The number and variety of machine learning (ML) models that utilize geospatial data (e.g. satellite imagery, airborne observations, and physical model estimates) is growing rapidly. Many of these ML models have the potential to be deployed to derive useful insights for many global problems, or become benchmark baselines that empower development of more accurate and complex models. For this to happen, these models should be made more discoverable and usable by ML practitioners and data scientists. The GMLMC specification aims to address this goal through a common metadata definition for ML models that operate on geospatial data.
At a high level, this specification should provide sufficient information to enable search and discovery of geospatial ML models and answer the following questions:
- Is this model applicable to my domain (e.g. land cover, agricultural monitoring, etc.)?
- What kind of input data are required to use the model?
- Is this model applicable to the geographic region I am interested in?
- How well does this model perform under the kinds of conditions under which I will be using it?
The best place to start is with the Model Metadata spec. This describes the top-level metadata document for a geospatial ML model. The Model Metadata spec describes the fields that may be present in the top-level metadata document with links more detailed descriptions and definitions of more complex fields.
The SpatioTemporal Asset Catalog (STAC) spec is a mature and well-defined standard for cataloging geospatial assets. STAC has mature specifications for describing remotely sensed data, as well as labeled data related to machine learning applications. The Electro-Optical, Projection, View, and SAR extensions, among others, provide a thorough description of remotely sensed data from a variety of sources. Combined with the Label extension, these provide a thorough description of the data required for training an ML model using remotely sensed imagery.
The Geospatial ML Model Catalog (GMLMC) specification takes advantage of the STAC specification to describe training data associated with a model, and we hope that discussions relating to the GMLMC spec will help guide development of those standards as well.
This specification is a work in progress and should be considered unstable until further notice. The specification is not currently versioned, but we plan to begin versioning once we have a basic working draft.
See the Contributing guidelines for more information on how you can contribute to the spec.