Opportunities and constraints for image-based weed recognition in large-scale production systems

Abstract

Site-specific weed control (SSWC) is an opportunity to transform the way in which weeds are managed. Yet, as an emerging technology, there are questions around the interactions of image-based weed recognition technologies, which enable SSWC, with weeds within large-scale cropping systems. The primary aim of this thesis was to explore this interdisciplinary domain, investigating the opportunities and constraints towards improved weed recognition posed by this technology-biology interaction. Three chapter-specific aims were formed: (1) develop and evaluate an open-source weed detection device, exploring the opportunity for open-source development in agriculture and plant science; (2) investigate the constraint of plant growth stage on deep-learning based, open-source object detection models; and (3) understand the risk of image-based weed recognition as a novel selection pressure, selecting for crop mimicry in weeds by repeated use of image-based algorithms for weed recognition.

The opportunity presented by open-source development was explored in depth in Chapter 3, through the review of open-source tools in agriculture and plant science (Chapter 3.1), and the successful development and field evaluation of a novel open-source fallow weed detection system (Chapter 3.2). Chapter 4 explored the challenge of multi growth stage recognition, finding significant confusion between visually similar classes and moderate generalising capabilities into unknown, yet visually similar growth stages. In Chapter 5, for the first time, it was shown experimentally that two generations of selection for wheat mimetic annual ryegrass (Lolium rigidum) resulted in heritable (h^2 = 0.43) changes in plant morphology towards crop mimicry.

This thesis provides unique, interdisciplinary insights where biology and image-based weed recognition meet, uncovering opportunities, and highlighting the constraints of working with highly, morphologically diverse, and adaptable weeds.