Plant perception · research model
Turn crop images into testable predictions.
Our visual classifier adapts a DINOv2 image encoder to field photographs and returns structured plant-problem predictions. Research teams can inspect the model version, scores and abstentions as they evaluate it for their own task. The current benchmark below establishes the starting point for this model product.
See the measured resultsBenchmark
Where the model stands.
Every human-judged photo in this comparison had been flagged by a hosted model. Confirmations inherit its problem name, which favours that model. Missed problems are not measured. The hosted comparator includes the model versions that originally read each photo.
The 95% interval resamples farms, not individual photos. The report records eight evaluations of the held-out farms, including reproductions and re-scoring. Repeated looks and the small sample limit how broadly this result can be interpreted.
Corrections remain an open problem
On a separate selection-fold diagnostic, the model returned the corrected name on 0 of 42 human-corrected photos. Those photos are not the held-out benchmark above. Expert labels and outcome training are next experiments, not a proven fix.
Download aggregate benchmark results · JSONSanitized aggregate results from the evaluation report generated 27 October 2026. Reviewed 4 November 2026; not a new benchmark run.
Next measurement
Evaluate an expert-labelled random sample, then test whether action and follow-up records improve results. Date not scheduled.
| Measure | Value |
|---|---|
| Evaluation set | 50 photos · 21 held-out farms |
| Our model | 23 of 50 · 46% |
| Hosted frontier model | 40 of 50 · 80% |
| Hosted advantage | 34 percentage points |
| 95% farm-bootstrap interval of the gap | 15 to 56 points |
| Agreement with hosted labels, not accuracy | AUROC 0.91 · 11 kinds · 2,704 photos |
| Measured | 27 October 2026 |
Research inference service
A prediction your software can use.
The private scorer returns a photo reference, model version, problem scores and an explicit abstention decision. These fields let a research pipeline compare predictions, track which model produced them and flag cases for review. The API contract is documented; customer access remains in development.
Developer referencePortable model research
Bring inference closer to capture.
A local ONNX export provides a starting point for running the model where crop images are captured. The goal is field inference with less dependence on connectivity. Phone integration, accuracy after export and on-device latency still require validation.
- Latency per photo
- Not yet measured on device
- Format
- ONNX, 16 plant and site categories
- artifact inspected 4 Nov 2026
Self-hosted inference · human review
Make uncertainty part of the record.
The product direction combines our own trained vision model with expert review. The research scorer already exposes abstention. The next step is to validate how those signals support a reviewer’s decisions, then evaluate whether verified follow-up observations improve the predictions.
Versioned model output
The scorer runs a model checkpoint through a local PyTorch or ONNX backend. Its structured response includes the model identifier and scores so results can be traced and compared.
Expert review of uncertain cases
The intended workflow presents uncertain cases to a person, preserves their corrections and links later observations. Reviewer time, missed problems and the value of additional models remain to be measured.
Research integration
Define a private model evaluation.
Tell us which growth, health or harvest-timing decision your model supports. We’ll agree on a benchmark, measure its errors and explain where evidence is missing. Your team receives a private report.
Tell Charles about your model, input format and proposed success metric. Private evaluation scope is agreed before any model or sensitive materials are exchanged.
charles@microhabitat.caOpens your email app. Nothing is sent from this website.