Crop vs weed discrimination dataset


These dataset were collected by the authors during multiple sessions at fields in Lincolnshire, UK. The dataset was acquired using a two-camera setup with RGB and NIR cameras (Teledyne DALSA Genie Nano) mounted 5 cm apart and deployed on a manually pulled cart over the bed of onion plants. The resulting images are of high resolution (2464×2056), however, since the optical centres of the RGB and NIR sensors were misaligned, and to create the dataset images, an additional step was required. The resulting aligned RGBN images are therefore smaller (Carrots – 2428×1985; Onions – 2419×1986) and can exhibit parallax errors, especially with high and mature plants.

The camera was mounted at a height of around 100 cm above the ground, which together with the employed optics resulted in 2.5 px/mm resolution covering a 100 x 85 cm patch of ground. Additionally, we did not use any lighting control mechanisms and relied only on the automated gain control built into the cameras to alleviate uneven lighting conditions.

The images were annotated by first segmenting the vegetation regions shown and annotating those regions, followed by applying pixel-based corrections to the annotations. We provide both types of ground truth, the pixel-precision ground truth providing a class for every pixel, and partial ground truth where the annotations are partially incomplete (region of mixed vegetation marked as a separate class) and the region boundaries imprecise.


Each of the datasets provides the following:

  • RGB and NIR images, registered and cut to the same image size
  • the NDVI image calculated from the NIR image and the red channel of the RGB image (aligned)
  • pixel-precision ground truth for each of the images (each pixel annotated as crop, weed or soil/non-vegetation)
  • partial ground truth for each of images (vegetation annotated per-region, regions of mixed vegetation marked as a separate class)




We provide:

  • Carrots 2017 dataset, collected 16/06/2017 in North Scarle (20 images)
  • Onions 2017 dataset, collected 11/04/2017 in South Scarle (20 images)


If you are considering using this data, please reference the following:

Petra Bosilj, Erchan Aptoula, Tom Duckett, and Grzegorz Cielniak: “Transfer learning between crop types for semantic segmentation of crops versus weeds in precision agriculture”, Journal of Field Robotics (2019)
    author = {Bosilj, Petra and Aptoula, Erchan and Duckett, Tom and Cielniak, Grzegorz},
    title = {Transfer learning between crop types for semantic segmentation of crops versus weeds in precision agriculture},
    journal = {Journal of Field Robotics},
    year = 2019,
    volume = "to be determined (published online)"

Recording platform:


Carrots 2017

Onions 2017

*If you want more information about the dataset, or are thinking of using these, please consider sending us an email (Petra Bosilj) telling a little more about your ideas.


This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported License.

Copyright (c) 2017 Petra Bosilj, Grzegorz Cielniak, and Tom Duckett.


This project was partially funded by BBSRC grant BB/P004911/1.