Radiant Earth Foundation at AGU 2021
On December 13–17, Radiant Earth’s team will present their latest research and milestones at the American Geophysical Union (AGU) Fall Meeting.
Monday, 13 December 2021, 5:09 PM EST
Abba Barde, our Data Scientist, will present on Semi-Supervised Crop Type Classification from Multispectral Earth Observations Using Generative Adversarial Networks. He’ll address the limitations and provide corresponding suggestions about improving the existing crop classification models, which contain batches of labeled and unlabeled data. In addition, he will present the results of adding synthetic labeled data to improve the accuracy of classification models in regions with limited labeled training data. The results are based on a high-quality ground reference dataset collected in Western Cape, South Africa.
Radiant Earth investigated semi-supervised learning both in built-from-scratch and transferred approaches of limited labeled data. The method relied on using Generative Adversarial Networks (GAN) to generate pairs of synthetic imagery and labels augmented with authentic satellite imagery and labels from ground referencing campaigns and unlabeled imagery from regions with no labels. This approach benefits from the combined power of semi-supervised learning and GANs to improve classification accuracy.
Tuesday, 14 December 2021, 9:28 AM EST
Radiant’s Tech Lead and Geospatial Software Engineer Jon Duckworth will present Open Specifications for Discoverable and Reusable Machine Learning Workflows in Earth Sciences. Jon will discuss the current state of work to develop a specification for describing geospatial machine learning (ML) workflows from training to deployment and evaluation, including early implementations, and provide a roadmap for future development. In addition, a virtual poster will showcase the strategies for combining the specifications and tools in the SpatioTemporal Asset Catalog (STAC) ecosystem to catalog ML workflows for Earth Science.
Radiant Earth is developing the STAC ML Model Extension in collaboration with partners in the geospatial machine learning community to address the need for open standards for cataloging ML model metadata. This extension is being developed in conjunction with the STAC Label Extension and other STAC extensions and tools to provide a comprehensive solution for cataloging ML workflows end-to-end.
Thursday, 16 December 2021, 9:15 AM EST
Hamed Alemohammad, our Executive Director and Chief Data Scientist, will present on Open Machine Learning-Ready Data and Standards to Support Decision Making based on Earth Observations. He will review Radiant’s existing efforts around reproducible and reusable machine learning workflows, including Radiant MLHub training data repository and standards for cataloging machine learning data and models using STAC specification. In addition, Hamed will demonstrate the value of developing and using these open and community-based standards and tools to empower solutions for sustainable development.
Friday, 17 December 2021, 5:09 PM EST
Kevin Booth, our Engineering Lead and Geospatial Software Engineer, will present on Radiant MLHub: Best Practices for Geospatial Machine Learning Training Data Publication and Lessons Learned. Radiant MLHub is the world’s first cloud-based open library dedicated to geospatial training data with machine learning models. He will discuss the motivations and driving forces behind building Radiant MLHub, challenges we’ve encountered during the publication process, and lessons learned along the way.
Radiant Earth established Radiant MLHub to allow anyone to discover and access high-quality EO training datasets and machine learning models. In addition to creating geospatial training data, Radiant welcomes others to register or share training data on Radiant MLHub.
