Overshooting cloud tops can cause severe weather conditions, such as aviation turbulence, lightning, strong winds, heavy rainfall, hail, and tornadoes. Due to hazards caused by overshooting tops, several methods have been developed to detect them. Convolutional neural networks are an approach to machine learning that performs well on image-based tasks such as object detection. This study uses convolutional neural networks to detect overshooting tops in GOES-14 satellite imagery from NOAA’s Geostationary Operational Environmental Satellite.

Visible and infrared images of GOES-14 satellite imagery are the primary source of input data. These images are divided into patches of size 31 x 31. The model takes in each patch and outputs its classification for that patch, whether it contains an overshooting top or not. The distribution of patches containing overshooting tops and those do not contain overshooting tops in the data is imbalanced. In this study, we first implement data sampling and cost-sensitive learning techniques to deal with imbalanced data. Next, we investigate approaches using tropopause temperature data to increase the performance of the model. The method we propose using tropopause temperature data in the preprocessing step performs better overall than other methods with a probability of detection 79.31%, false alarm ratio 90.94%, and critical success index 0.088. Most of the false alarms are located close to overshooting occurrence, and using a preprocessing step decreases the testing time significantly. This thesis compares the performance of different imbalanced learning techniques on satellite images.