Skip to content

Agricultural AI Research

Exploring Synthetic Data for Agricultural AI

Investigating crop and weed image generation, annotation quality and robust perception across diverse agricultural environments.

Research direction. Imagery is schematic; no datasets or benchmark results are published yet.

Dataset sampleSchematic · static fixture
Schematic agricultural dataset sample shown as a source image, a class mask with soil, crop and weed, and its metadata record. Static illustration, not a real field image.

Research challenges

What makes field perception difficult

Seven sources of difficulty the research is designed to separate and measure.

Evidence: Conceptual · R-007
  • Crop and weed similarity

    Look-alike species share colour, texture and leaf shape, especially when young.

  • Lighting variation

    Sun angle, shadows and glare change appearance throughout the day.

  • Seasonal differences

    Colour, residue and canopy change across the season.

  • Crop growth stages

    Seedlings and mature plants are effectively different visual classes.

  • Soil and field backgrounds

    Soil colour, moisture and residue cover alter contrast with plants.

  • Limited labelled datasets

    Pixel-level labels are slow to produce and rarely cover every condition.

  • Generalisation across farms

    Models trained on one region or cultivar may fail on another.

Experimental pipeline

From field scenario to failure review

The intended experimental pipeline; real and synthetic sources follow the same validation and evaluation steps.

  1. 01

    Field Image or Synthetic Scenario

    Licensed field imagery or a parameterised scenario description.

  2. 02

    Image Generation or Selection

    Generate variations, or select real images to cover target conditions.

  3. 03

    Crop and Weed Labelling

    Pixel-level classes from scene ground truth or annotators.

  4. 04

    Segmentation Mask Validation

    Check class consistency, boundaries and empty or merged regions.

  5. 05

    Dataset Preparation

    Field-level splits so no farm appears in both training and test.

  6. 06

    Model Evaluation

    Per-class and per-condition metrics on held-out real imagery.

  7. 07

    Failure Review

    Inspect look-alike confusions and condition-specific misses.

Dataset explorer

Agricultural Dataset Explorer

One schematic dataset sample viewed five ways. It explains how pixel-level labels and dataset metadata fit together. Static fixture — not a real field image.

Illustrative Dataset Explorer — Static Schematic Fixture

Illustrative Demonstration
View mode
  • 0 · soil / background
  • 1 · crop
  • 2 · weed

Combined Segmentation

The full mask: each pixel holds exactly one class index — soil 0, crop 1, weed 2.

Pixel-level labels

A segmentation mask is an image the same size as the input, where each pixel stores a class index instead of a colour. Colours here are only for display.

Dataset structure

dataset/
  images/      RGB samples
  masks/       class-index PNGs
  splits/      field-level train/val/test lists
  metadata.json
  dataset_card.md

Static fixture: one original schematic sample and an example record. It is not a real field image and not part of a published dataset.

Scenario variation

How one scenario changes with conditions

Vary lighting, growth stage and task to see how the scenario description and what it tests change. Illustrative only: no generative model or field simulation runs in your browser.

Illustrative Scenario Variation — Not a Live Simulation

Illustrative Demonstration
Lighting
Growth stage
Task
ILLUSTRATION · NOT A SIMULATION OUTPUT

lighting=low · stage=early · task=weed

Weed Detection · Early stage · Low Light

This variation combines small seedlings that look most like weeds with long shadows that can be mistaken for plants. It tests how many weeds are missed or confused with crops.

Held fixed
Crop type, field layout, camera height and annotation rules
Varied
Lighting (low light) and growth stage (early)
Compare against
The same task under bright light at an intermediate stage

Illustration only. No generative model, field simulation or dataset generation runs in your browser.

Research evaluation

How agricultural models would be evaluated

Metrics and comparisons the research uses. Benchmark values will be published only when verified and approved — none are yet.

Segmentation metrics

Intersection over Union|A ∩ B| / |A ∪ B|
Per-class overlap; reported separately for crop and weed.
Dice Score2 |A ∩ B| / (|A| + |B|)
Overlap measure closely related to IoU.
Precision and RecallTP/(TP+FP), TP/(TP+FN)
Weed false alarms versus missed weeds — they carry different costs in the field.

Robustness and data quality

Performance by field conditionmetric | lighting, stage, field
The same metric computed per condition, not one aggregate.
Real-versus-synthetic comparisonreal-only vs real+synthetic
Controlled comparison on the same held-out real test set.
Annotation qualityagreement, flagged masks
Label consistency checked before results are trusted.

Suitability depends on context. TP, FP and FN are true positives, false positives and false negatives; A and B are predicted and reference regions. A metric that suits detection may mislead for segmentation, and class imbalance, label quality and the cost of each error type all change which measures matter.

Comparison protocol

Evidence: Unavailable · R-008

What changes

Trained on real labelled field images only.

Held fixed

Field-level splits, architecture, budget, seeds.

Question answered

What does real data alone achieve?

Results matrix

Results by configuration and metric. Cells without verified, approved evidence state that no result is published.
ConfigurationIoU (crop)IoU (weed)DicePrecisionRecall
Real-only baselineNo verified resultNo verified resultNo verified resultNo verified resultNo verified result
Real + syntheticNo verified resultNo verified resultNo verified resultNo verified resultNo verified result
Synthetic-onlyNo verified resultNo verified resultNo verified resultNo verified resultNo verified result
Unseen-farm testNo verified resultNo verified resultNo verified resultNo verified resultNo verified result

No results are published: no experiment has been verified and approved in the evidence register. Values will appear here only when they are.

Applications

Where this research is relevant

Research relevance, not deployments. Outputs listed are technical artefacts a study could produce.

  • Precision agriculture

    Mapping crop cover and weed pressure at fine resolution.

    Possible research output

    • Per-condition segmentation analysis
  • Agricultural robotics

    Perception for robots operating in variable field conditions.

    Possible research output

    • Lighting-sweep evaluation sets
  • Automated weeding

    Separating weeds from crops at close range.

    Possible research output

    • Look-alike evaluation subsets
  • Crop monitoring

    Tracking growth stage and canopy over time.

    Possible research output

    • Growth-stage annotated samples
  • Plant health analysis

    Investigating visible stress indicators with agronomists.

    Possible research output

    • Expert-defined label schema