---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': grass
          '1': white_clover
          '2': wild_lettuce
  - name: frame_type
    dtype: string
  - name: split
    dtype: string
  splits:
  - name: train
    num_bytes: 19485356
    num_examples: 5005
  download_size: 19428341
  dataset_size: 19485356
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: mit
task_categories:
- image-classification
size_categories:
- 1K<n<10K
---
# Grassfield Weed Classification

Grass-field weed patches released with a two-stage weed identification method. Sub-frames cut from images of grass fields are labelled as grass, irregular weed (white clover, Trifolium repens) or regular weed (wild lettuce, Lactuca virosa); the published split has 4,776 sub-frames (3,182 grass, 1,193 clover, 401 lettuce). Larger region frames (229) carry the same three classes and are included too, marked by a frame_type column, with the published train/test split. Images were taken with two cameras under sunny, cloudy and shaded conditions at different locations. The dataset contains 5,005 images across 3 classes: grass, white_clover, wild_lettuce.  
Images per class:
- grass: 3,295
- white_clover: 1,275
- wild_lettuce: 435

This dataset is indexed on https://project-agml.github.io/ as part of the AgML python library.

The original train/test/val split has been preserved in the `split` column.

## Citation

```bibtex
@article{calderaracea2024two,
  title={A two-stage deep learning strategy for weed identification in grassfields},
  author={Calderara-Cea, Felipe and Torres-Torriti, Miguel and Auat Cheein, Fernando and Delpiano, José},
  journal={Computers and Electronics in Agriculture},
  volume={225},
  pages={109300},
  year={2024},
  publisher={Elsevier}
}
```


*This dataset was reformatted from its original format to match HuggingFace standards.*