ICEtormer is a new deep learning tool developed by CEOSpaceTech (Romania) and IFREMER (France), which incorporates a methodology to automatically segment sea ice from radar satellite images with unprecedented accuracy. It consists of a neural network based on transformer architecture – the same technology behind modern AI models – which has been specially adapted for Earth observation.

ICEtormer uses a module called Multi-Head Transposed Attention (MHTA), to focus attention across different channels in an image, enabling more precise identification of ice features. It includes innovative features like pixel shuffle/unshuffle operations that preserve spatial detail better than traditional downsampling methods.

The development of this model marks a step forward in enabling real-time, automated monitoring of Arctic conditions.

This paper has been presented at the 12th International Workshop on sea ice modelling, assimilation, observations, predictions and verifications at ESA-ESRIN, organised by the International Ice Charting Working Group.

It has been funded by ESA through the SARWAVE project and the European Union through the National Recovery and Resilience Plan, “PNRR-III-C9-2022 - I5 Establishment and operationalization of Competence Centers” competition, “Competence Center for Climate Change Digital Twin for Earth forecasts and societal redressment: DTEClimate” project, contract no. 760008/30.12.2022, code 7/16.11.2022.

More information

N.-C. Ristea, A. Anghel, A, Mouche, F. Nouguier, A. Grouazel, M. Datcu, “Transformer Neural Networks for Sea Ice Segmentation of Sentinel-1 GRD Images,” IICWG - DA - 12 Workshop 2024, ESA-ESRIN. Read the poster here.