←
Locamage16 posts
Locamage avatar

Locamage

@Locamage

Understanding the Earth

◷ Joined October 2026▧ 7 Photos and videos
Locamage avatar
Locamage@Locamage·Oct 26

Meet Finch 1, our first step toward one model that works with any sensor, at any resolution, anywhere on Earth. Trained on imagery from 50+ sensors across satellites, aircraft and drones.

0:34
Locamage avatar
Locamage@Locamage·3h

With it, we're releasing the Finch embeddings of every sample of all 34 benchmark tasks we report, 207 GB under CC BY 4.0, with checksums. Blog: locamage.com/blog/finch-1-0-… Model card: locamage.com/finch-1.0-model… Embeddings: embeddings.locamage.com

Locamage avatar
Locamage@Locamage·3h

Finch also handles sensors it has never seen. On 369-band NEON hyperspectral imagery, it matches UniverSat-B (46.9 mIoU), which was pretrained on that sensor. Here: a full Wyvern hyperspectral scene, also absent from training.

Post attachment
Locamage avatar
Locamage@Locamage·3h

With a frozen encoder and a linear or kNN readout: • Cropland from radar alone (CropHarvest China): 70.8%, best published 63.7 • Marine debris and oil spills (MADOS): 72.0 mIoU vs 68.5 • On par with OlmoEarth and UniverSat-B across 27 suite tasks

Post attachment
Locamage avatar
Locamage@Locamage·3h

The Finch 1.0 model card is out: a full evaluation of our Earth observation foundation model on 34 public benchmark tasks. One encoder for 50 satellite, aerial and drone sensors at native resolution. Frozen embeddings + a linear layer. 🧵

Post attachment
Locamage avatar
Locamage@Locamage·Oct 28

Day 3 of Finch 1: what changed. Crop types in France through a whole season, and crop mapping from radar alone. Same frozen model, same single linear layer.

Post attachmentPost attachment

Day 2 of Finch 1: what it sees. Land cover in France, buildings in Khartoum, roads and roofs around the Chesapeake Bay, and a wildfire scar in the US, from 20 cm aerial photos to 30 m satellite imagery. Every map comes from one frozen model and a single linear layer.

Post attachmentPost attachmentPost attachmentPost attachment
Locamage avatar
Locamage@Locamage·Oct 27

Day 2 of Finch 1: what it sees. Land cover in France, buildings in Khartoum, roads and roofs around the Chesapeake Bay, and a wildfire scar in the US, from 20 cm aerial photos to 30 m satellite imagery. Every map comes from one frozen model and a single linear layer.

Post attachmentPost attachmentPost attachmentPost attachment

Meet Finch 1, our first step toward one model that works with any sensor, at any resolution, anywhere on Earth. Trained on imagery from 50+ sensors across satellites, aircraft and drones.

0:34
Locamage avatar
Locamage@Locamage·Oct 26

If you are a researcher and want to work with us, email contact@locamage.com. Thanks to a16z Speedrun Alpha (@speedrun) and Google's TPU Research Cloud and TPU Builder Program for supporting this work.

Locamage avatar
Locamage@Locamage·Oct 26

It also handles aerial imagery from new providers and regions, like TreeSatAI's tree species mapping, with no changes to the model. With just a linear probe, it beats TerraMind-L and its ~20M-parameter trained decoder on DynamicEarthNet.

Locamage avatar
Locamage@Locamage·Oct 26

A few highlights from 32 public tasks (encoder frozen): • +12.6 vs OlmoEarth v1.2 on Sentinel-1 radar crop mapping • +11.3 vs AlphaEarth on PASTIS crops, Sentinel-2 only • +6.2 vs UniverSat-B on Landsat deforestation drivers • +4.9 vs OlmoEarth v1.3 on MADOS marine debris

Post attachment
Locamage avatar
Locamage@Locamage·Oct 26

Most geospatial foundation models learn from a few low-resolution satellites. We think the variety of imagery is worth scaling. Finch reads all of it, from 4.4 mm drone photos to 250 m MODIS, with one set of weights.