Advancing Earth Observation with Radiant Earth: The ACCESS-19 Project

Principal Investigator: Kevin Booth (Radiant Earth)

In the ever-evolving field of Earth observation, Radiant Earth is making significant strides through innovative initiatives, prominently highlighted in the ACCESS-19 project. Spearheaded by Principal Investigator Kevin Booth, this project has made groundbreaking contributions aimed at enhancing access to Earth observation data and machine learning (ML) models. At its core, ACCESS-19 revolves around three major enhancements that promise to revolutionize how researchers and developers utilize geospatial data.

Project Overview

The ACCESS-19 project aims to democratize access to training data and machine learning models that harness Earth observation capabilities. This is crucial for a world increasingly reliant on environmental data to tackle challenges like climate change, urban planning, and biodiversity conservation. Radiant Earth’s focus on three pivotal objectives ensures that this initiative not only delivers high-quality data but also enhances usability and accessibility.

Project Objectives

1. Produce a Multi-Mission Global Land Cover Training Dataset: This aims to provide users with robust datasets that are crucial for training machine learning models in land cover classification.

2. Develop an Open API for Registering and Retrieving ML Models: Making ML models easy to access and implement is essential for fostering innovation in the field.

3. Develop a Python Client to Increase Usability of Radiant MLHub: This tool simplifies the process of interacting with the repository, allowing users to easily retrieve datasets and models.

Exciting Updates

The project took off in late 2019 with the development of Radiant Earth’s open-access repository, designed specifically for Earth observation training data and machine learning models. A cornerstone of this achievement is LandCoverNet, a multi-mission global land cover training dataset that consists of an impressive 8,941 image chips, each measuring 256 x 256 pixels. This wealth of data is derived from 300 geographically diverse tiles of Sentinel-2 imagery and spans multiple continents, covering Africa, Asia, Australia, Oceania, Europe, North America, and South America.

But it doesn’t stop at just a dataset. Alongside these labeled land cover chips, users can access a yearly time series of corresponding Sentinel-1, Sentinel-2, and Landsat-8 imagery, enriching the resources available for researchers. Since its publication in 2020, LandCoverNet has quickly gained traction and become one of Radiant Earth’s most downloaded datasets, showcasing its significance in practical applications.

Expanding the Radiant MLHub

Another landmark achievement of the ACCESS-19 project is the expansion of Radiant MLHub. Initially designed for hosting training datasets, it now facilitates the publishing and retrieving of machine learning models. These improvements are neatly organized through the STAC (SpatioTemporal Asset Catalog) API, allowing users to access model documentation, code, and weights through a web interface.

To streamline the experience further, a new Python client has been developed. This not only allows users to programmatically search for and access ML training datasets but also simplifies the overall interaction with Radiant MLHub. With this client, even those who may not have extensive technical backgrounds can engage with the platform effortlessly.

Looking Forward: Source Cooperative

As Radiant Earth continues innovating, the team has embarked on the development of the next generation of Radiant MLHub, named Source Cooperative. This upcoming platform aims to better support extremely large training datasets and expand the capabilities for publishing datasets beyond just machine learning. Currently in a private beta, the public beta is anticipated to launch in Q3 of 2023, promising exciting new features and enhanced functionalities.

Major Accomplishments

The ACCESS-19 initiative has led to several noteworthy milestones:

• Production of a Multi-Mission Global Land Cover Training Dataset: A substantial contribution to the available resources for research and model training.

• Expansion of Radiant MLHub: Now includes capabilities for machine learning model publishing, further enhancing its utility.

• Development of a Python Client: Streamlines API interactions, making it accessible to a broader audience.

Further Resources and Research Publications

Researchers interested in the datasets produced can find valuable publications documenting the various facets of the LandCoverNet project:

• LandCoverNet Africa: A Geographically Diverse Land Cover Classification Training Dataset. DOI: 10.34911/rdnt.d2ce8i

• LandCoverNet Asia: DOI: 10.34911/rdnt.63fxe5

• LandCoverNet Australia: DOI: 10.34911/rdnt.0vgi25

• LandCoverNet Europe: DOI: 10.34911/rdnt.7s12zu

• LandCoverNet North America: DOI: 10.34911/rdnt.jx15e8

• LandCoverNet South America: DOI: 10.34911/rdnt.6a27yv

For those looking to dive deeper into the available resources, the Radiant ML Hub serves as an essential gateway.

Radiant Earth’s ACCESS-19 project exemplifies a significant commitment to enhancing the utilities of Earth observation data and machine learning, paving the way for innovative research and applications in a rapidly changing world.