EARTHZI
One satellite sees the Earth in radar. Another sees it in light. EARTHZI learns to recognise the same ground beneath both — retrieving the matching Sentinel-1 SAR or Sentinel-2 optical scene for any query, in either direction.
Two sensors, one Earth
Synthetic Aperture Radar penetrates cloud and night, but reads like noise to the eye. Optical imagery is intuitive, but useless after dark or under cloud. EARTHZI bridges the two — matching a radar patch to its optical counterpart, and back again, by embedding both into a shared space.
A fine-tuned RemoteCLIP model is trained with contrastive learning on paired Sentinel-1 / Sentinel-2 patches — pulling true geographic pairs together in embedding space, and pushing everything else apart.
At inference, a single query image is embedded and compared by cosine similarity against a pre-computed gallery, returning the five closest matches with rank and score.
Evaluated across 4,000 queries in both directions, on a held-out test split.
What EARTHZI does
Four capabilities, built around one shared embedding space.
SAR → Optical
Upload a Sentinel-1 radar patch and retrieve the optical scene of the same location — useful when the only available imagery is radar.
Optical → SAR
Upload a Sentinel-2 optical patch and retrieve the matching radar acquisition — the higher-performing direction at 35.7% Rank@1.
Fine-tuned RemoteCLIP
A remote-sensing-adapted CLIP backbone, contrastively fine-tuned on paired SAR–optical patches to align both modalities in one space.
Top-5 Ranked Matches
Every query returns its five nearest neighbours by cosine similarity, each with filename, similarity score, and rank position.
From query to match
Upload a patch
A SAR or optical image patch is submitted, along with the desired retrieval direction.
Generate embedding
The fine-tuned RemoteCLIP encoder maps the image to a vector in the shared SAR–optical embedding space.
Cosine similarity search
The query vector is compared against every pre-computed gallery embedding from the opposite modality.
Return top-5
The five closest matches are returned with rank, filename, and similarity score — rendered below.
Run a live retrieval
Upload a Sentinel-1 or Sentinel-2 patch, choose the retrieval direction, and EARTHZI will return its top-5 cross-modal matches with similarity scores.
Connects to POST /search on the EARTHZI inference backend.
Data, model & context
Keep It Free, Keep It Growing
Our services will always be free.
Donations help us improve, maintain, and expand them for everyone.