ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

Multi-species weed detection using supervised (CNN and transformer-based) and semi-supervised learning approaches in corn field

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2600368.(doi:10.13031/aim.202600368)
Authors:   Utsav Bhandari, Saroj Burlakoti, Rhonda Miller, Ph.D., Sierra Young, Ph.D., Eric Westra, Ph.D., Aaron Etienne, Ph.D.
Keywords:   Multi-species weed detection, Precision agriculture, Pseudo-labeling, Semi-supervised object-detection, UAV, YOLO

Abstract. The widespread adoption of site-specific weed management (SSWM) using unmanned aerial vehicles (UAVs) relies on accurate, real-time object detection models. However, the requirement for vast amounts of expert-annotated training data remains a critical bottleneck. This study evaluates the efficacy of supervised and semi-supervised object detection (SSOD) frameworks for multi-species weed detection in a commercial forage corn field. A high-resolution UAV dataset comprising three dominant weed species (common lambsquarters, redroot pigweed, and green foxtail) was generated, and 27 supervised detectors spanning the YOLOv8, v9, v10, 11, 26 and RT-DETR architecture families were systematically benchmarked. Results indicated that architectural scale does not linearly predict accuracy in data-constrained agricultural environments; lightweight architectures remained highly competitive. YOLOv9s was identified as the optimal baseline for edge deployment, achieving an mAP@0.5 of 0.84 and demonstrating superior minority-class detection within a highly constrained 14.5 MB footprint. Subsequently, YOLOv9s was utilized to evaluate six state-of-the-art SSOD frameworks using 8,991 unannotated images. Notably, complex SSOD architectures (e.g., MixPL, Consistent-teacher) experienced significant performance degradation (mAP@0.5:0.95 of 0.34–0.36) on the complex multi-species imagery, whereas foundational consistency-based frameworks (Mean teacher) maintained baseline-level performance (mAP@0.5:0.95 of 0.47). These findings demonstrate the viability of lightweight CNNs for real-time UAV applications. They also reveal critical transferability gaps for advanced SSOD methods in dense, visually complex agricultural environments.

(Download PDF)    (Export to EndNotes)