Automated Weed Segmentation: A Knowledge- Based Approach to Support Machine Learning Training

Read the full article See related articles

Listed in

This article is not in any list yet, why not save it to one of your lists.
Log in to save this article

Abstract

Accurate landscape feature classification is a critical component of precision agriculture, enabling targeted on-farm management practices such as weed control and variable rate applications. Machine and deep learning models, including Convolutional Neural Networks (CNNs) and Random Forests (RF), have shown promise for real-time applications like weed detection. However, a major bottleneck remains: the generation of large, representative labeled datasets required to train these models, especially deep learning algorithms, is both time-consuming and labor-intensive. This study presents and evaluates an automated feature-labeling workflow developed using eCognition software (version 9.5) for Unmanned Aerial Vehicle (UAV). The workflow was tested on a ~ 2000 m² research field at the University of Saskatchewan, Canada, using high-resolution UAV imagery (0.88 mm spatial resolution). The field included strips of kochia, wild oat, wild mustard, and false cleavers seeded between wheat rows (30.5 cm spacing). The workflow integrated a series of spatial algorithms - including image segmentation, line detection, distance mapping, convolution filtering, morphological filters, local extrema detection, and image thresholding. Key inputs included the Color Index of Vegetation and Excess Green Index, which were effective in distinguishing green vegetation (crops and weeds) from the soil background. Using randomly distributed labeling points and a confusion matrix for accuracy assessment, the workflow achieved an overall accuracy of 87% (kappa = 0.81), even under a scenario without manually provided training samples. The automated workflow presented in this paper offers the potential for automated image labeling or sample collection for image classification in the domains of machine or deep learning. The workflow would greatly decrease the time and labour resources needed to collect such extensive labels for model training and validation. Future work should aim to enhance the workflow towards the generalization of the algorithms’ parameters and for use with multiple date/field imagery, thus ensuring the transferability of the workflow to other agronomic experiments.

Article activity feed