L. Drees: Temporal Prediction and Evaluation of Brassica Growth in the Field using Conditional Generative Adversarial Networks

My name is Lukas Drees, 26 years old, and I have been working for 1 year as a Phd student in the remote sensing group of the University of Bonn under the supervision of Ribana Roscher.
I am employed in the research project PhenoRob where I am engaged in data analysis of agricultural field experiments. In particular I am involved in the analysis of images taken from the field by UAVs or ground robots. Methodically I use machine learning and deep learning techniques with current research focus on multi-modal deep learning and generative models.

Major goals of farmers are the control of pests and plant diseases, and the assessment of the maturity of their crops.
To do this at an preferably early stage, the phenotype of a plant needs to be forecasted into the future and analyzed.
For this, the use of generative models has recently increasingly gained relevance because they have proven their strength especially for changing data representation, data augmentation, and domain adaptation.
In this work, we use a conditional generative adversarial network to predict the future plant stage of cauliflowers based on RGB images.
In order to evaluate the quality of our generated images, we apply a Mask-R-CNN for cauliflower instance segmentation to both the generated and the real images.

The comparison of the cauliflower instances shows that our trained generator is suitable to generate realistic temporal predictions several weeks into the future under different field conditions for all growth phases.