Blue River Technology, a subsidiary of US-based farming equipment maker John Deere, is using deep machine learning tools to control weeds and reduce the cost of crop spraying. The intelligent spraying system can differentiate weeds from crop plants.

The company is using the PyTorch machine-learning framework, originally built by Facebook AI, to train robotic crop sprayers to map weeds as they move through a field. Powered by Python programming language, PyTorch is an open-source deep-learning framework that is flexible and modular for research. The library allows developers to build, deploy and add to new AI models.

Weed control is a bit trickier than pest control. As weeds and crops are both plants, it is difficult to identify a weed from a crop. As a result, the sprayer needs to be intelligent to spray herbicide on a weed but doesn’t hit a crop plant. For this, the machine needs to make real-time decisions to identify a crop from a weed. 

Using a high-resolution camera and the system scans the field. The data from the camera is sent to a convolutional neural network using PyTorch, an open-source machine learning library, to identify weed from the crop, create a pixel-accurate map of the field and thus guide the robot sprayer to hit only the proper targets. This entire process takes mere milliseconds of time so that the robot sprayer can traverse the field as quickly as possible and still accurately target weeds while leaving crops untouched. 

 According to Chris Padwick, director of computer vision and machine learning at Blue River Technology, PyTorch was used to train its See & Spray robotic farming system owing to its flexibility and the fact that it was easy to debug. To train it, the machine learning system receives a giant variety of images featuring cotton plants, to identify good plants, and then a huge variety of pictures of pigweed, to identify bad plants. After training, the computer vision system can then identify any plant it sees and rates its confidence that it’s a crop or a weed.