AgriSynth

Computer vision models for agriculture learn from labelled images: crops at different growth stages, across species and conditions. Collecting those photographs in the field is slow and expensive and depends on the season, and labelling every plant in every image by hand takes hours.

Our client, an AgriTech company building computer vision models for precision agriculture, wanted synthetic training data generated on demand instead – realistic field scenes with exact, consistent labels, covering any species at any growth stage, from seedlings to mature plants, across monocots (grasses and cereals) and dicots (broadleaf plants).

Rather than use an existing plant-modelling technique such as L-systems, we wrote a custom parametric growth model that builds plant geometry from mathematical parameters – species characteristics, growth stage, leaf shape – instead of pre-built assets. One framework handles both monocots and dicots, from narrow cereal blades to broadleaf structures.

Around the plants, the generator varies the rest of the scene – species mix, plant density, growth stage, soil, lighting, weather and camera angle – and renders each frame together with its labels. It runs inside the client's cloud ML training pipeline, so new batches are generated on demand.

The project was funded by Innovate UK, the UK's innovation agency – a strong validation of both the technical approach and its commercial potential.

The client can generate training data when they need it rather than when the crop is in the field, with every image labelled automatically and consistently – taking the seasonal wait and the manual labelling step out of building their datasets.

Because every scene is built from parameters, a batch covering different species, growth stages, densities or conditions is a change of settings rather than another round of field photography.