Geopera is a data and research company. We work on the hard problems of turning satellite measurements into data you can check. When we claim an accuracy, we publish the validation.
Research areas: Bathymetry · Imagery Processing · World Models · User Experience
Seabed depth from optical imagery, checked against ICESat-2 laser photons and airborne LiDAR, with an uncertainty estimate on every pixel.
The automated pipeline behind every Geopera order: ortho, pansharpening, atmospheric correction, and mosaicing in one gapless flow we call Legato.
We train JEPA-style self-supervised models on our imagery archive to learn how Earth's surface looks and changes. That underpins our change detection.
Field studies of how practitioners find, order, and use imagery. What we learn shapes how Pera Portal works.
Jervis Bay, NSW: our depth product beside an airborne LiDAR survey, and the difference between them.
Bathymetry · Featured
We validate our satellite-derived bathymetry against measurements that didn't train it. In clear water it agrees with held-out ICESat-2 laser photons to within 38 cm (Bahamas lagoon, validation published June 2026). In turbid water it holds to about a metre against an independent airborne LiDAR survey at Jervis Bay, NSW. Every pixel ships with its own uncertainty estimate, and depths resolve to 15 to 20 m in clear water.
The accuracy starts with the inputs. The same automated processing that cleans every Geopera order feeds the depth model.
Bathymetry · Interactive
Drag the divider: on the left is the satellite capture of the Bahamas validation scene, on the right is the seabed depth we computed from it, with an uncertainty estimate on every pixel.
Imagery Processing
No image leaves Geopera raw. Every order runs the same automated chain (ortho correction, pansharpening, colour balancing, mosaicing) with nothing manual between stages and nothing skipped. The seamless result is the proof the process ran.
Imagery Processing
A distorted satellite view once convinced the internet a mega-dam was buckling. We reconstructed how the geometry error happened, and what correct terrain modelling looks like.
World Models · In progress
We are pretraining joint-embedding predictive models on our processed imagery archive. Instead of predicting pixels, the model learns to predict how a scene's representation evolves. That is the groundwork for change detection and automated data QC that generalise across sensors. We will publish the method and the validation when the results hold up.
User Experience · Ongoing
Remote sensing has a usability problem, and we treat it as a research question. We run interviews and workflow studies with GIS analysts, surveyors, and scientists, and the findings shape the ordering experience in Pera Portal. If you buy imagery for a living, we'd like to watch you work.
Methods, error analyses, and explainers from the team: the working notes behind how we process and validate imagery.
We're a small team working on bathymetry, imagery processing, and machine learning. If the problems above sound like yours, we should talk.
