DeMeTrA: A Two-Stage Coarse-to-Fine Deep Learning Framework for Medicane Detection and Tracking from MSG SEVIRI Image Sequences

Medicanes are rare Mediterranean tropical-like cyclones characterized by small spatial scales, rapid evolution, and in particular a warm core, a cloud free eye, and a closed ring of strong winds, leading to potentially severe coastal impacts, which require robust near-real-time detection and tracking from high-frequency geostationary satellite imagery. Their automatic detection and tracking remain challenging because labelled events are scarce, cyclone morphology is highly variable, and satellite-based signatures may be confused with other organized cloud systems.

This work presents DeMeTrA, a deep-learning framework for medicane detection and rotation center localization exploting MSG SEVIRI Rapid Scan Service (RSS) Airmass RGB image sequences. The system combines self-supervised spatiotemporal representation learning based on VideoMAE pretrained vision transformer model, with supervised downstream modules for cyclone presence estimation and center tracking. The architecture was developed through a two-scale detection and tracking strategy motivated by subsequent analyses of basin-scale inference. High-resolution video Transformer backbones provide powerful spatiotemporal representations, but their use in this context requires the Mediterranean domain to be processed through fixed-size small local crops compatible with the pretrained input geometry. This local formulation may limit the availability of full-domain spatial context during the initial detection stage. Under realistic Mediterranean conditions, where organized cloud systems, frontal structures, and peripheral spiral-like patterns can resemble medicane signatures, independent local decisions may lead to false alarms or to multiple candidate centers that are not physically consistent at basin scale. To address this limitation, DeMeTrA integrates a lightweight first-pass module operating on the full Mediterranean basin at reduced spatial resolution. This module estimates cyclone presence and provides a coarse cyclone-center location, which is then used to guide the high resolution VideoMAE tracking stage over a physically consistent region of interest. By separating basin-scale event screening from local center refinement, the framework preserves the advantages of pretrained VideoMAE representations while restoring the large-scale contextual information required for robust detection. This two-scale design limits false-alarm generation, avoids inconsistent candidate centers from independent spatial crops, and provides a single coherent center-track estimate for each detected event, supporting near-real-time medicane monitoring from geostationary infrared satellite observations.

Applications of DeMeTra to documented medicane cases will be shown to analysie its capabilities and skills during the storm development and mature phases. The methodology is currently being extended to the MTG Flixible Combined Imager FCI) for future applications.

Keywords: medicanes; MSG SEVIRI; deep learning; vision transformer; cyclone detection; cyclone tracking