Mycology Classification

Project ID: AI-002 · Stage: Field trials with citizen science partners

Mission Profile

The mycology classifier augments field researchers and foragers with instant, high-accuracy identification. It couples a transformer backbone with curated datasets from herbariums, community science uploads, and lab-grown exemplars, producing species-level predictions and contextual safety notes even in offline conditions.

Core Capabilities

Hierarchical taxonomy: Predicts kingdom through species, surfacing genus-level confidence intervals when ambiguity is detected.
Risk envelopes: Maps each identification against toxicity databases and generates a color-coded advisory with edible, caution, or danger flags.
Habitat intelligence: Merges geospatial layers—soil composition, host trees, and climate history—to recommend likely co-located species.

Model Development

Phase	Dataset Composition	Outcome
Pretraining	2.4M public fungi images, herbarium scans, synthetic augmentations	Established foundational morphological embeddings.
Fine-tuning	180k lab-curated images with spore print spectra	Lifted macro accuracy from 82% to 93% on held-out evaluations.
Active learning	12k community submissions flagged for review	Added 37 rare species and reduced false positives on amanitas by 63%.

Deployment Architecture

Mobile bundle: Core model distilled to a 65 MB ONNX runtime optimized with quantization for Android and iOS native wrappers.
Offline cache: Region-specific model deltas and toxicity guides pre-sync when the device regains connectivity.
Knowledge loop: Confirmed identifications upload anonymized image embeddings plus environment notes for retraining cycles.

Ethical and Safety Considerations

All high-risk predictions require a secondary confirmation prompt before sharing edible recommendations.
Transparency module exposes top attention heatmaps so users can learn diagnostic features.
Partner botanists review ambiguous clusters monthly to avoid reinforcing dataset bias.

Current Results

93.2% top-1 accuracy across the 120-species validation battery; 98.7% top-3 accuracy.
Offline inference latency averages 148 ms on mid-tier devices; 72 ms on flagship models.
Shared taxonomy service now powers seasonal alerts on the Mushroom Forecasting project.

Next Milestones

Ship multilingual interface layers for Spanish, Japanese, and German partners.
Expand spore print recognition from lab captures to crowd-sourced macro lens attachments.
Publish an interpretability brief outlining key morphological differentiators per class.

Collaboration Signals

Program sponsor: Biodiversity Initiative
Primary engineer: Applied CV Lead, Biosensing Pod
Contact: lab@jessicawiedeman.com

Document updated: 2024-05-12