COSMO: Contrastive Satellite Image Embedding Techniques Reflecting Expert Weather Bulletins
IEEE Access · 2026 · 14 · 92385-92398
DOI: 10.1109/access.2026.3702845 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Min-Hyoung Kim und 1 weiteren Autor:innen. Die Arbeit nennt 38 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Institute of Information and Communications Technology Planning and Evaluation, Korean Government [Ministry of Science and ICT (MSIT)] [AI Star Fellowship (Kookmin University)].
Abstract (Originalsprache)
A similarity search of geostationary meteorological satellite imagery is a basic tool for analyzing atmospheric conditions. However, if one simply reuses vision encoders pretrained on natural-image datasets, similarity is computed in the embedding spaces they induce; thus, the resulting representations emphasize visual appearance and dataset-specific semantics rather than the meteorological similarity implied by expert analyses. To mitigate this mismatch between visual and meteorological similarities, we propose COSMO, a contrastive representation learning framework that reshapes satellite image embeddings using operational weather bulletins as weak textual supervision. COSMO keeps image and text encoders frozen and learns lightweight projection heads that align images and bulletins in a shared embedding space with a multi-positive contrastive loss regularized by an auxiliary reconstruction term that encourages the projected image embeddings to preserve the structure of the original visual features. Using a time-aligned corpus constructed from COMS geostationary imagery and robustly preprocessed weather bulletins provided by the Korea Meteorological Administration (KMA), we evaluated the learned embeddings on retrieval-style tasks, including image-to-image and image-to-text searches based on bulletin-derived semantic similarity. COSMO consistently outperforms purely visual baselines and non-contrastive ablations on these retrieval tasks, and analyses of the learned embedding space further suggest that bulletin-guided contrastive learning makes satellite image embeddings more consistent with meteorological similarities.
Förderung
- Institute of Information and Communications Technology Planning and Evaluation
- Korean Government [Ministry of Science and ICT (MSIT)] [AI Star Fellowship (Kookmin University)]
Zitieren
APA
Kim, M., & Park, H. (2026). COSMO: Contrastive Satellite Image Embedding Techniques Reflecting Expert Weather Bulletins. IEEE Access, 14, 92385-92398. https://doi.org/10.1109/access.2026.3702845
Vancouver
Kim M, Park H. COSMO: Contrastive Satellite Image Embedding Techniques Reflecting Expert Weather Bulletins. IEEE Access. 2026;14:92385-92398. doi:10.1109/access.2026.3702845
BibTeX
@article{Kim2026,
author = {Kim, Min-Hyoung and Park, Ha-Myung},
title = {COSMO: Contrastive Satellite Image Embedding Techniques Reflecting Expert Weather Bulletins},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {92385--92398},
doi = {10.1109/access.2026.3702845}
}
EUVIMED