Publication
Image-based BBCH-Stage Estimation of Maize Plants in the Leaf Development Phase using Multi-Task Learning
Justus Bracke; Jana Kombrink-Lübbe; Marcel Storch; Martin Atzmueller
In: 2026 IEEE International Workshop on Metrology for Agriculture and Forestry (MetroAgriFor). IEEE International Workshop on Metrology for Agriculture and Forestry (MetroAgriFor), November 9-11, Potsdam, Germany, IEEE Xplore, 2026.
Abstract
The standardized BBCH-scale describes the growth stage of a plant and is an important parameter in agricultural practice and plant breeding. Various computer vision-based approaches for an automated estimation of the BBCH- or similar growth describing scale-stages using image data were investigated in the past for different crop types. The comparability of the approaches especially for maize plants is difficult as there are almost no appropriate public datasets available. To address this data scarcity, we present an extension of the RocasMaize image dataset with a new layer of plant-specific BBCH-stage annotations of maize plants in the leaf development stage. The annotations are provided for 461 images, each containing five to seven maize plants in the BBCH-stages 10 to 17. Furthermore, based on the dataset we investigated the potentials of a new multi-task learning based approach for the plant-specific BBCH-stage estimation and compared it to a recently published instance segmentation based approach. The results show for the multi-task learning based approach on average over all BBCH-stages an F1-score improvement of 4.35%. The annotated dataset and the code scripts of the experiments are publicly available.