Date of Award


Document Type


Degree Name

Master of Science (MS)


Ocean Engineering and Marine Sciences

First Advisor

Steven Lazarus

Second Advisor

Michael E. Splitt

Third Advisor

Pallav Ray

Fourth Advisor

Eraldo Ribeiro


Machine learning is a rapidly expanding technology that has proven to be highly useful for image classification. Ground-based camera networks are an emerging resource for aviation weather information with near real-time imagery available online for public viewing and download. While raw web camera imagery can be analyzed by aviators, high pilot workload motivates the use of machine learning to autonomously interpret cloud type information from images that is relevant to aviation weather hazards. In particular, transfer learning is a machine learning approach by which elements of a pre-trained machine learning model are refitted for new tasks. By employing transfer learning for image classification, the outermost layers of complex convolutional neural networks (CNNs) can be quickly retrained for new classes of imagery as organized by the user. In this research, a tiered methodology is employed whereby machine transfer learning is used to develop cloud image classification schemes of increasing complexity. To achieve accurate results using transfer learning, a large and diverse training dataset is required. During May of 2022, four publicly accessible ground-based web cameras were installed at FIT (Florida Institute of Technology) Aviation – located at the Melbourne International Airport (KMLB). These data are used to develop an extensive cloud image archive. Image categories of interest include a variety of cloud types and sky conditions, prioritizing those relevant to aviation safety hazards such as towering cumulus, cumulonimbus, and precipitation. Utilizing Google’s TensorFlow machine learning platform in Python, transfer learning was conducted with their Inception v3 convolutional neural network for deep learning. Several iterations of models were developed and tested for accuracy to assess the impacts of training data organization, application to different camera sites, and time of day. As proof of concept, the models were used to classify the FIT Aviation web camera imagery in real time to supplement weather information for potential users while simultaneously allowing for the near real-time tracking of model performance. Results reveal that cloud-type image classification using transfer learning is a viable method for extracting high-temporal-resolution information from growing web camera resources while minimizing the human component of weather information processing. Classification output demonstrates that the model correctly identifies hazard-related image content for a large percentage of cases, especially when raw model output is adapted to optimize model hit rates while minimizing false alarms and missed events.


Copyright held by author

Included in

Meteorology Commons