Automated weed segmentation with knowledge based labeling for machine learning applications

Accurate classification of landscape features is essential for precision agriculture, supporting targeted practices such as weed control and variable-rate applications. Although machine and deep learning models show strong promise for real-time weed detection, they require large labelled datasets, which are costly and time-consuming to produce. This study develops and evaluates an automated feature-labelling workflow using eCognition (v9.5) for Unmanned Aerial Vehicle (UAV) RGB imagery. The workflow was tested on a ~ 2000 m² research field at the University of Saskatchewan using high-resolution imagery (0.088 cm). The field contained strips of kochia, wild oat, wild mustard, and false cleavers seeded between wheat rows. The workflow combines multiple spatial algorithms, including segmentation, line detection, distance mapping, convolution filtering, morphological operations, and thresholding. Vegetation indices such as the Colour Index of Vegetation and Excess Green Index effectively separated crops and weeds from soil. Using randomly distributed labelling points and a confusion matrix, the workflow achieved 87% overall accuracy (kappa = 0.81) without manual training labels. This automated workflow demonstrates strong potential for accelerating dataset creation for machine learning and deep learning applications, reducing manual effort while maintaining accuracy. Future work will focus on improving its transferability across fields, dates, and experimental conditions.