Mycology Classification
An image model that suggests likely fungus species from a photo.
Concept. Not built; this page describes a proposed design.
What it would do
- Take a photo of a mushroom and return the most likely species, with a confidence score.
- Fall back to genus or family when it cannot tell species apart.
- Flag species known to be toxic, and lookalikes of toxic species.
- Work offline on a phone.
How it would work
- Start from an image model pretrained on general photos (for example a vision transformer).
- Fine-tune it on labeled fungus photos.
- Predict at several levels (family, genus, species) so a low-confidence species guess can fall back to genus.
- Look up each prediction in a toxicity table and show a warning when it matches a toxic species or a known lookalike.
- Shrink the model (distillation, quantization) so it runs on a phone.
Data it would need
| Data | Possible source | Notes |
|---|---|---|
| Labeled photos | iNaturalist research-grade observations, Mushroom Observer | Labels are crowd-sourced; some are wrong |
| Expert-checked test set | Herbarium records, mycologist review | Needed to measure accuracy honestly |
| Toxicity information | Field guides, poison control references | Must be reviewed by an expert |
| Location and date | Photo metadata | Can narrow down likely species |
Targets (assumptions, not measurements)
- Top-3 suggestions include the right species for most common species in the region.
- Response under 200 ms on a mid-range phone.
- Model file small enough to download on mobile data (under 100 MB).
Safety rules
- Never say a mushroom is safe to eat.
- Always show the top few candidates, not just one.
- Show a toxicity warning when any candidate in the top few is toxic or a known lookalike.
- Tell the user to confirm with an expert or a local mycological society.
Risks and open questions
- Many species cannot be told apart from a photo. Spore prints, smell, or microscopy may be needed.
- Training photos over-represent common, photogenic species and some regions.
- A confident wrong answer about a toxic species could cause serious harm.
- Should the app ask for several photos (cap top, gills, stem base) before suggesting anything?
- How to measure accuracy: an expert-checked test set the model never saw during training.