Dzung Phan, Vinicius Lima
INFORMS 2023
The 2025 CVPR EarthVision Data Challenge organized by the Embed2Scale consortium invites researchers to push the boundaries of AI-based compression for Earth observation by developing lossy neural encoders that reduce SSL4EO-S12 (v1.1) data cubes to 1024-dimensional embeddings while preserving downstream analytical utility. Participants engaged in a three-week development phase (March 10–31 2025) submitting solutions via Eval.AI, followed by a one-week testing phase (April 3–5 2025) on a held-out dataset to establish the final leaderboard. The winners’ solutions and insights on the evaluation framework will be presented at the CVPR EarthVision workshop 2025. By rigorously evaluating self-supervised embeddings across diverse hidden tasks, this challenge seeks to advance the state of the art in lossy neural compression and promote transparent, open-science practices in Earth observation foundation modeling.