The Classification of Tropical Storm Systems in Infrared Geostationary Weather Satellite Images Using Transfer Learning

Abstract

The work performed in this study evaluated the application of generalized pretrained object detection models for the identification and classification of tropical storm (TS) systems through transfer learning. While the majority of literature focuses on developing bespoke models for this application, these typically require significantly more training data, compute resources, and time to train the models due to the large number of parameters the model has to tune to achieve similar results. These models also required additional preprocessing steps, such as extracting the storm from the image, and used a limited number of classes to describe the intensity of the storms. The approach presented here used considerably less data than the majority of other work (2–10x fewer input images) and a larger number of classes. The accuracies of the produced models trained on four different experimental datasets (varying the amount of data and number of classes) through this approach were 75%, 82%, 69%, and 89%. Overall, the models produced promising results, performing approximately equal to the bespoke models with scope to improve the performance of the model.

Keywords

• Machine learning (ML)
• remote sensing
• transfer learning
• tropical storm (TS)