OlmoEarth v1.3: embeddings for the whole planet (huggingface.co)

OlmoEarth has released v1.3 of its satellite embedding model, which includes precomputed, annual embeddings for the entire planet, available in 10-meter resolution and 64- or 128-dimensional formats. This update is accompanied by the OlmoEarth Embeddings Explorer, a user-friendly, browser-based tool that allows users to visualize and interact with the embeddings without the need for technical expertise or GPU access. This new model streamlines the process from fine-tuning to quick analysis, enabling applications such as monitoring environmental changes, analyzing land cover, and understanding agricultural practices across various regions. The significance of OlmoEarth v1.3 lies in its enhanced efficiency and accessibility for the AI/ML community, boasting a 53% reduction in training time and a 50% decrease in carbon emissions compared to the previous version. The model's performance has been validated against competitors across multiple evaluation tasks, demonstrating superior generalization capabilities, especially in unprecedented locations. With open access to model weights, data, and code, users can easily utilize these embeddings to train classifiers for practical applications like tracking illegal mining in Peru or assessing the impact of land-use changes, significantly reducing the barriers to leveraging advanced satellite data.