Image-based Weed Detection in Soybean using Machine Learning Algorithms

Loading...
Thumbnail Image

TR Number

Date

2026-06-11

Journal Title

Journal ISSN

Volume Title

Publisher

Virginia Tech

Abstract

Weed management remains a major challenge in soybean production because weeds compete with crops for light, nutrients, water, and space, often causing substantial yield losses. Problematic grass and broadleaf species in soybean production systems such as Texas panicum, common lambsquarters, Palmer amaranth, and common ragweed require timely detection and management planning. Manual scouting is time-consuming, labor intensive, and spatially coarse, limiting its usefulness for site-specific weed management. Recent advances in unmanned aerial vehicles (UAVs), ground-based sensing, deep learning, and web-based geospatial tools offer new opportunities to automate weed detection, quantify spatial patterns, estimate yield impacts, and support precision weed management. The current research presents a series of studies that develop and evaluate deep learning– based weed detection and decision-support systems for soybean production. First, a high resolution UAV dataset of 10,442 annotated images of soybean and Texas panicum was used to systematically assess modern You Only Look Once (YOLO) architectures (YOLOv8–YOLOv11; nano, small, and medium variants) with 10-fold cross-validation. All models achieved mAP@50 values above 85%, with YOLOv9-small and YOLOv11- medium reaching 88.4%, while YOLOv8-nano provided the fastest inference (6.5 ms per image on an NVIDIA® V100 GPU) with less than a 1% reduction in mAP@50, indicating strong potential for real-time deployment. Second, species-specific models for common lambsquarters were developed using 5,218 UAV images and 5,538 ground-based images across growth stages and field conditions. YOLOv9–YOLOv12 models achieved mAP@50 values above 80% on both platforms, with YOLOv9-tiny performing best. The relationship between lambsquarters density and soybean yield followed a quadratic pattern (R² = 0.85; RMSE =282.244) and weed counts from YOLOv9-tiny detections were strongly associated with yield at 35 days after planting but weaker at 50 days, likely due to canopy occlusion. Self-supervised based yield prediction models gave strong performance, the combined SSL model, trained using images from both 35 and 50 DAP, produced the strongest overall performance. Using the 512-dimensional SSL-extracted image features as inputs to a ridge regression model, the approach achieved R² values of 0.72 and 0.63 for yield prediction on 35 and 50 DAP images, respectively. Thirdly, a web-based platform (WeedGeoLocator) integrating UAV RGB orthomosaics with a YOLOv8-nano model trained on 23,557 annotated images of common ragweed, Palmer amaranth, and Texas panicum was developed to generate georeferenced weed maps and summaries through an accessible client–server interface, advancing practical tools for precision weed management. Overall, this research demonstrates the potential of integrating advanced technologies like drone imagery, ground-based sensing, deep learning, self-supervised learning, and web-based mapping tools to advance weed detection and decision-making in soybean production. These technologies provide a foundation for precision weed management systems.

Description

Keywords

precision agriculture; UAV imagery; soybean weed detection; YOLO; deep learning; Texas panicum; common lambsquarters; decision support tools

Citation

Collections