On-Orbit Demonstration of Backpropagation-Free Neural Compression for Earth Observation

Satellites increasingly carry low-power hardware for onboard AI, yet short ground-station passes and limited bandwidth keep downlink the bottleneck of Earth observation (EO). As a result, onboard compression largely determines how much usable data reaches the ground. Existing autoencoder-based compression requires large training datasets and must be retrained whenever the sensor or channel count changes. We propose a GPU-optimized neural compression pipeline that reduces the entire onboard computation to a single matrix multiplication and requires neither training nor a solver. The pipeline has been deployed and verified aboard a commercially operated, NVIDIA Jetson-based 6U CubeSat for blue-carbon monitoring, where it compresses on-orbit captures ${\sim}20\times$ while preserving image quality. We report experiments on diverse EO data under one frozen configuration, together with results from the on-orbit demonstration.