Who loves to weed their lawn?

This scalable environmental robotic project takes these budding student engineers through the research and development process, showing them how to take ideas and bring them to life in the real world.  The project is multistage, so there is plenty of work for future greener students. This blog focuses on the weed identification component of the system.

[Footnote: The Project Sponsor did stipulate this was a supervised industrial placement project, as the state of the art in this field is far more advanced than our ‘Envirobot’ system].

The broader scope of the project was to build an autonomous lawn care robot that identifies weeds in a lawn and applies a minimal and targeted dose of herbicide directly onto each weed. The goal is to minimise herbicide use, reducing costs and environmental impacts, and reducing human exposure to herbicides from manual application. The principle can be applied to crops, bushland regeneration and remediation of mine sites.

One aspect of this project was using “Transfer Learning” to develop our neural network. We start with a pre-trained generalised network, we add extra layers, and then train only these extra layers. Think of it like learning to drive a ride-on mower and then wanting to drive a tractor. You don’t need to re-learn everything about driving; you just need to learn the specifics of driving a tractor, like the more significant turning circle, slower breaking, and where the gear stick is located. This is brilliant because training neural networks is computationally intensive, so it involves significant time and energy savings.

So, back to the weed identification component of the project. The students had to train this neural network, a type of machine learning model, as follows.

Step 1. Make a camera cart and software to collect hundreds of field images of lawns and weeds, focusing on a single broad-leaf weed called plantain.

Step 2. Mark the location of this weed in every image. As it takes many images to train a neural network, the augmentation technique was applied to increase the number of images systematically, for example, rotating a single image to make multiple training images.

Step 3. Splitting the data into a training set and an independent validation set.

Step 4. They trained the network on the in-house GPU server using only the training set.

Step 5. The network was tested on the validation set, and the weed prediction was valid 78% of the time.

The Industrial Placement program at ResTech offers final-year undergraduate engineers an opportunity to gain experience on challenging research projects in a commercial setting.

There is a lot of hype about AI, and this project was a fun way for both students and staff to separate the wheat from the chaff and explore the capabilities of this technology in a fertile field!

 For the Semantic Scholar subscriber, the abstract can be found here