Klassificering och Rörelseanalys av Mesoskaliga Molnmönster i Satellitbilder med Maskininlärning

Clouds represent one of the largest sources of uncertainty in climate modeling because of their heterogeneous spatial structures and dynamic behavior. This thesis investigates the potential of artificial intelligence (AI) to classify and track mesoscale low-level cloud patterns, particularly those observed in the trade wind regions of the Atlantic Ocean. Focusing on cloud structures such as Sugar, Gravel, Fish, and Flowers, the thesis explores whether deep learning approaches can provide reliable classification and tracking using satellite imagery. A dataset provided by Swedish Meteorological and Hydrological Institute (SMHI), comprised of SEVIRI Level 1.5 RGB satellite images from 2021 was used. The research applies a range of deep learning techniques including a custom convolutional neural network (CNN), a pre-trained ResNet50 fine-tuned on EuroSAT and the self-supervised Vision Transformer-based model DINOv2. DINOv2 embeddings were used in two separate approaches, one based on annotated labels and the other using K-means clustering, to examine whether classification accuracy could be improved under conditions of limited labeled data. Results demonstrate that the DINOv2 model, when combined with a multilayer perceptron (MLP) classifier and trained on annotated labels, outperformed other configurations in cloud pattern classification. Although the clustering-based approach using DINOv2 features did not reach the same level of accuracy, it still showed a promising alternative in scenarios with limited labeled data. RAFT-based motion tracking was able to visualize cloud movement across satellite images, offering exploration of temporal cloud behavior. This thesis contributes to the EU-funded AI4PEX project and illustrates how AI-based methods can assist meteorologists and climate scientists in cloud monitoring tasks. The proposed methods may contribute to better cloud representation in climate models and support more accurate long-term climate projections.