Treat underwater imaging as a data-quality problem before computer vision
A vision model cannot recover useful evidence when turbidity, backscatter, lighting geometry, or working distance makes the subject unreadable.
Context
The aquaculture monitoring initiative needs an honest route from pond imagery to future software-assisted observation without claiming an operational vision system.
Experiments and investigation
- ✓Define a repeatable tank matrix for turbidity, camera distance, visible light angle, and infrared illumination.
- ✓Record environmental and camera settings with each sample so results remain comparable.
- ✓Score human-readable visibility before considering model accuracy.
Decisions
- ✓Make image acquisition a standalone validation gate.
- ✓Do not select a production model or publish accuracy expectations before representative samples exist.
- ✓Preserve failed samples because they reveal operating limits.
Result and current status
A staged imaging protocol and evidence boundary are designed; an integrated field result has not yet been established.
Failure or limitation
No verified pond dataset or controlled comparison result is published yet.
Lessons learned
- ✓Optics and illumination are part of the data pipeline.
- ✓A negative visibility result can prevent expensive premature AI work.
Next step
Build the controlled tank test, capture labelled conditions, and publish only reproducible findings.