Plant Instance Segmentation for Crop Vision
Leaf-and-plant instance masks traced on organic, irregular boundaries with AI assistance - labeled under real field conditions so models survive deployment.
The challenge
Plants are the hard case for segmentation: leaves overlap and occlude each other, boundaries are organic and irregular, and the same species looks different at every growth stage. Manual polygon tracing at canopy density is so slow that agricultural datasets historically stayed small.
Field imagery adds capture noise that lab datasets never see - harsh and shifting sunlight, motion blur from handheld or vehicle-mounted cameras, dust on lenses, occlusion by neighboring foliage. A model trained only on clean imagery fails in its first week outdoors.
Disease and condition labeling compounds it: 'early blight versus nutrient burn' is a judgment call that different annotators will make differently unless the taxonomy is pinned down with visual decision rules.
Why this data matters
Crop-vision models are moving into production across precision spraying, autonomous weeding, disease early-warning and yield estimation - and each of those needs segmentation-grade data, not just classification labels, because acting on a plant means knowing exactly where it is.
Research institutions and product teams consume this data on different rhythms: research groups need versioned, citable datasets for publications; product teams need rolling versions that grow with the season. One pipeline can serve both if versioning is first-class.
Seasonality is a real constraint - the dataset that matters is the one that grows with the crop, capturing each growth stage as it happens. Miss a stage and you wait a year.
How we did it
- 1
Smart polygons on organic shapes
Smart polygons find the exact leaf or plant boundary from a click or rough box - an order of magnitude faster than manual tracing on irregular shapes - and annotators refine from there. Every mask still passes human review.
- 2
Locked disease taxonomy
A visual decision guide per condition - what early blight looks like versus nutrient burn, when to call a leaf 'healthy' - is agreed before scale-up, so annotators apply the same call. Ambiguous cases escalate to review rather than being guessed.
- 3
Real-condition labeling
Datasets deliberately keep the noise the deployed model will see - lighting shifts, blur, occlusion - because robustness comes from the data distribution, not from wishing the field looked like a lab.
- 4
Growth-stage coverage
Capture and annotation are planned across the season so the dataset covers each growth stage, instead of over-representing whichever week the project started.
- 5
Plant-level tracking
Greenhouse and row video is tracked plant-by-plant over time, turning footage into longitudinal per-plant records - the substrate for growth-rate and treatment-response models.
- 6
Versioned releases
Each release is immutable with fixed splits and quality metrics; research teams cite a version, product teams roll forward as the season adds data.
Inside the dataset
- Instance masks per plant and per leaf, human-verified
- Condition / disease labels applied against a locked visual guide
- Growth-stage coverage across the season
- Field-noise retained deliberately (lighting, blur, occlusion)
- Longitudinal plant-level tracks from greenhouse / row video
Quality control
Two review layers: mask quality (does the polygon hug the true organic boundary?) and taxonomy consistency (would a second annotator make the same disease call?). Double-annotation samples measure agreement per condition class, and low-agreement classes get their decision guides tightened before more data is labeled.
Versions ship with per-class counts, agreement statistics and growth-stage coverage, so agronomists and ML engineers can both see what the dataset actually contains.
What was delivered
- Instance masks per plant and leaf, human-verified
- Condition labels applied against a locked visual guide
- Field-noise retained deliberately for deployment robustness
- Versioned exports with quality metrics
Exported as COCO segmentation, YOLO-seg and PNG masks, with train/valid/test splits per version.
What the data enables
- Disease-detection and early-warning models
- Precision spraying and autonomous weeding (act-on-plant segmentation)
- Counting and yield-estimation models
- Longitudinal growth and treatment-response analysis
Want a dataset like this?
Send us your images or video - we annotate, review, and deliver a training-ready dataset in your format.
Book an intro call