A Vision-Language Framework for Multispectral Scene Representation Using Language-Grounded Features

Enes Karanfil, Nevrez Imamoglu, Erkut Erdem, Aykut Erdem {enes.karanfil,nevrez.imamoglu}@[email protected]@ku.edu.tr · 2025

Read A Vision-Language Framework for Multispectral Scene Representation Using Language-Grounded Features by Enes Karanfil, Nevrez Imamoglu, Erkut Erdem, Aykut Erdem {enes.karanfil,nevrez.imamoglu}@[email protected]@ku.edu.tr (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-10144.txt.

Description

Enes Karanfil, Nevrez Imamoglu, Erkut Erdem, and Aykut Erdem wrote this 2025 paper on teaching satellite scene models to use more than visible color. Remote sensing scenes, farmland, coasts, snowfields, can fool ordinary red-green-blue descriptions. Their framework, Spectral LLaVA, aligns multispectral satellite data with a vision-language model.

Built on Sentinel-2 imagery from a large public archive, the system keeps a multispectral vision backbone frozen and trains only a light projection layer to line its features up with language. Experiments cover classifying scenes by probing and jointly classifying and describing them in words. Adding the extra spectral bands sharpened both jobs, especially where color alone misleads, like haze or snow. The resulting descriptions came out more detailed and accurate, the authors report.

Catalog record

CREATOREnes Karanfil, Nevrez Imamoglu, Erkut Erdem, Aykut Erdem {enes.karanfil,nevrez.imamoglu}@[email protected]@ku.edu.tr
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE18 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-10144

Browse the full library · About the Tower of Babel