RoWeeder: Unsupervised Weed Mapping through Crop-Row Detection

This is official implementation for our CVPPA 2024 paper.

RoWeeder is an innovative framework for unsupervised weed mapping that combines crop-row detection with a noise-resilient deep learning model. By leveraging crop-row information to create a pseudo-ground truth, our method trains a lightweight deep learning model capable of distinguishing between crops and weeds, even in the presence of noisy data.

Evaluated on the WeedMap dataset, RoWeeder achieves an F1 score of 75.3, outperforming several baselines. Comprehensive ablation studies further validated the model's performance.

By integrating RoWeeder with drone technology, farmers can conduct real-time aerial surveys, enabling precise weed management across large fields.

Installation

Prepare the environment

conda create -n SSLWeedMap python=3.11
conda activate SSLWeedMap
# Install from environment.yml
conda env update --file environment.yml

Preprocessing

For each field (000, 001, 002, 003, 004)

Download and extract the dataset

wget http://robotics.ethz.ch/~asl-datasets/2018-weedMap-dataset-release/Orthomosaic/RedEdge.zip -d dataset/
unzip dataset/RedEdge.zip -d dataset/

Rotate the images

python3 main.py rotate --root dataset/RedEdge/000 --outdir dataset/rotated_ortho/000 --angle -46 &
python3 main.py rotate --root dataset/RedEdge/001 --outdir dataset/rotated_ortho/001 --angle -48 &
python3 main.py rotate --root dataset/RedEdge/002 --outdir dataset/rotated_ortho/002 --angle -48 &
python3 main.py rotate --root dataset/RedEdge/003 --outdir dataset/rotated_ortho/003 --angle -48 &
python3 main.py rotate --root dataset/RedEdge/004 --outdir dataset/rotated_ortho/004 --angle -48

Patchify the images

python3 main.py patchify --root dataset/rotated_ortho/000 --outdir dataset/patches/512/000 --patch_size 512 &
python3 main.py patchify --root dataset/rotated_ortho/001 --outdir dataset/patches/512/001 --patch_size 512 &
python3 main.py patchify --root dataset/rotated_ortho/002 --outdir dataset/patches/512/002 --patch_size 512 &
python3 main.py patchify --root dataset/rotated_ortho/003 --outdir dataset/patches/512/003 --patch_size 512 &
python3 main.py patchify --root dataset/rotated_ortho/004 --outdir dataset/patches/512/004 --patch_size 512

Generate the pseudo GT

python3 main.py label --outdir dataset/generated --parameters parameters/row_detect/69023956.yaml

Train the RoWeeder Flat model

python3 main.py experiment --parameters=parameters/folds/flat.yaml

Get the model for inference from HuggingFace hub

from roweeder.models import RoWeederFlat
model = RoWeederFlat.from_pretrained("pasqualedem/roweeder_flat_512x512")

Citation

If you find this work useful, please consider citing our paper (in press):

@inproceedings{roweeder,
  title={RoWeeder: Unsupervised Weed Mapping through Crop-Row Detection},
  author={Pasquale De Marinis, Gennaro Vessio, Giovanna Castellano},
  booktitle = {Proceedings of the IEEE/CVF European Conference on Computer Vision (ECCV) Workshops},
  year={2024}
  note={in press}
}

pasqualedem/RoWeeder