Compressed learning based semantic compression utilizing CCSDS123.0-B-2

Kurzfassung

Downlink is a central bottleneck in Earth observation. Semantic compression transmits only the features required by a downstream task rather than the full imagery. Learned feature coders, however, require per task trained entropy models and can be computationally expensive to run onboard. We present a study of the standardized CCSDS-123.0-B-2 codec as a task-agnostic, near-lossless entropy coder for neural features one of the few standards offering a tunable, bounded per-sample error up to the input bit depth. We treat a feature as a width-one multispectral cube and cast its compression as a multiple measurement vector (MMV) problem with a compressed learning linear sketch, considering both a random Gaussian and a learned sketch matrix. To enable training, we approximate the codec’s quantization with a uniform straight-through quantizer and show that this surrogate tracks the real CCSDS bitrate across the absolute error limit sweep. We then investigate neural feature properties, in particular spectral axis smoothness, that benefit the predictor, and use a residual ratio penalty to learn smoother sketch measurements. Finally, on EuroSAT-MS we benchmark CCSDS based feature compression against a learned, task-specific coder, and discuss the trade-offs that make a standardized, task-agnostic codec a practical onboard choice.