Skip to content
LexDataLexData
PlatformIndustriesCustomers
DocsThe Field GuideBlogWhy models drift
AboutCareersSecurityContact
Log inStart now

// AGRICULTURE

Produce Sorting & Grading

Produce Sorting & Grading. What it takes to run it.

Each piece boxed and tracked for one grade, sized off the box, a bruise masked for its own area, before the diverter acts.

Start with your footageBook a demo →All of agriculture →

// 01 · The question

“Which grade is this piece of produce, sized and classed before the diverter has to act?”

// 02 · Why it is hard

The belt sets the clock, not the plant. The grade for a piece has to exist before it reaches the diverter, well under a second of belt travel, which is why the model runs beside the belt rather than anywhere with a round trip in it. A new cultivar or the first week of a new harvest moves what counts as a defect without moving a single setting on the line.

// 03 · What the labels have to be

Bounding box + tracking

Decided before the first frame

A grade decided on one frame and lost on the next is a wrong call reaching the line's own diverter, because the same piece crosses the camera for a dozen frames before the jet. A track keeps one identity on each piece so one grade reaches the diverter once. The box's own width across the belt gives a size without a mask; a mask only earns its labeling cost on the one class where severity is itself an area, a bruise, and the rest of the grade stays a box question.

This is the decision that is expensive to reverse. The geometry has to match what the answer contains, and finding out it does not means labelling the set a second time.

// 03a · What it finds

Each piece of produce on the belt

a box on every piece as it passes the camera over the conveyor, with a track that keeps one identity on it from the camera to the diverter

The size

read off the box's own width across the belt, no separate measurement station needed

The defect class

a second label on the same box, the grader's own words for what the line sees, a bruise, a split, a russet mark, sunburn, insect damage or clean

A piece the model doubts

a mark that could be a bruise or a stem scar, held for the grader's table rather than scored either way

// 03b · What it checks

  1. 01

    The grade's track held to one identity before it reaches the diverter, so a piece seen on a dozen frames is graded once, not decided twice.

  2. 02

    The defect class against the grading sheet already in use on this belt, in the grader's own words, not a generic defect list.

  3. 03

    The bruise class alone against a size, since severity there is an area and a mask is what carries it; every other class stays a box question.

  4. 04

    A doubted piece sent to the grader's table rather than the reject lane, with the frame attached, since a piece sent to juice by a model that was unsure is lost revenue nobody counts.

// 03c · What the site provides

  • The camera already over the belt, mounted to see the full width where pieces pass singly or in a known lane; no new scanner, no new diverter.
  • Frames from this belt and this season's crop, not a stock produce photograph; the model learns this packhouse's own light and this cultivar's own marks.
  • The grading sheet in the packhouse's own words, since "leather" or "blush" are the grader's terms and the class list has to carry them.
  • A webhook into whatever already reads the line's diverter controller, and a grader's table for the piece the model will not call either way.

// 04 · How we run it

Three parts of one loop, on this job.

LexAnnotate

Field Precision

Annotate crops, equipment, obstacles, and pathways with polygons, polylines, and semantic segmentation. Tag environmental factors like dust, shadows, and occlusions. 80k+ image annotations in production deployments.

LexInsight

Yield Intelligence

Transform plant-level detections into field-level intelligence. Map disease patterns across zones against a healthy baseline the grower names, and count what the camera sees to estimate yield.

LexAlert

Navigation Alerts

Monitor autonomous navigation models against seasonal drift. Retrain as fields change to keep navigation reliable through the seasons.

How the loop fits together →

// 05 · What breaks it after launch

Covariate shift

The season turned

→Covariate shift →

A new cultivar arrives with a different ground colour, an early or late harvest changes how much of the crop carries a given mark, and the light in the packhouse itself changes through the season. None of that is a line fault; it is the crop changing under a model trained on last season's fruit, and the grader's table is where the mismatch shows up first.

It is not the only one that can get this use case, it is the one that usually gets it first. All ten conditions.

// 06 · In agriculture

GPS-Denied Autonomy

Full write-up ↓Every case study →

The Challenge

Unreliable GPS under tree canopies and low visibility. Models failed in unstructured, obstacle-heavy terrains.

Our Execution

Annotated thousands of images with polygons and polylines. Tagged diverse environmental factors like dust, shadows, and occlusions.

The Result

Enhanced autonomy in GPS-denied environments. Faster deployment of AI-powered navigation.

80k+

Image annotations in production

// 06a · Questions we get asked

Does this replace the optical sorter or the mechanical diverter already on the line?

No. It gives the diverter's own controller a grade and a position per piece over a webhook; the diverter, the jet and the lanes stay the plant's own equipment. What changes is what decided the grade.

Can it grade a defect whose severity depends on size, like a bruise?

Only on the one class where severity is itself an area. The box gives presence and the piece's own size; a mask goes on the bruise class alone, since a bruise above a certain area is a different grade from the same mark below it.

What happens to a piece the model is not sure about?

It goes to the grader's table at the end of the belt, with the frame attached, rather than to the reject lane. A piece sent to juice by a model that was unsure is lost revenue nobody counts, so doubt costs a review, not the product.

What happens when a new cultivar or harvest arrives?

The ground colour, the mark pattern and the packhouse light itself all move together. The graders' own corrections landing in one direction, overriding a grade thirty times before lunch, is the signal, and a short window of those frames retrains the model mid-harvest with no downtime on the belt.

Read next

Automated sorting on the packhouse belt, from the box on the apple to the diverter →Surface defect detection: the same severity-by-area argument on a machined part instead of a piece of fruit →What a new cultivar or an early harvest does to a model trained on last season's fruit →
Start with your footageBook a demo →

// 07 · Also in agriculture

Crop Health & Disease Detection→Weed Identification & Classification→Autonomous Navigation→Yield Estimation→Field Boundary Segmentation→Livestock Monitoring→All forty-one→
Crop Health & Disease DetectionWeed Identification & ClassificationAutonomous NavigationYield EstimationField Boundary SegmentationLivestock MonitoringAll forty-one →

Send us a week of this footage. We’ll show you what comes back.

Start with your footageBook a demo →
LexData
LexData

Product

  • Platform
  • Industries
  • Use cases

Resources

  • Docs
  • The Field Guide
  • Blog
  • Why models drift

Industries

  • Energy & utilities
  • Oil & gas
  • Agriculture
  • Manufacturing
  • Insurance
  • Retail
  • Robotics

Company

  • About
  • Customers
  • Careers
  • Contact

Trust

  • Security
  • Privacy
  • Terms

Stay updated

What we learn running vision models in production.

See everything.
Miss nothing.

Stay updated

What we learn running vision models in production.

Terms of use & Privacy policy

© 2026 LexData Labs · All rights reserved