Nathan's DIY Weed-Killing Drone Takes Homebrew Agricultural Automation to New Heights

Mononymous maker Nathan, of YouTube channel NathanBuildsDIY, is aiming to make agriculture a more autonomous process — with an artificially intelligent drone that can spot weeds and give them a highly-directed dose of weedkiller, without having to flood the whole field with chemicals.

"The concepts are simple," Nathan explains by way of introduction to the project. "Use AI to spray only the weeds that are in a field. That reduces the chemicals that are applied when you compare it to spraying the entire field, which is what's often done today. That means cleaner food. Second, use automation to reduce the labor that's put into a field."

That latter feature is a big part of Nathan's second-generation drone, which builds on successes with a more manual first-generation experiment. The revised design is built with automation in mind — using a clever battery replacement system that can accept a new battery and eject the old one into a charging station without human intervention, effectively allowing it to operate constantly without tying up anyone's time.

In operation, the drone works like its predecessor: the drone flies on a pre-planned path over a field of crops, capturing images and using a machine learning model to classify the plant life below as either a desired food or an invasive weed. When a weed is spotted, a sprayer is activated to douse the weed — and as little of the surrounding crops as possible — in an herbicide.

The 3D-printed drone is powered by a Raspberry Pi Zero 2 W single-board computer, connected to four electronic speed controllers driving four motors in a traditional quadcopter layout. There's a low-cost LED-based "lidar" distance sensor and Raspberry Pi's Global Shutter Camera Module — well-suited to machine learning work, as it captures the whole image at once and avoids distortion caused by rolling-shutter capture on fast-moving subjects.

Surprisingly, the machine learning part of the project runs on the Raspberry Pi Zero 2 W itself using TensorFlow Lite — and while there's a radio link to a more powerful laptop, this acts only as a mission planning station. The autonomous ground station, meanwhile, is under the control of an Arduino UNO — linked via USB to the mission-planning laptop.

The project is detailed in the video embedded above and on Nathan's YouTube channel; source code, printable STL files, and a bill of materials have been published to GitHub under an unspecified license.