Ai2 Releases OlmoEarth v1.3: A New Satellite Map Embedding Model
Ai2, the Allen Institute for Artificial Intelligence, has announced the launch of OlmoEarth v1.3, a satellite map embedding model along with precomputed annual embeddings for the entire planet at various resolutions. The release includes a browser-based map-making tool designed to simplify access to satellite data for a broader range of users.
The OlmoEarth v1.3 model is engineered to compress a full year of satellite observations into a numerical representation, or fingerprint, of each 10-by-10-meter patch of land. Users can also access data at larger resolutions of 64-square meters and 128-square meters. This numerical fingerprint captures patterns that can be used to differentiate various types of land cover, including crops, forests, and mines.
Previously, utilizing OlmoEarth required users to fine-tune in OlmoEarth Studio, which involved training a specific model for each task using extensive GPU resources, alongside the need to download and store massive amounts of imagery. The introduction of the OlmoEarth Embeddings Explorer allows users to bypass complex setup processes and turn data into actionable insights without requiring sign-in or extensive resources.
The concept of embeddings represents aggregated learning outcomes derived from millions of satellite images. OlmoEarth processes data from several satellite missions, including Sentinel-1, Sentinel-2, and Landsat, which provide complementary views of Earth. Specifically, Sentinel-1 measures radar reflections, while Sentinel-2 and Landsat monitor the visible and infrared light spectrum.
According to Ai2, version 1.3 is not only the most cost-effective OlmoEarth model to date but also significantly reduces computational demands. It requires only 53% of the computing time compared to the previous version (v1.2) and just 18% of the computing requirement of the original version (v1). Carbon emissions have also decreased, with reductions of 50% and 22% for the respective previous versions. Despite these optimizations, the model’s strength in fine-tuning tasks means v1.2’s performance remains highly beneficial for research and partner labs.
Applications of OlmoEarth Models
The practical applications of these satellite models span various fields, aiding in data compression across time and geographical space. This capability proves particularly useful for monitoring environmental changes and human activities. For instance, illegal gold mining in Peru’s Madre de Dios region was responsible for the deforestation of over 1,200 hectares around La Pampa during its peak in 2017 and 2018. Following an operational crackdown in 2019, embeddings derived from OlmoEarth allowed researchers to visually analyze and confirm these changes over time.
The embeddings facilitate straightforward comparisons of satellite pixels across years without requiring extensive model training. For example, analyzing changes from 2016 to 2020 indicated a significant eastward expansion in mining activities, amounting to approximately 34 square kilometers of change, which then dropped to just 9 square kilometers from 2020 to 2024 post-intervention.
Moreover, agricultural analysts can leverage this technology to monitor the expansion of oil palm plantations and assess the extent of forest cover loss. Such assessments play a critical role in understanding the environmental impacts of agricultural practices globally. Ai2 emphasizes that for many users, accessing and testing ideas against OlmoEarth has been a resource-intensive process that now takes seconds in the Exploratory phase, easing the experimental burden significantly.
By eliminating the need for high-performance GPUs to analyze a region based on a limited set of labels, the OlmoEarth v1.3 model improves efficiency. For tasks necessitating fine-tuning, the embeddings can assist in determining where to allocate computing resources effectively.