AI-generated illustration for Galileo Whispers. Generic spacecraft imaging hardware above imagined icy terrain; this is neither a mission photograph nor a compression test.
Spacecraft image compression changes how an image is represented before transmission. Lossless compression permits exact recovery of the input pixel values. Near-lossless compression permits changes within a specified error bound. Lossy compression accepts a wider class of irreversible changes, judged against the intended use of the image.
Those distinctions concern reconstruction of the recorded data. They are separate from correcting radio transmission errors, and from whether the instrument measured the original scene accurately.
Follow the numbers, not just the picture
A scientific detector produces numerical samples, often described in digital numbers, or DN. An image may contain several wavelength bands, each with its own noise, calibration, and scientific purpose. Two reconstructions can look almost identical on a screen while differing in the quantities a researcher uses.
Lossless compression exploits patterns or predictable relationships without throwing away the information needed to reverse the encoding. For example, an encoder can represent the difference between a sample and a prediction instead of writing every sample independently. The decoder must recover the original values exactly.
The historical CCSDS lossless-compression specification describes compression as source coding: reducing data volume to ease storage and transmission. Exact recovery is a property of the encoding–decoding pair when the necessary encoded information arrives intact.
It does not remove detector noise. A noisy recorded sample is recovered as the same noisy sample.
Near-lossless means a defined permitted difference
Near-lossless is technically a form of lossy compression, but with a deliberately constrained reconstruction error. In an absolute-error mode, the promise can be expressed as |reconstructed DN − original DN| ≤ E, where E is a selected bound.
Consider an invented pixel value of 1,000 DN and an illustrative bound of 2 DN:
| Reconstruction contract | Allowed outcome for this example |
|---|---|
| Lossless | Exactly 1,000 DN |
| Near-lossless with maximum absolute error 2 DN | Any integer from 998 through 1,002 DN |
| A lossy method without that per-pixel bound | Acceptance depends on its other specified criteria |
The interval follows directly from 1,000 ± 2. It is not an instrument measurement, a recommended mission setting, or a promise that a codec chooses every value in that interval equally often.
Eon and colleagues’ 2025 Landsat Next compression study examines absolute and relative error controls using proxy data. Relative modes can vary allowed error with signal level. The paper’s question is whether chosen settings meet the particular instrument’s scientific requirements; it is not whether any small-looking distortion is universally harmless.
A harmless-looking error can matter to a measurement
Suppose two illustrative samples are 1,000 and 1,003 DN. If each may change independently by up to 2 DN, their reconstructed difference could range from −1 to 7 DN. The subtraction combines the two possible errors. A small per-pixel bound therefore does not preserve every small contrast exactly.
Likewise, 2 DN is a 0.2% change relative to 1,000 DN, but 20% relative to 10 DN. These are arithmetic examples, not predictions for a particular codec’s error distribution. They explain why dark regions, faint features, ratios between bands, and thresholds deserve scientific tests beyond visual inspection.
Choosing a bound below an instrument’s typical noise does not make the added error literally zero. Researchers must check the relevant retrievals, spatial patterns, spectral relationships, and aggregation. A compression ratio alone cannot demonstrate scientific acceptability.
Compression and channel protection do different jobs
Compression intentionally changes the representation to save bits. Radio channel coding adds structured redundancy so a receiver can detect or correct some errors introduced during transmission. A mission can use both in sequence.
A correctly delivered near-lossless bitstream still reconstructs the permitted changed values. Error correction cannot restore information deliberately discarded before transmission. Conversely, lossless encoding does not ensure that corrupted or missing compressed bits can be decoded correctly.
The CCSDS historical document explicitly treats channel coding and packetization as separate error-control measures. Compressed bit errors can propagate into multiple reconstructed samples, depending on the format. The related explanation of spacecraft error-correcting codes follows that transmission problem.
The question to ask of a scientific image is therefore precise: exact relative to which input, bounded in which units, tested for which measurement, and protected against which transmission failures? “Compressed” by itself answers none of those questions.
