EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Hugging Face Blog·Aug 12·5 min read
Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis
TL;DR
OlmoEarth Studio now enables users to compute and export custom embedding vectors from Earth observation data, enhancing downstream analysis capabilities. These embeddings facilitate tasks like similarity search and segmentation, with strong performance metrics reported.
✦ Why It Matters
Start using OlmoEarth Studio today to compute custom embeddings for your Earth observation projects.
Key Takeaways
- Compute custom embeddings for specific areas and time ranges.
- Use embeddings for similarity search and land-cover mapping.
- Export lightweight COGs for easy sharing and analysis.
- Leverage few labeled pixels for effective classification tasks.
Full Summary
OlmoEarth Studio has introduced a feature that allows users to compute and export embedding vectors, which are compact numerical representations of Earth observation data. These embeddings can be generated from various parameters, including area of interest, time span, and imagery sources like Sentinel-2.
The embeddings have demonstrated strong performance in tasks such as similarity search and land-cover mapping, with a logistic regression classifier achieving a weighted F1 score of 0.84 using only 60 labeled pixels. Users can generate embeddings on demand, reflecting specific conditions without relying on pre-computed data.
The exported Cloud-Optimized GeoTIFFs (COGs) are lightweight and compatible with various geospatial tools, making them easy to share and analyze. This capability allows for detailed exploration of seasonal dynamics and landscape structures without extensive training data.