Like animals, plants transmit information through electrical signals. Bombadil learns to read them without tags and detects seven indicators of stress, even in never-before-seen species.
Living plants circulate electrical signals from their roots to their branches, and these signals can be picked up with a simple pair of electrodes planted in the conductive tissues. Until now, the machine learning models that used them were designed for a single task, a single species or a single environment, which prevented scaling.
From the laboratory to the forest
James Kirk's team presents Bombadil, an unsupervised pre-trained foundation model on more than 4 million hours of recordings from 100 plant species, in ten types of environments: laboratories, greenhouses, commercial orchards and natural forests. From these signals alone, the model tracks the evolution of seven physiological indicators of stress, and manages to do so for species and types of stress that had no label in the training data.
The authors are targeting applications in agriculture, forestry and horticulture, including for the care of plants in private homes: a plant could "signal" a lack of water or an attack before symptoms become visible.
Read the full article on the original source
Go to the source →