Skip to content

Private evaluations

What can your crop model predict?

We test a model against reviewed private observations and return a report covering its performance and limits. Choose a growth, health or harvest-timing question. The public example below is fictional. Our ten years of raw field data stay in-house.

See a sample report
Available to explore
Explore the fictional timeline and its synthetic JSON example.
Internal preparation
We are checking planter identities and labels before using matched visits in a private evaluation.
Planned
Wearable-camera video. No footage captured yet.

Fictional example

What your team receives

A report compares your model versions on the same cases. The figures here illustrate the format. Photos, individual observations and raw data stay at MicroHabitat.

Choose an evaluation question

Which version predicts harvest readiness more closely?

Mean absolute error over 60 forecasts

Version A
3.2 days
Version B
2.1 days

Version B was closer to the grower-confirmed readiness date by 1.1 days on average. Review early and late predictions before a field trial.

Synthetic demonstration records. No field photos or video are included.

Explore a plant’s record over time.

Explore a fictional planter’s dated observations, care actions and recorded changes. These public examples illustrate the format. Private evaluations run on separate, reviewed observations.

  1. 01Choose a plant

    Pick a crop and planter from the list.

  2. 02Follow the visits

    See what was observed, what was done and what changed.

  3. 03Inspect the example

    Read the fictional observations in a structured JSON file.

Synthetic example

Lettuce · Rooftop A

P-1872 · 8 dated visits. Read the observation, action and recorded change at each visit. A change does not prove the action caused it.

  1. Week 1

    nitrogen deficiency

    Action · treated
    First record
  2. Week 2

    no problem found

    Action · no action
    recovered
  3. Week 3

    no problem found

    Action · watered
    unchanged
  4. Week 4

    no problem found

    Action · fed
    recovered
  5. Week 5

    no problem found

    Action · no action
    recovered
  6. Week 6

    no problem found

    Action · watered
    unchanged
  7. Week 7

    no problem found

    Action · harvested
    recovered
  8. Week 8

    leaf scorch

    Action · no action
    worse

Fictional example · one planter over time

Use the fictional example

The JSON contains only the 20 fictional planters shown in this demonstration. Raw field observations, photographs and videos stay in-house.

Every record in this demonstration is fictional. Matching real visits and verifying labels are part of the private evaluation process.

Three questions we can evaluate

Crop growth

Question
Can the model estimate change in size or condition across two visits to the same planter? We check identity, dates, viewpoint and whether the target can be measured reliably.
Evaluation method
Compare predictions with reviewed measurements on held-out farms. Report error by crop, growing condition and time between visits.
Available evidence
16,244 photos, 196 farms, Jun to Nov 2026
source inventory reported 3 Nov 2026
Sequences with eight or more consecutive weeks
Not yet verified from source records

Plant health

Question
Does the model detect a worsening plant, and how many healthy plants does it flag unnecessarily? Include unflagged images when estimating misses.
Evaluation method
Compare predictions with independent expert judgments. Return detection rates, false alarms and cases that need human review.
Available evidence
Historical diagnosis benchmark on 50 problem photographs from 21 held-out farms
27 Oct 2026

Harvest timing

Question
How closely does a model’s predicted harvest window match the recorded harvest for the same crop? Farm-level totals alone cannot establish when an individual plant was ready.
Evaluation method
Agree on the decision window and timing tolerance. Evaluate early and late predictions separately. First verify enough linked harvest dates and visits to support a comparison.
Farmer-reported plant issues
952
source inventory reported 3 Nov 2026
Harvest records with weight
1,673 records, 3,678 kg
Apr to Oct 2026
Weekly follow-ups
1,507
since 11 Aug 2026

How we check private evidence

Review data handling and permitted uses
GPS
Location metadata is reported removed in the farm upload pipeline. Verify this for each evaluation subset.
Faces
Review private images for identifiable people and restrict access. Public examples exclude clients.
Permitted use
Data rights in progress. MicroHabitat authorizes internal model training and evaluation. Before a customer engagement, we check the permitted use of each private subset. The customer receives findings; raw records stay in-house.
Flagged photos
Human assessment is incomplete. Record who reviewed each label, include unflagged cases and report how selection affects the result.
Duplicate frames
The historical training export contained 16,658 records; 14,126 remained after deduplication. Each new evaluation requires its own checked inventory.
export of 26 Oct 2026

Define your private evaluation

Tell us the crop, decision and model you want tested. We agree on reference measurements, success criteria, price and timeline before starting. The customer receives a report; our raw archive stays in-house.

Tell us which crop you work with and what your model needs to predict: growth, health or harvest timing. We’ll discuss the reference measurements, test protocol and findings your team needs. Raw field data stay in-house.

charles@microhabitat.ca

Opens your email app. Nothing is sent from this website.

Website assistant

What would you like to know?

Ask about our crop AI, Farm Bot or field evaluations.

Enter to send · Shift + Enter for a new line

Anthropic receives your question, relevant earlier messages from this conversation and relevant source passages. Privacy