Abstract
Despite advances in weather forecasting technology, poor visibility on the open ocean remains a significant safety hazard. Rapidly changing maritime weather conditions can cause visibility to plummet within minutes, leading to potential accidents and shipping delays. In recent years, deep neural networks (DNNs) have demonstrated strong performance in land-based visibility prediction; however, such approaches typically rely on large historical datasets and provide forecasts for fixed geographic areas. Oceanic visibility forecasting is more challenging due to the absence of stationary weather stations, constantly changing vessel positions, and the highly dynamic nature of marine environments. To address these challenges, this thesis introduces self-adaptive deep learning (SADL) models for real-time marine visibility forecasting using time-series data collected from onboard sensors and drone-based atmospheric measurements. The proposed models incorporate a continuous online learning mechanism that updates model weights in real time, enabling adaptation to both short-term weather variations and long-term environmental trends. Real-time training is performed using small data batches, ensuring computational efficiency while maintaining high data relevance. The SADL models are built upon the Long Short-Term Memory (LSTM) architecture, which is well-suited for capturing the temporal dependencies in visibility forecasting. The models are evaluated using high-fidelity simulated data derived from real-world datasets provided by NOAA and Copernicus. Case studies, including a realistic simulated storm scenario, demonstrate that the proposed approach achieves high accuracy and robust performance under diverse and extreme conditions. These results highlight the potential of integrating self-adaptive deep learning with real-time sensor data to enhance maritime safety through accurate visibility forecasting.