AIDE: Trusted Deep-Learning Super-Resolution for MTG FCI within Destination Earth

Meteorological imagery occupies a special role in Earth observation: its high temporal frequency and broad spectral coverage are indispensable for weather forecasting and climate modelling, while its spatial resolution remains limited. Technological advances, however, are driving a trend toward higher spatial detail and open new application domains beyond traditional meteorology.

The Flexible Combined Imager (FCI) aboard Meteosat-12 represents the latest generation of geostationary weather sensors and images the entire Earth disk at 10-minute intervals. Its 16 spectral channels span the visible to longwave infrared and offer native spatial resolutions ranging from 2000 m down to 500 m — a configuration well suited to super-resolution techniques.

The AIDE project develops a methodology to increase the spatial resolution of all FCI channels to up to 500 m while preserving the radiometric integrity of the data. The approach employs a purpose-built deep-learning model that is augmented with an estimate of its own predictive uncertainty, thereby enabling safe downstream use of the enhanced products in demanding applications and quantitative analyses. The method is implemented as a demonstrator within the Destination Earth (DestinE) Data Lake, demonstrating that appropriately designed machine-learning approaches can be deployed reliably in critical operational contexts.

This contribution presents the concept and development of the method, summarizes promising project results, and discusses the limitations and potential of super-resolution approaches in Earth observation.