Getting Out of the Weeds with John Deere's Revolutionary AI Technology

Engineering & simulationNVIDIA GPU (Xavier/Orin)300 teraflops of compute; 411ms processing time per image

An embedded NVIDIA GPU solution processes images from 36 cameras in real-time to classify crops versus weeds, enabling the system to control 97 independent nozzles for precise herbicide application.

“we take those pictures, we put them through the machine learning algorithm on the brain of the system where this is an NVIDIA GPU embedded solution that we've developed with NVIDIA, and it's a really robust computer that can sit there in like 115 degrees Fahrenheit with no act of cooling or anything and it can do just fine.”

Engineering & simulationconvolutional neural networks

The team used supervised classification with convolutional neural networks to train the computer to distinguish between green crops and green weeds, achieving human-level accuracy on complex visual tasks.

“But, uh, you know, with the advent of deep learning and it's kind of specifically convolutional neural networks with supervised classification, we've actually been able to teach the computer the difference between crop and weeds at a, at a really, really high performance level.”

Engineering & simulation

The team hired weed science experts to label images, creating a multi-continental, multi-year, multi-season dataset to ensure the machine learning model had sufficient variety and quality for reliable performance in unknown fields.

“So what we've managed to do is sort of try to build a team and code that knowledge, train out our labeling workforce, and, uh, you know, just kind of to how hard this problem is, it seems like it takes a new labeler about 3 weeks before they kind of start making the right decisions, becoming productive…”

Engineering & simulationDino V2 (Meta)

The speaker mentions Meta's release of the Dino V2 model for image interpretation and segmentation as part of the broader landscape of AI developments influencing their field, though not explicitly stated as the tool used in the final product.

“So, uh, meta has been working on this problem, and they released recently released a model called Dino V2, which does a pretty good job, an image interpretation, and some image segmentation, and then sales force also, you know, has a, has some really interesting models they called blip…”