Overview
EN ISO 19178-1:2025 - "Geographic information - Training data markup language for artificial intelligence - Part 1: Conceptual model" defines a conceptual UML model for describing training data used in Earth Observation (EO) AI/ML workflows. Published by CEN/ISO TC 211, the standard focuses on maximizing interoperability and usability of EO imagery training data by standardizing metadata, task/label definitions and quality/provenance information.
Key topics and requirements
- UML conceptual model: Establishes a formal UML schema (TrainingDML-AI) to represent training datasets, dataset items and associated metadata to support consistent exchange and integration.
- AI/ML task taxonomy: Specifies EO AI tasks used in supervised learning, including scene-level, object-level and pixel-level tasks - enabling clear definition of labels and expected outputs for models.
- Metadata essentials: Requires description of dataset-level attributes such as persistent identifier, version, licence, training data size, and the measurement/imagery sources used for annotation.
- Labeling model: Defines classes and relations for labels and labeling activities (see sections like AI_Label and AI_Labeling) to capture label semantics and annotation provenance.
- Data quality & provenance: Specifies how to describe quality aspects (e.g., labeling errors, representativeness, quality measures) and provenance (agents performing labeling, labeling procedures, lineage).
- Modularity & extension: Provides principles for modularization and guidance for extending TrainingDML-AI to accommodate project-specific needs while preserving interoperability.
- Conformance & integrity: Addresses conformance rules and modelling principles including data integrity, authenticity and non-repudiation considerations.
Applications and users
This conceptual standard is relevant for organizations and professionals involved in EO AI/ML datasets:
- Data scientists and ML engineers preparing or consuming EO training data for model development.
- Remote sensing and GIS specialists organizing labeled imagery for land cover, object detection, segmentation and change detection.
- Satellite data providers, EO data platforms and data curators** packaging training datasets with standardized metadata and licences.
- Annotation teams and QA managers capturing provenance and quality metrics for labeling workflows.
- Standards bodies and platform vendors implementing interoperable APIs, catalogs and dataset repositories.
Practical benefits include improved dataset discoverability, reproducibility of ML experiments, easier dataset federation and clearer legal/licence traceability for model training.
Related standards
TrainingDML-AI references and interoperates with existing ISO geographic information standards, for example:
Keywords: Geographic information, training data markup language, EO imagery, AI/ML, supervised learning, UML model, data provenance, data quality, labels, training dataset interoperability.