AstroLoc: Robust Space to Ground Image Localizer

TL;DR

We present AstroLoc, an accurate, robust localization model for space-based imagery of Earth. Trained jointly on astronaut photos from the International Space Station and images from Earth observing satellites, AstroLoc localizes via image retrieval over a database of satellite imagery. AstroLoc sets a new SOTA on existing space-based localization tasks like astronaut photography localization and Lost in Space orbit determination. Check out our demo to see it in action!

Demo

Goal

GIF courtey of EarthLoc&Match.

Astronaut Photography Localization (APL) is the task of identifying the location on Earth depicted in an image taken by astronauts in space with handheld cameras. In the microgravity environment, the astronauts can orient the cameras to image any portion of the visible Earth area (20M sqkm). When using high zoom lenses, the images may contain only a few sqkm of land area, making localization akin to finding a needle in a haystack.

Training Pipeline

AstroLoc is the first APL model to be trained on astronaut photos themselves. After converting 300k weakly labeled astronaut photos into a fully labeled training set, we pair astronaut photos with corresponding satellite images and train using a contrastive loss. At the same time, we jointly train on batches of clustered satellite images (following the method in EarthLoc), to ensure our model has a complete view of Earth's landforms.

Results

AstroLoc excels in various space-to-ground localization tasks. While trained for astronaut photography localization, AstroLoc can also be used to identify satellite location/orbit from imagery (the "Lost in Space" problem), and localize historical space shuttle photography, which was originally taken on film and digitized and doesn't come tied to trajectory data. Examples from each domain, which contain notable visual differences (domain gaps), are below.

BibTeX


        @misc{Berton_2025_astroloc,
        title={AstroLoc: Robust Space to Ground Image Localizer}, 
        author={Gabriele Berton and Alex Stoken and Carlo Masone},
        year={2025},
        eprint={2502.07003},
        archivePrefix={arXiv},
        primaryClass={cs.CV},
        url={https://arxiv.org/abs/2502.07003}
        }