ground Lightning Polarity Occurred over South Korea, Journal of the Korean Earth Science Society, 31(4): 335-347 (in Korean with English abstract). https://doi.org/10.5467/JKESS.2010.31.4.335
  18. NMSC(National Meteorological Satellite Center), 2019. GK2A Level 1B data user manual, https://nmsc.kma.go.kr/homepage/html/base/cmm/selectPage.do?page=static.satllite.introGk2a, Accessed on Feb. 17, 2021.
  19. Oh, S.G., M.S. Suh, and Y.J. Lee, 2010. On the Relation Between Cloud-to-Ground Lightning and Rainfall During 2006 and 2007 Summer Cases, Journal of the Korean Earth Science Society, 31(7): 749-761 (in Korean with English abstract). https://doi.org/10.5467/JKESS.2010.31.7.749
  20. Petrova, S., R. Mitzeva, V. Kotroni, J. Latham, and E. Peneva, 2009. Analysis of summer lightning activity and precipitation in the Central and Eastern Mediterranean, Atmospheric Research, 91(2): 453-458. https://doi.org/10.1016/j.atmosres.2008.06.013
  21. Piepgrass, M.V., E.P. Krider, and C.B. Moore, 1982. Lightning and surface rainfall during Florida thunderstorms, Journal of Geophysical Research, 87(C13): 11193-11201. https://doi.org/10.1029/JC087iC13p11193
  22. Price, C. and D. Rind, 1994. Possible implications of global climate change on global lightning distributions and frequencies, Journal of Geophysical Research, 99(D5): 10823-10831. https://doi.org/10.1029/94JD00019
  23. Seong, M.G. and M.S. Suh, 2014. Characteristics of Infrared and Water Vapor Imagery for the Heavy Rainfall Occurred in the Korean Peninsula, Remote Sensing, 30(4): 465-480 (in Korean with English abstract).
  24. Sheridan, S.C., J.F. Griffiths, and R.E. Orville, 1997. Warm season cloud-to-ground lightning-precipitation relationship in the South-Central United States, Weather and Forecast, 12(3): 449-458. https://doi.org/10.1175/1520-0434(1997)012<0449:WSCTGL>2.0.CO;2
  25. Tapia, A., J.A. Smith, and M. Dixon, 1998. Estimation of convective rainfall from lightning observations, Journal of Applied Meteorology and Climatology, 37(11): 1497-1509. https://doi.org/10.1175/1520-0450(1998)037<1497:EOCRFL>2.0.CO;2
  26. Wilks, D.S., 2011. Forecast Verification, Statistical Methods in the Atmospheric Sciences: 3th Edition, Academic Press, Waltham, USA, pp. 308-318.
  27. Zipser, E.J., 1994. Deep cumulonimbus cloud systems in the Tropics with and without lightning, Monthly Weather Review, 122(8): 1837-1851. https://doi.org/10.1175/1520-0493(1994)122<1837:DCCSIT>2.0.CO;2