Contrasted-Fertilization Wheat Ear Dataset 2020

Overview

This dataset contains 701 wheat RGB images in which all the ears have been manually labelled with bounding boxes. Its primary function is to provide training or test data for deep learning models aiming at the detection of wheat ears. In terms of diversity, it integrates wheat images from two varieties, at all the key development stages from heading to maturity, and for four or eight contrasted nitrogen fertilization management. 

Field experiments and image acquisition

Images were acquired during the 2020 season in two trial fields located in the Hesbaye area, Belgium. The images were captured by two RGB cameras in nadir position. Those cameras were positioned on the cantilever beam to avoid shadows from the phenotyping platform. The auto-exposure algorithm was tuned to prevent, as possible, image saturation. All the details regarding the field experiments and the image acquisition can be found in the related paper: "Dandrifosse S., Ennadifi E., Carlier A., Gosselin B., Dumont B. & Mercatoris B., 2022. Deep learning for wheat ear segmentation and ear density measurement : From heading to maturity. Comput. Electron. Agric. 199(June), DOI:10.1016/j.compag.2022.107161."

Image pre-processing

2048 x 2560 pixel images were acquired in the field, but each of them was converted to four square sub-images of 1024 x 1024 pixels. The labelled images in this dataset are the sub-images.

Ear bounding boxes

The dataset contains a total of 77657 bounding boxes stored in csv files as Python-style lists of [xmin, ymin, width, height]