Embedding workflows for Earth Observation tasks

Abstract

Efficient data handling is essential for managing and unlocking the growing volume of Earth Observation (EO) archives. Recent advances in machine learning enable neural embeddings, that is: compact, meaningful representations to distill information into small vectors for downstream tasks and near real-time applications. This workshop demonstrates how modern Foundation Models can generate EO embeddings that preserve task-relevant information. These embeddings allow lightweight decoders up to two orders of magnitude smaller, accelerating training and inference.