Automated Detection of the invasive Chinee Apple and Water Hyacinth via UAV Imagery

Abstract

Weeds are detrimental to agriculture and open pastures, competing with crops and native vegetation for essential resources such as soil moisture, sunlight, and space. Weeds in Parkland disrupt native ecosystems, reduce pasture productivity, hinder stock management, and limit livestock access to water. Similarly, weeds in waterways degrade water quality, affect native aquatic life, and block infrastructure. For improved management of pests in general, including weeds, increasing interest in artificial intelligence (AI), particularly in deep learning (DL), has led to research on automated weed detection and mapping using Unmanned Aerial Vehicles (UAVs) imagery. However, major challenges in the domain adaptation of DL models for automated weed detection are the lack of reliable datasets for specific weed types and the limited groundwork for the initial training of DL models to detect such weeds.
To overcome these hurdles, this study presented here-in first invests in creating an unmanned aerial vehicle (UAV) imagery dataset of Chinese apple (Ziziphus mauritiana), a significant woody tree that has widely infested grazing lands of North Queensland, Australia, resulting in a reduction of carrying capacity. Furthermore, the study leverages a publicly available UAV dataset of another weed, water hyacinth (Eichhornia crassipes), to lay the groundwork for a DL segmentation model. Water hyacinth is a significant weed of Australian waterways and of great concern worldwide. Secondly, the study trains and validates trending YOLO (v11 and v26), a popular real-time instance segmentation DL model on the Chinese apple and water hyacinth UAV datasets. The work carried out in this study lays the groundwork for the automatic detection and mapping of specific weeds in Australia, creating the potential for future studies.