Mycology · Field note
A camera adds a sense the sensors don't have: sight. The honest version of that sentence is longer, and worth reading.
Four sensors tell a Mycophyte chamber everything about its air. They tell it nothing about its mushrooms. A camera changes that: it can watch the crop the way a grower does, by appearance. But appearance is subtle, lighting is fickle, and the difference between "healthy pin set" and "the beginning of something bad" is exactly the difference a computer is worst at — which is why the vision module is framed honestly as an experimental flagging system, not a diagnostic oracle.
The hardware is deliberately modest: a USB camera pointing at the shelves, frames captured on a schedule, inference run on the device by a lightweight object detector from the YOLO family (or faster, with a pocket TPU). The detector is trained for chamber-relevant classes — mushrooms at various stages, substrate, casing, and anomalies. When it finds something it wasn't asked to see, or finds the substrate changing in an unexpected way, it records a flag. The flag is a message to a human: "look at this, right now." It is not a verdict.
Why the caution is warranted: molds and healthy mycelium share a visual vocabulary. Both are white fuzz to a first approximation; the differences are texture, color temperature, growth rate, and context — and some molds are expert at looking like something else. A detector can learn to separate them on a fixed camera with fixed lighting and a fixed strain; the moment any of those change, the detector's error rate rises and its "confidence" becomes less trustworthy. Any product that promises to "diagnose mold from a photo" is selling the demo, not the deployment.
What vision reliably adds is surveillance and history. The module can notice that a region of the casing changed color between Tuesday and Wednesday in a way the human eye, looking twice a day, would miss. It can count pins and track canopy development over a flush, which is genuinely useful data for a lab comparing strains. It can log an image of every anomaly event so the human review happens with evidence, not memory. These are real capabilities, and they compound: every flagged-and-reviewed image can be logged as a labeled example, slowly improving the detector's training set under human supervision.
There is also the question of what the images contain, which is a privacy question. Frames are processed on-device and never leave it by default. A grower who opts into sharing training examples does so explicitly, per image. The federated learning layer shares model updates, not frames — a separation documented in the docs.
The practical takeaway for anyone building this at home: mount the camera once and never move it, light the chamber from a fixed angle, and treat every flag as an instruction to open the chamber and look. The day the detector is right about an anomaly you didn't see — that is the day the vision module earns its place. Until then, it is a diligent intern who is wrong sometimes, and the system is built on the assumption that it will be.