The authors of the WeedMaize dataset conducted their experiments in fields located in Badajoz, Spain (39°1’14.42", -6°3’40.69"). They captured images in April 2020 within maize crops (Zea mays L.M) at V1 and V3 growth stages, without any weed treatment, 10 days prior. Using an “M” trajectory, zenithal images were taken every 2 meters from a height of 1.3 meters. These images featured three weed species (Cyperus rotundus L., Echinochloa crus-galli L., and Solanum nigrum L.) in low-growth stages (Vc and V1 or greater), as well as crop plants (Zea Mays L.). In total, 312 images were manually labeled, including a “Ve” class for unidentified emerging plants. The labeling process involved the use of LabelImg software, and the data was saved in PascalVOC format.

A Canon PowerShot SX540 HS camera with a resolution of 5184 px X 3886 px was used, with a shutter speed of 1/1000 and automatic ISO calibration to maintain image quality under changing lighting conditions. The images represented various configurations, including weeds within and between crop lines, with significant overlap between weeds and maize plants, especially those in the V3 growth stages.

Here is how EPPO codes for plants generally work: codes consist of 5 letters = 3 (genus) + 2 (species). For example, Alopecurus myosuroides -> ALOMY. More information on that here

Plant labels include SOLNI_Vc for Solanum nigrum L. at the embryonic stage, SOLNI for more developed Solanum nigrum L. plants, CYPRO_min for multi-leaf Cyperus rotundus L. plants with center-focused bounding boxes, CYPRO for single-leaf Cyperus rotundus L. plants (>V1), ECHCG for Echinochloa crus galli L., and ZEAMX for Zea mays L. plants.

The dataset was split as follows: 70% of the images formed the train_set, while the remaining 30% comprised the validation_set. Additionally, a random sample of 30% of the training set images constituted the test_set, used to assess training progress. Training and validation (including testing) were performed using a GeForce GTX 1080 GPU, but due to GPU limitations, image sizes had to be reduced. This reduction involved scanning each image to generate 74 smaller sub-images, with parameters set as follows: width 3886 px, height 1926 px, and an overlap of 1900 px. The corresponding XML files defining plant labels were adjusted accordingly for each sub-image.

During scanning, some bounding boxes near the sub-image edges might have been cut, resulting in incomplete data. To ensure data integrity, these boxes were eliminated from training. The chosen large overlap during scanning ensured that if a plant was removed due to touching the edge, it would appear in the next sub-image, ensuring all plants had labels in the training set. After scanning the three image groups (training, test, and validation sets), 4368 training sub-images, 2181 validation sub-images, and 1310 test sub-images were obtained. Sub-images without bounding boxes were removed, and labeling errors were corrected by experts through visual examination. The final data set consists of 7859 sub-images with 60436 bounding boxes.