Radiant MLHub is a common repository of geospatial training data and trained ML models to serve as a resource for a community of practice around benchmark training datasets and models to advance applications of ML techniques on Earth Observations.
In the first phase, Radiant MLHub will host open source training datasets generated by Radiant Earth Foundation's team as well as other training data catalogs contributed by Radiant Earth’s partners. All of the training datasets will be stored using SpatioTemporal Asset Catalog (STAC) standard, and exposed through a common API. In the second phase, model hosting and Python client applications will be added to MLHub services.
Training datasets include pairs of imagery and labels for different types of ML problems including image classification, object detection, and semantic segmentation. Labels will be generated from ground reference data or annotation of imagery.
The Radiant Earth team is working on two training datasets that will be published in the Summer and Fall 2019. The first dataset consists of crop types in Africa. This dataset is built from ground reference data contributed by Radiant Earth partners, processed and cataloged by Radiant Earth. Each of the ground reference data points is matched with corresponding Sentinel-2 scenes (across the whole growing season) that will be published in an ML-ready training data catalog.
The second dataset contains image chips for global land cover classification based on Sentinel-2 surface reflectance observations (L2A data). These chips are selected to be representative of global land cover classes, and the labels are generated using a hybrid approach of ML prediction and crowdsourcing.
Join Radian MLHub community on Slack to be part of this collaborative open source group. If you are looking for new training data or models, or are interested in collaborarting with us to develop new ones, get in touch!
Subscribe to Radiant Earth Foundation's newsletter to get the latest updates for workshops and hackathons around machine learning and Earth observations.