Towards a Representation Framework for EO Deep Learning Models in SDI

ABSTRACT The rapid adoption of deep learning (DL) in Earth Observation (EO) has raised concerns about model sharing and reuse. Existing DL models (DLM) in EO are not well described in a unified and interoperable approach. To address this issue, a comprehensive and interoperable descriptive framework is needed to operationalize the FAIR (Findable, Accessible, Interoperable, Reusable) principles for EO DLM. The paper proposes a formal representation for EO DLM. Key considerations for formalizing the model are analyzed, including model lifecycle, model provenance, model inference, model constraints, and model quality. The conceptual and content models are then proposed, followed by an implementation schema. Use cases demonstrate the applicability of the approach for sharing and reusing DLM in Spatial Data Infrastructures (SDIs). The results help transform the traditional SDI into an AI‐ready SDI.