We are pleased to welcome Nicolas Houdré and Hugo Riffaud de Turckheim (Université Paris Cité) for the next MILO seminar on Tuesday 10 February from 13:30 to 14:30 by videoconference.
Title: Conquering EO Heterogeneity: From Atomic Scalars to Resolution-Adjustable Foundation Models
Abstract: The rapid proliferation of Earth Observation (EO) satellites has created a data landscape of extreme heterogeneity, producing vast amounts of data with varying spatial, spectral, and temporal characteristics. Traditional deep learning architectures, however, are not built to address the challenges inherent to this diversity; they typically rely on fixed input formats that struggle to accommodate multiple spatial resolutions, diverse spectral configurations carrying distinct semantics, and variable image sizes without costly retraining or sensor-specific encoders. To address this, we present two architectures built to master this diversity. First, Atomizer adapts the data representation itself by decomposing images into sets of scalars enriched with metadata such as wavelength, bandwidth, and resolution enabling a single encoder to generalize to unseen sensor configurations without retraining. Second, RAMEN (Resolution-Adjustable Multimodal Encoder) establishes a unified foundation model that treats resolution as a controllable parameter, allowing users to dynamically trade off spatial precision and computational cost across diverse modalities such as optical, radar, and elevation data. Together, these methods demonstrate that explicitly encoding physical attributes paves the way for next-generation, modality-agnostic EO models capable of scaling to future satellite missions.
Inscription to the mailing list: https://groupes.renater.fr/sympa/info/milo-seminars
Looking forward to seeing you on Tuesday 10!
Pour participer à cet événement, l’inscription est obligatoire.
Partager :
