Reading a field like a model does
What a neural network actually sees when it looks at a crop — and why that changes how we collect training data.
When we say a model 'reads' a field, we mean something precise: it converts pixels and spectra into structured signals — vigour, stress, stage, anomaly — and then into decisions. Understanding what the model sees is the difference between a tool farmers trust and a black box they ignore.
Beyond green pixels
The human eye and a computer vision model attend to different things. A model may weigh texture, canopy temperature proxies and spectral indices more heavily than overall colour. In practice this means a stressed canopy can look 'fine' in true colour while the model sees it early — but only if it was trained on that stress.
Why ground truth is the bottleneck
The gap between the field and the model is closed entirely by data. If a training set contains mostly healthy canopies photographed in good light, the model learns a world without problems. Our research protocol exists to prevent that: plots are labelled by agronomists, symptoms are photographed as they appear, and negative examples are collected as carefully as positive ones.
Trust built from transparency
Models earn trust the same way people do — through demonstrated reliability and honest limits. That is why we publish our validation methods, share error analyses and let agronomists audit model behaviour on their own fields.