Industry field guide
Harvest imagery for AI from strawberry fields and greenhouses
Useful harvest footage shows why a picker chooses one fruit and leaves another.
- 2022 NAICS
- 111333 · 111419
- Practice pack
- Example and answer key (JSON)
- Updated
What this guide tests
Which fruit has a supported pick label, why is F2 unresolved, and what weakness does the draft evaluation split have?
See the specimen and fields · Work through the fictional records
Task and source context
Useful harvest footage shows why a picker chooses one fruit and leaves another. USDA has documented research on automated strawberry assessment and a greenhouse tomato project that includes machine-learning vision for harvesting. These are public research examples. A buyer would still need to specify the crop, harvest task and acceptance rules before a grower records a collection. USDA strawberry research · USDA greenhouse tomato project
Show the decision in context
For strawberries, capture differences in ripeness, fruit partly hidden by leaves, and examples a picker leaves untouched. Record the visibility of the fruit and a reviewer’s verified label. For covered tomato crops, include the cluster, visible stem and occlusion before and after the harvest decision. Tag the camera position and lighting. Confirm whether the project needs fruit detection, pick readiness or an action sequence. Each output needs different labels.
Hold back an entire row, cultivar or greenhouse section where appropriate. Nearby frames from the same plant can make evaluation look better without testing a new growing condition.
Agree what the grower can share
Get permission from the grower and the workers participating. Keep farm coordinates, proprietary cultivar trials, supplier information and people outside the approved frame unless those elements are authorized. Agree on biosafety, operating boundaries and whether cameras can be used near equipment.
Start with one approved row and a short task description. State which crop, views and labels are feasible, and what is excluded. Describe a possible collection
Industry research
The U.S. Census Bureau’s industry definitions establish which activities these codes cover. The workflow sources in this guide describe the task or record type. Fiund’s pilot below uses fictional material to test a specific decision within that scope.
111333 · Strawberry Farming
This U.S. industry comprises establishments primarily engaged in growing strawberries.
111419 · Other Food Crops Grown Under Cover
This U.S. industry comprises establishments primarily engaged in growing food crops (except mushrooms) under glass or protective cover.
Build a pilot sample
Set up this pilot
Film two consenting practice takes with paper fruit and an occluded pick. Keep all views from one plant out of the training split when it is used for evaluation.
Prepare the test packets
- Start with the worked case. Use the fictional record types below as a guide to the source files you need. Give each packet a stable ID and save its source versions with it.
- Change one fact. Make one packet with enough evidence to answer the question and one where the evidence points to a different answer. Keep the other fields fixed so a reviewer can tell what changed.
- Remove the decisive evidence. Give the reviewer a third packet with a plausible near-match in its place. Have them write what can still be established and what evidence would resolve the rest.
Fields the reviewer needs
- Plant, fruit and sequence ID
- Supports object and split grouping.
- Visibility and expert decision basis
- Separates image sufficiency from readiness.
- Action confirmation and evaluation partition
- Connects the demonstration to a checkable task.
Use Plan a farm, forest or fishery AI data sample to turn the specimen into a brief with a defined task, source and review owner.
Fictional exercise
Fruit visibility and an expert's pick decision are separate labels
A proposed strawberry task uses original textual plant and fruit annotations. No footage or physical picking is performed, and fictional labels do not establish agronomic harvest rules.
| Record ID | Source | What it records |
|---|---|---|
| HC1 | Plant A frame annotations | Fruit F1 fully visible, reviewer decision PICK under a stated original practice rubric. F2 partly leaf-occluded, reviewer decision UNRESOLVED because the needed surface is hidden. |
| HC2 | Action sequence | Confirmed fictional action removes F1; F2 remains. Camera stays on the same plant, with adjacent frames linked to the same sequence. |
| HC3 | Evaluation split draft | Training uses frame 1 of Plant A; test uses neighboring frame 2 of Plant A. A separately staged Plant B sequence is available but unused. |
Question for the model. Which fruit has a supported pick label, why is F2 unresolved, and what weakness does the draft evaluation split have?
Show the reference answer
HC1 supports F1's reviewer-assigned PICK label and HC2's corresponding action. F2 is unresolved because relevant surface evidence is occluded; remaining on the plant does not prove unripe status. HC3's neighboring frames share the same plant and sequence, making them a weak test of transfer to another growing condition. Hold out the separately staged Plant B sequence and preserve visibility, decision and action as distinct labels.
How to check the result
Ask for the answer, the record IDs that support it and any fact the evidence leaves unresolved. Have a reviewer compare the response against the reference before deciding whether the sample is ready to expand.
Errors to flag in this case
- Treating every unpicked or occluded fruit as a confirmed negative readiness label.
- Testing on adjacent frames of the training plant and claiming generalization to new conditions.
Change the inputs and run it again
Use a different authorized row or cultivar with its own reviewed rubric. Report task transfer separately from performance on familiar neighboring frames.
Score the decision and the evidence separately
Save the reviewer’s expected answer, supporting record IDs and unresolved fields in a separate answer key. Score the model on the decision, its citations and its handling of the deliberate gap. A correct answer citing a different case counts as an evidence error.
Before expanding collection, have a second reviewer run the same checks on the held-out packet. Keep related versions, views and exports of the same case together so they stay on one side of the training and evaluation split.
Check permission and scope
Use approved rows or synthetic task descriptions. Clear grower and participant rights; exclude coordinates, cultivar trials and unapproved people or operating equipment.
Remove precise farm coordinates and operator identifiers from a review copy. Keep the linkage between lot, sample and outcome intact.
Describe an authorized archive through Fiund’s existing-data assessment, or describe a new capture capability through the collection brief. The contributor agreement covers the terms for participation.
Sources and further reading
Resources for this workflow
Company and program reviews
Use these reviews to compare sourcing, capture or evaluation approaches. Confirm the applicable program, accepted sources and terms with the provider.
Compare the approach
truelabel review for teams collecting physical AI data
Compare how a capture team specifies devices, processing and quality review for an outdoor recording pilot.
Compare the approach
Graviti review for workplace video and multimodal training data
Compare a catalog with commissioned multimodal collection when a particular crop, setting or viewpoint matters.
Related industry examples
Industry guide · NAICS 111411
Mushroom tray and harvest-stage data for AI
A tray photo can show caps at several stages.
Industry guide · NAICS 111334
Berry harvest and handling images for AI review
Two photos can show that a berry changed appearance.
Industry guide · NAICS 111339
Stone fruit size and handling data for AI evaluation
A peach can look large in a close-up and still fall into a different size class when measured.