Dockweed (Rumex spp.) Detection — YOLO26

Two YOLO26 models for detecting dockweed (Rumex obtusifolius) in RGB images: an object-detection model producing bounding boxes, and an instance-segmentation model producing pixel-level masks. Both are intended for precision-agriculture use cases such as weed mapping, spot-spraying, and robotic/mechanical weeding.

Model summary

Model file	Task	Output	Base checkpoint
`Rumexv8.pt`	Object detection	Bounding boxes	`yolo26m.pt` (YOLO26-medium)
`Rumexv8-seg.pt`	Instance segmentation	Bounding boxes + pixel masks	`yolo26m-seg.pt` (YOLO26-medium)

Despite the "v8" in the file names, both models are fine-tuned from Ultralytics YOLO26 (medium size) checkpoints, not YOLOv8.

Intended use

Detecting/segmenting dockweed (Rumex obtusifolius) plants in field or grassland imagery for precision agriculture, weed monitoring, and automated weeding systems.
Both models were trained only on Rumex obtusifolius. They have not been validated on Rumex crispus (curled dock) or other dock/Rumex species, and may not generalize well to them.
Not intended for use outside agricultural weed-detection contexts, and not validated for safety-critical or regulatory decision-making.

Training data

Own dataset: 250 images collected and manually annotated by the model author.
Public dataset: additional images of Rumex obtusifolius (only this species, Rumex crispus excluded) drawn from a public online source.
A subset of the combined dataset is published at JacobsFarmextra/Dockweed on Hugging Face.
Split: 2,610 images in the training set, 871 images in the validation set (same dataset.yaml used for both models).

See training_log.txt for the original dataset-preparation log.

Training procedure

Both models were fine-tuned from their respective YOLO26-medium checkpoints using Ultralytics with identical hyperparameters and augmentation settings, image size 640, for up to 100 epochs with early stopping (patience 20).

Setting	Value
Base checkpoint	`yolo26m-seg.pt` (seg) / `yolo26m.pt` (det)
Image size	640
Batch size	8
Epochs	100 (max)
Patience (early stopping)	20
Checkpoint save period	every 20 epochs
Device	single GPU (`device=0`)
Dataloader workers	0
Augmentation	enabled
Rotation (`degrees`)	10°
Translation (`translate`)	0.1 (10%)
Scale (`scale`)	0.5 (0.5–1.5×)
Horizontal flip (`fliplr`)	0.5 (50% chance)
HSV hue (`hsv_h`)	0.015
HSV saturation (`hsv_s`)	0.7
HSV value/brightness (`hsv_v`)	0.4
Mosaic (`mosaic`)	1.0

Full training scripts are included in this repository:

train_seg.py — used to train Rumexv8-seg.pt (this is the original script used, with the local file path anonymized).
train_det.py — used to train Rumexv8.pt (reconstructed from train_seg.py: identical settings, using the non-segmentation yolo26m.pt base checkpoint).

from ultralytics import YOLO

model = YOLO("yolo26m-seg.pt")  # or "yolo26m.pt" for the detection model

model.train(
    data="dataset.yaml",
    imgsz=640,
    batch=8,
    epochs=100,
    patience=20,
    save_period=20,
    workers=0,
    device=0,
    augment=True,
    degrees=10,
    translate=0.1,
    scale=0.5,
    fliplr=0.5,
    hsv_h=0.015,
    hsv_s=0.7,
    hsv_v=0.4,
    mosaic=1.0,
)

How to use

Requires ultralytics:

pip install ultralytics

Object detection (bounding boxes):

from ultralytics import YOLO

model = YOLO("Rumexv8.pt")
results = model.predict("image.jpg")
results[0].show()

Instance segmentation (masks + boxes):

from ultralytics import YOLO

model = YOLO("Rumexv8-seg.pt")
results = model.predict("image.jpg")
results[0].show()

Limitations

Trained on Rumex obtusifolius only — accuracy on Rumex crispus or other look-alike species is unknown and likely lower.
The own-collected portion of the dataset (250 images) is relatively small; performance may vary with lighting, growth stage, background vegetation, and camera/sensor characteristics not well represented in training.
No formal held-out test-set metrics (precision/recall/mAP) are published with this card; users should validate on their own data before deployment.
Not evaluated for fairness/bias beyond species coverage, and not intended for any use outside weed detection in agricultural imagery.

License

Released under AGPL-3.0, consistent with the Ultralytics YOLO license terms. See the Ultralytics licensing page for details on AGPL-3.0 vs. Enterprise licensing options.

Citation

If you use these models, please cite Ultralytics YOLO:

@software{yolo26_ultralytics,
  author = {Ultralytics},
  title  = {Ultralytics YOLO26},
  url    = {https://github.com/ultralytics/ultralytics}
}

And, if used, the accompanying dataset:

JacobsFarmextra/Dockweed, https://huggingface.co/datasets/JacobsFarmextra/Dockweed

Downloads last month: -

JacobsFarmextra
/

Dockweed