Automated Weed Detection in Red Beet (Beta vulgaris L., Conditiva Group, cv. Kestrel F1) Using Deep Learning Models

Weed competition in red beet (Beta vulgaris L. Conditiva Group) directly reduces crop yield and quality, making detection and eradication essential. This study proposed a three-phase experimental protocol for multi-class detection (cultivation and six types of weeds) based on RGB (red-green-blue) colour images acquired in a greenhouse, using state-of-the-art deep learning (DL) models (YOLO and RT-DETR family). The objective was to evaluate and optimise performance by identifying the combination of architecture, model scale and input resolution that minimises false negatives (FN) without compromising robust overall performance. The experimental design was conceived as an iterative improvement process, in which each phase refines models, configurations, and selection criteria based on performance from the previous phase. In phase 1, the base models YOLOv9s and RT-DETR-l were compared at 640 × 640 px; in phase 2, the YOLOv8s, YOLOv9s, YOLOv10s, YOLO11s, YOLO12s and RT-DETR-l models were compared at 640 × 640 px and the best ones were selected using the F1 score and the FN rate. In phase 3, the YOLOv9 (s = small, m = medium, c = compact, e = extended) and YOLOv10 (s = small, m = medium, l = large, x = extra-large) families were scaled according to the number of parameters (s/m/c-e/l-x sizes) and resolutions of 1024 × 1024 and 2048 × 2048 px. The best results were achieved with YOLOv9e-2048 (F1: 0.738; mAP@0.5 (mean Average Precision): 0.779; FN: 28.3%) and YOLOv10m-2048 (F1: 0.744; mAP@0.5: 0.775; FN: 27.5%). In conclusion, the three-phase protocol allows for the objective selection of the combination of architecture, scale, and resolution for weed detection in greenhouses. Increasing the resolution and scale of the model consistently reduced FNs, raising the sensitivity of the system without affecting overall performance; this is agronomically relevant because each FN represents an untreated weed.