Autoweed: detecting Harrisia cactus in the Goondiwindi region for selective spot-spraying

Abstract

Innovations in the field of robotics have sought to combine technology with farming to provide efficient solutions for the ever-present problem of invasive weeds. Research has focused on the development of robotic solutions for use in croplands, while rangeland weed management techniques have remained largely unchanged. In response, we developed a prototype robotic spot-sprayer for pass over foliar spraying in Australian rangelands that we call AutoWeed. In partnership with North West Local Land Services, we take aim at Harrisia cactus (Eriocereus martini)in the Goondiwindi region to determine its suitability as a potential target. We collected an image dataset of Harrisia cactus comprising over 2,000 images of target and non-target plant life in the Goondiwindi region. Using state-of-the-art convolutional neural networks (CNNs) to detect the Harrisia cactus, we achieved an average classification accuracy of 98.14%, which is an encouraging result. To understand the model's learning, class activation maps were utilised to generate heatmaps that identify regions of interest. It was discovered that the CNN models could correctly discriminate between Harrisia cactus and its surroundings despite the highly variable nature of the environment. These results prove the target suitability of Harrisia cactus for future field trials of the AutoWeed spot-sprayer robot.