Aircraft icing is a hazardous phenomenon which has potential to cause fatalities and socioeconomic losses. It is caused by super-cooled droplets (SCDs) colliding on the surface of aircraft frame when an aircraft flies through SCD rich clouds. When icing occurs, the aerodynamic balance of the aircraft is disturbed, resulting in a potential problem in aircraft operation. Thus, identification of potential icing clouds is crucial for aviation. Satellite remote sensing data such as Geostationary Operational Environmental Satellite (GOES) series have been widely used to detect potential icing clouds. An icing detection algorithm, operationally used in the US, consists of several thresholds of cloud optical depth, effective radius, and liquid water path based on the physical properties of icing. On the other hand, there is no operational icing detection algorithm in Asia, although there are several geostationary meteorological satellite sensors. In this study, we proposed machine learning-based models to detect icing over East Asia focusing on the Korean Peninsula using two geostationary satellite sensors—Meteorological Imager (MI) onboard Communication, Ocean and Meteorological Satellite (COMS) and Advanced Himawari Imager (AHI) onboard Himawari-8. While COMS MI provides data at 5 channels, Himawari-8 AHI has advanced capability of data collection, providing data at 16 channels. Instead of simple thresholding approaches used in the literature, we adopted two machine learning algorithms—decision trees (DT) and random forest (RF) to develop icing detection models based on Pilot REPorts (PIREPs) as reference data. Results show that the COMS icing detection model by RF produced a detection rate of 88.67% and a false alarm rate of 14.42%, which were improved when compared with the result of the direct application of the GOES algorithm to the COMS MI data (a detection rate of 20.83% and a false alarm rate of 25.44%). Although much higher accuracy (a detection rate > 95%) was achieved when Himawari AHI data were used, the model was not robust due to the very limited number of training data. Incorporation of MODIS-derived icing reference data may improve the reliability of the machine learning models for Himawari AHI data.