Bionote

Andreas Hund is senior scientist in the group of Crop Science at ETH Zurich, Switzerland. The group aims to develop new field phenotyping and modelling techniques. Andreas Hund employs these techniques in the context of physiological wheat breeding. He is active in larger networks, such as EMPHASIS, global wheat (www.global-wheat.com) or the alliance for wheat adaptation to heat and drought (AHEAD) of the Wheat Initiative. These activities aim to capitalize on large and connected datasets and the improvement of phenotyping technologies towards a better understanding of crop adaptation to stresses.
Presentation Abstract
Computer vision is increasingly used in daily life and will certainly be a game changer in agriculture. In the public, computer vision and digitalization in agriculture is probably known best as smart farming. For example, the technology aims to improve the efficiency of crop production while reducing its negative impacts on the environment. Less well-known are high-throughput phenotyping technologies to improve data collection in field experiments. Such methods may target the screening of hundreds or even thousands of genotypes. Drones are still the most widely adopted carrier system in this community, but there is a trend towards field robots. These ground-based vehicles enable proximal imaging to quantify traits at an organ scale rather than the plant or canopy level, as is the case for most drone imaging setups. The most important game-changers for such phenotyping approaches are the constantly improving deep learning techniques. They offer unpreceded precision in extracting targeted traits from images of complex canopies. Starting from segmented anatomical features, such as leaves, flowers, or fruits, there is a wide range of new possibilities to quantify the effect of stresses on yield formation and product quality. I will showcase some of these possibilities based on our own work and discuss how to drive this development further. Robust feature extraction across a wide range of field conditions requires sufficiently large training data. I will use the example of the global wheat datasets (www.global-wheat.com) to demonstrate the benefit of such diverse training sets. As data to train such models are currently extremely sparse, I will discuss how the community may advance to collect sufficiently large data for feature extraction in a wide variety of plants.