Διόρθωση προγνώσεων σημαντικού ύψους κύματος στη Μεσόγειο θάλασσα με χρήση μοντέλων U-Net και δορυφορικών μετρήσεων
View/
Keywords
U-NetAbstract
Marine weather forecasting, particularly significant wave height prediction, is a critical issue for shipping, fisheries, and coastal protection. The problem is characterized by complex, dynamic systems described by nonlinear differential equations, making traditional numerical weather prediction computationally demanding and sometimes insufficiently accurate. This thesis aims to improve the accuracy of significant wave height forecasts in the Mediterranean Sea by leveraging artificial intelligence techniques. Specifically, convolutional neural networks of the U-Net type and the XGBoost machine learning algorithm are employed, trained on numerical model forecasts, satellite observations, and reanalysis data. The U-Net models utilize spatially organized multichannel inputs, combining forecasts and observations on geographical grids, while XGBoost provides an efficient baseline using local features without spatial context. Model training was performed using advanced computational resources such as Nvidia A100 GPUs via the Google Colab Pro+ platform. Data includes historical forecasts from Copernicus Marine Service, satellite wave height measurements, and reanalysis data serving as ground truth. The methodology involves spatial correction of numerical forecasts through learning from real observations, with separate models trained for each forecast lead time. Key findings demonstrate that U-Net models significantly outperform XGBoost and the initial forecasts across all accuracy metrics, especially for longer lead times. This approach highlights the strong potential of deep neural networks to enhance accuracy and reliability in marine weather forecasting, with practical applications in route optimization and maritime safety.
Postgraduate Studies Programme
Τεχνητή Νοημοσύνη - Artificial IntelligenceDepartment
Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Ψηφιακών ΣυστημάτωνCorporate Department
Εθνικό Κέντρο Έρευνας Φυσικών Επιστημών «Δημόκριτος»Number of pages
114Language
GreekCollections
Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα