Thunderstorms with lightning, torrential rain, hail, and gust are one of the most difficult meteorological phenomena to detect and predict, as they are not only diverse in scale but also highly variable in space-time. Recently, researches are being conducted to detect rapidly developing thunderstorms using geostationary orbit satellites with improved observation performance in terms of channel number, spatial resolution and frequency, such as MSG, Himawari-8, and GK2A and using numerical weather prediction model output. The National Meteorological Satellite Center(NMSC) has also introduced the detection algorithm of rapidly developing thunderstorm (RDT) developed by EUMETSAT, to detect thunderstorms occurring in Korea and is testing the GK2A algorithm(K-RDT) and the Himawari algorithm(H-RDT). In this study, optimization of the K-RDT was performed to improve the detect level of thunderstorms occurring in South Korea. To this goal, GK2A/AMI data, that was launched in December 2018, RADAR data, lightning data, and numerical model data were used. And the thunderstorm cases that occurred in the summer of 2019 and 2020 were used in this study. Information on thunderstorm morphology and physical characteristics was collected from satellite data, atmospheric stability index data from numerical model output and precipitation intensity data from RADAR data. Lightning data was used as a dependent variable for machine learning, and lightning and radar data were used to validate the thunderstorm detection algorithm. For considering the developing stages and duration time of thunderstorms, and data usability of GK2A/AMI, a total of 72 types of detection algorithms (category: 6 (0-5) x depth: 6 (0-5) x day and night: 2 (0, 1) = 72) were developed(KU). For selecting discrimination parameters that do not have multicollinearity through various statistical techniques and have a large influence on the occurrence of thunderstorms, we used various statistical techniques and machine learning method of logistic regression and stepwise variable selection. In the machine learning process, about 80% of the convective and non-convective cloud cells (about 22,000) including thunderstorms were used as training data, and the optimal regression model was selected from the remaining cloud cells. To validate the performance of the KU, qualitative and quantitative validations were performed with the two other thunderstorm detection algorithms (K-RDT, H-RDT) currently operated in the NMSC. As a result of qualitative and quantitative varidation of the three thunderstorm detection algorithms using lightning and radar data, it was found that the performance of the thunderstorm detection algorithm (KU) improved in this study was superior regardless of the frequency of thunderstorms and rainfall intensity. For a higher level of thunderstorm detection, it is necessary to add training cases used for machine learning, and studies on the time continuity of day and night. In addition, the validation method should be refined in consideration of the thunderstorm and the spatiotemporal characteristics of the varidation data.