Multi-Camera AI Cuts Labelling for Crop Phenotyping

Researchers have found a way to teach computers to measure crops with far less hand labelling. The team used unlabeled pictures from six cameras mounted on a field robot that photographed cotton bolls from different angles. They trained two contrastive learning systems called SimCLR and MoCo to treat different views of the same plant as a matching pair. The best multi-camera versions performed about 14 percent better than standard image distortion training. That gain matters because manual annotation is slow and costly for large fields. With less labelling, breeders can track growth and stress across many plants more quickly. The work points toward cheaper automated phenotyping for cotton and other crops.

image
advertisement
"We will never ask you to login, or register through a third-party advertisement."