In-orbit computing can preserve more useful land-cover information when a satellite has little bandwidth to send imagery to Earth. An October 22, 2026 preprint by Yi Zhang and colleagues at Beijing Institute of Technology compares sending analytical results, compressed scene representations, and intermediate outputs for processing on the ground.
Their land-cover experiments used 4,940 images. At 0.1 bits per pixel, transmitting a classification map produced an overlap score of 0.60, compared with 0.28 for transmitting imagery before classification, using BPG compression. Higher scores indicate closer agreement with reference labels.
The result comes from dataset experiments, not an orbital trial. It also concerns a specific task: identifying land-cover categories. Sending a finished answer sacrifices information that other analyses might need.
The three operating modes address different constraints. Completing analysis onboard minimizes the downlink payload. Sending a compact representation retains more scene information. Splitting the model lets ground computers finish work that exceeds the satellite’s resources.
How Compression Preserves a Task
ESA’s Φsat-2 image-compression application provides a concrete example of processing distributed between orbit and Earth. An onboard encoder compresses single-band camera images; a dedicated ground decoder reconstructs them after transmission. GEO-K, an Italian company, developed the application.
ESA describes a development test centered on building detection. Researchers generated building masks from both original and reconstructed images, then compared the outputs. This tests whether compression preserves information needed for the application, rather than merely whether the reconstructed image looks similar.
That distinction explains why the smallest file is not automatically the best product. As New Space Economy’s coverage of onboard processing notes, some users still need traceable source imagery even when an automatically generated result would arrive sooner.
Selecting What to Send
ESA’s PhiFireAI application uses a different approach. It examines Φsat-2 imagery for smoke, burned ground, water, and unaffected areas. Its stated design suppresses downlink when neither fire nor burned areas are detected, reducing image volume and the work required of human analysts.
These applications make the processing choice concrete. A building-analysis service needs reconstructed information suitable for mapping. A wildfire-monitoring service can prioritize images containing relevant events. The useful transmission depends on the requested output.
Useful Books Available on Amazon
