The problem
we solve
Not everything can, or should, be processed onboard. Once Earth Observation data reaches the ground, it arrives in growing volumes and in different forms: panchromatic and multispectral imagery, data from different sensors, acquisitions taken at different times. On their own, each of these is a partial view; the value comes from combining them.
Doing that well means handling fusion, alignment and enhancement as one connected problem, at high resolution and at the scale a mission actually produces.
Use cases
_Image fusion
Combining multiple sources, panchromatic, multispectral, and data from different sensors, into a single, more complete product.
- Multi-source combination: panchromatic, multispectral and multi-sensor inputs merged into one coherent output.
- Unified pipeline: fusion, alignment and enhancement handled together in a single pass.
- Complementary detail: information present in one source but missing from another is recovered in the combined product.
_Super-resolution
Reconstructing detail beyond what the sensor natively captures, on the ground, where power and memory limits no longer apply. Where image fusion brings different sources together into one complete picture, super-resolution targets resolution specifically, including through pansharpening, to extract detail the sensor alone couldn’t capture.
- Pansharpening: high-resolution panchromatic detail combined with lower-resolution multispectral bands to sharpen the final product.
- Beyond native resolution: detail reconstructed computationally, past what the raw acquisition provides.
- Ground-scale compute: heavier models than onboard constraints allow, run across full mission volumes.
_Change detection
Identifying and dating structural changes across sequences of satellite imagery, processing each new acquisition incrementally as it arrives.
- Incremental processing: each acquisition updates the model directly, without reprocessing the full history.
- Continuous state: context maintained across the time series, keeping detection accurate as new data arrives.
- Scale: efficient across large areas and long time spans.
How we do it
#Computer Vision
#Adaptive AI
#Predictive Intelligence
Ground processing removes the hard power and memory limits of onboard hardware, but it introduces a different challenge: scale. A pipeline that works on a single scene has to keep working when the input is thousands of acquisitions, arriving continuously, without a proportional blow-up in cost or time.
We design for that from the start. Our work builds on adaptive machine learning research: models that process data as a stream and update incrementally as new acquisitions arrive, keeping processing efficient as volume grows and letting a system stay current with incoming data. The same approach generalises across sensor types and resolutions, so a pipeline is built around the structure of the problem rather than tuned to a single dataset.