Real-Time Weed Detection and Removal Using Deep Learning and Laser

Real-time weed detection and removal using deep learning and laser

 

Herbicides have been the main weed controlling technique in agriculture for decades. However, due to environmental, health, and economic concerns on the impact of applying excessive amount of these chemicals, and the fact that weeds are becoming herbicide resistant, alternative approaches to weed control are crucial. Advances in robotics and deep learning technologies provide an opportunity to explore alternative approaches to weed control. Creation of an affordable real-time automated weed detection and removal robotic system using deep learning for detection and laser for removal is one of these alternative approaches being explored.

In this research, the machine vision weed detection models and the weeds’ response to laser treatment are investigated.

Three deep learning models, YOLO-v3, SSD, and Faster-RCNN were trained on thirteen weed species that are common in cotton fields. The process involved collecting weed species images from the field, labelling, performing some preprocessing augmentation to increase the dataset, and then train the models to learn the image features. Testing the models on the test images dataset showed different accuracy and speed, with SSD being the most accurate in detecting the weeds and YOLO-v3 being the fastest.

The two laser experiments were conducted. Experiment 1 involved treating two species of weeds with four different laser powers (1.2W, 1.35W, 4.2W, 4.5W) to determine the time it takes to completely cut the weed stem. Experiment 2 involved monitoring another two species of weeds for a week after treating them with two different lasers at constant application times (1s, 2s, and 3s). The results showed that the diameter of the weed stem, laser power, and the exposure time play a significant role in survival of the weed, with the distance from laser source to the weed (up to 15cm) having no significant impact.