AI for Medicanes Science

Applying advanced machine earning and deep learning techniques

Satellites gather vast amounts of environmental data daily, making remote sensing a key player in Big Data and Digital Twin initiatives. To efficiently analyze this data, experts are leveraging artificial intelligence (AI) tools.

In recent years, Artificial Intelligence has leapt forward thanks to three main drivers: larger models (Deep Learning Transformers), more data (including unlabeled archives), and more compute (GPUs and supercomputing). In vision, modern video models learn rich spatio-temporal representations: they capture rotation patterns, symmetries, and transitions in video data.

In meteorology and remote sensing, this translates into concrete use cases:
  • Storm and cyclone detection from satellite sequences (IR, microwave, visible) with supervised and self-supervised approaches.
  • Nowcasting/forecasting of high-impact phenomena (convective cells, squall lines, Medicanes) via models that predict cloud-field evolution.
  • Segmentation and classification of upper-level structures, fronts, and organized systems.
  • Index estimation (symmetry, compactness, warm-core proxies) to infer intensity and phase transitions (from extra-tropical to tropical-like).

    • The key is integration: self-supervised pre-training (to “digest” large unlabeled archives) followed by supervised fine-tuning on targeted tasks like detection, tracking, and intensity estimation.

    The Role of Machine Learning

    Machine learning algorithms are capable of learning from data, managing complexity, and integrating information from various sensors across different platforms. They utilize the full spectrum of available channels to solve complex, non-linear problems. However, ML implementation in remote sensing is relatively new and faces with some challenges:

    Data Availability: Data are the fuel

    In our case, they originate from IR and microwave satellite sensors in continuous observation; they must be collected, cleaned, recalibrated, and often reprocessed multiple times. It’s essential that the training set is representative of the data seen during operations (same channels, spatial coverage, cadence).

    When variability is high and volumes are “big,” human understanding combines with AI in an iterative interplay:
  • data profiling (statistics, outliers, temporal gaps),
  • unsupervised pre-training to extract general features,
  • targeted data curation (class balancing, definition of edge cases),
  • supervised fine-tuning with reliable labels,
  • deployment checks, and renewed refinement.

    • In short: AI doesn’t replace domain expertise—it amplifies it, accelerating cycles of “hypothesis → test → improve”, and thus increasing scientific output.

    Computational Resources

    Significant resources are required for big data processing and management: modern GPUs, HPC (High Performance Computing) clusters. These requirements are today very important for AI and especially for Deep Learning in Earth Science, in order to leverage proprerly big data available and big models needed for this amount of data.This infrastructure brings advanced video models closer to near-real-time operation, with tangible benefits for weather early warning.

    Applications of Machine Learning in Remote Sensing

    Project Focus: Geostationary IR

    In this project, we focus on IR bands from a geostationary satellite continuously observing the Mediterranean. The Airmass RGB composite (highlighting dry/moist air and stratospheric intrusions) is especially useful to capture upper-level dynamics tied to cyclogenesis. Example Airmass RGB composite during a cyclonic system.

    Detecting Medicanes

    Our first goal is to recognize the presence of a Mediterranean tropical-like system from short video sequences. The model learns to detect:
  • Cyclonic rotation patterns in mid-/high-level clouds,
  • Symmetry and compactness of the structure,
  • Hints of an eye when present or during its formation/development.

    • Tracking and Phase Transition

    Once detected, the next step is to localize the rotation center and track it over time.
    We obtain a trajectory of the center (tracking), and hopefully indicators of evolution from extra-tropical to tropical-like as an early-warning signal of cyclone intensification. A cyclone automatic tracking during its evolution phase.

    In a Nutshell

    AI for Medicanes Science combines geostationary IR, advanced video models, and a pipeline to move from observation to timely detection, tracking, and intensification warning. It's both scientific and engineering work: meticulous data care, robust modeling choices, and the compute backbone that makes AI useful in the field, especially during extreme events.

    Phase 2 — New Frontiers in AI-Based Intensity Estimation ★ New

    Phase 2 of the MEDICANES project significantly expands the AI toolkit, introducing new methodologies for objective intensity estimation and structural transition detection. These advances go beyond tracking: the aim is to determine how strong a medicane is, and how it is evolving thermodynamically.

    Passive microwave radiometry at 50–60 GHz and 183 GHz opens powerful new diagnostic capabilities. The 50–60 GHz oxygen absorption band characterises the depth, symmetry, and vertical coherence of the cyclone's warm core, while the 183 GHz water vapour absorption channels reveal upper-tropospheric humidity, dry intrusions, and moisture structures linked to potential vorticity anomalies. Together, they enable a process-oriented diagnostic of warm-core formation and tropical transition — something that visible and infrared imagery cannot achieve alone.

    Building on these diagnostics, Phase 2 integrates and adapts state-of-the-art deep-learning frameworks:

    • D-PRINT CNN — a convolutional neural network for short-term intensity change prediction from infrared imagery, adapted for the hybrid structures of medicanes
    • D-MINT183 — a CNN framework using 183 GHz microwave observations to estimate current intensity, suited to the shallow and rapidly evolving thermodynamic structure of medicanes
    • SATCON-adapted consensus algorithms — combining infrared, microwave, and sounding-based estimates into optimised values of maximum sustained winds (Vmax) and minimum sea-level pressure (MSLP) for the Mediterranean basin

    Given the limited number of medicane events, training strategies leverage transfer learning from global tropical cyclone datasets, synthetic data augmentation, and physics-informed constraints. This work is developed in collaboration with the Tropical Cyclone research group at CIMSS, University of Wisconsin-Madison, and with support from the ESA Earth System Science Hub on hybrid physics-data-driven approaches and multi-sensor data fusion.

    Synergy with Digital Twins: The synthetic medicane datasets generated by the Digital Twin work will provide physically consistent training examples spanning complete medicane lifecycles — including storm phases currently absent from the observational record — directly enriching the AI training pipeline.