Emma Bryce at Anthropocene: This is more than just another AI gimmick, the University of Bonn researchers say: their new tool could help guide farmers on how to raise crops more sustainably without sacrificing yields—for instance, pinpointing places in the field where adding less fertilizer could result in the same productivity, or more.
To build their tool, they trained a machine learning algorithm—often described as ‘artificial intelligence’ or AI—with thousands of images of growing crops. These images were gathered from previous studies on three different sets of crops, photographed by overhead cameras or drones: thale cress test plants grown in a lab; a field of growing cauliflowers; and a mixed field of wheat interspersed with faba beans. Overall, the researchers plugged more than 100,000 images into the algorithm, which captured the three sets of growing crops over a period of months to years.
Through this data-crunching, the algorithm learned to link certain visual features of a crop in its early stages, with how its growth unfolded over time. The researchers also trained the algorithm to accurately identify specific crop traits, such as leaf area and estimated biomass, which can be linked to yields.
After this intense training period, the researchers made a striking discovery: if they presented the algorithm with a single image of a crop in the early phases of its growth, it could use this as a foundation to generate dozens of artificial images, which predicted how the crop would look at different stages of its future growth. These artificial images were strikingly accurate, closely comparing to real images of those crops in the field, says Lukas Drees, a doctoral student at the University of Boon, and lead author on the new research.
More here.
