Detection and Tracking of Medicanes Through DeMeTrA Self-Supervised Vision Transformer

Medicanes are mesoscale cyclones that develop over the Mediterranean Sea and display tropical-like cyclone characteristics, including a warm core, spiral cloud organization, and deep convection over warm sea surfaces. Since their structure and position can change rapidly on short lead times before coastal impact, robust near-real-time tracking algorithms are essential for timely warning and operational decision support. To advance this research direction, this work introduces the Deep Learning Medicane Tracking (DeMeTrA) Algorithm, an end-to-end deep learning framework for medicane detection and rotation-center localization from SEVIRI Rapid Scan Airmass RGB imagery. The proposed methodology consists of a three-stage VideoMAE v2 architecture encompassing the following: (i) self-supervised domain specialization on unlabeled satellite image sequences, (ii) supervised binary classification of cyclone versus non-cyclone events, and (iii) supervised coordinate regression for rotation-center tracking. The training corpus spans several time windows of Meteosat Second-Generation observations across the Mediterranean basin, with ground-truth annotations derived from a consensus cyclone-track reference. On event-based splits, cyclone detection reaches 91% balanced accuracy on a balanced validation set and 89% on an unbalanced test set representative of operational conditions. The tracking results show generally low localization errors (mostly below 20 km), with limited outliers in the most complex cases. These findings support the use of Transformer-based video models for operational medicane monitoring and establish a baseline for future developments.