Creation of new Extra-Tropical Cyclone fields in the North Atlantic with Generative Adversarial Networks: A deep learning framework to generate new synthetic atmospheric variables fields from the learned original sample data distribution

• Main Author: Dainelli, Filippo (author)
• Other Authors: Taormina, R. (mentor), Bricker, J.D. (graduation committee), Glassmeier, F. (graduation committee), Caires, Sofia (graduation committee), Delft University of Technology (degree granting institution)
• Format: Master Thesis
• Language: English
• Published: 2020
• Subjects: Extra Tropical Cyclones; Generative Adversarial Network; Deep Learning; North Atlantic
• Online Access: http://resolver.tudelft.nl/uuid:baa63b40-8318-4d46-bf72-0372d9ef13db

Extra-Tropical Cyclones (ETCs) are major storm system ruling and influencing the atmospheric structure at mid-latitudes. These events are usually characterized by strong winds and heavy precipitation and cause considerable storm surges with threatening wave systems for coastal regions. The possibility to simulate these storms or to increase the amount of significant data available is crucial to optimize risk assessment and risk management for construction projects and territorial plans which might get damaged by events of this kind. The project addresses the possibility to learn the distribution of cyclones atmospheric fields of pressure, wind and precipitation in the North Atlantic by training a Generative Adversarial Network (GAN). The ETCs tracks are extracted from the ERA5 reanalysis dataset in the domain with boundaries 0°-90°N, 70°W-20°E and period going from 1st January 1979 to 1st January 2020. A GAN tries to learn the distribution of a training set based on a game theoretic scenario where two network competes against each other, the generator and the discriminator. The former is trained to generate new examples which are plausible and similar to the real ones by having as input a vector of random Gaussian values. The random vectors domain is called latent space. The latter learns to distinguish whether an example is coming from the dataset distribution or not. The competition set by the game scenario makes the network improve until the counterfeits are indistinguishable form the original. The generative models trained on the ETCs dataset are validated to understand if they are able to generate new samples of fields presenting similar atmospheric characteristics to those of the original dataset. To train the GAN two different loss function are considered, the Wasserstein distance and the Cramèr distance. The Cramèr Gan (CGAN) shows better performance in representing the distribution of the atmospheric fields, generating images that on average look similar to the original ones. The Wasserstein GAN (WGAN) ...