Deep Learning for Large Scale Plant Phenotyping

Abstract

Crop breeding needs phenotyping to assess the physical and biochemical characteristics of plants and identify desirable traits for selection. Traditional techniques for evaluating plant traits involve manual measurement of properties such as plant height, leaf length, and biomass in the field. Imaging approaches offer an alternative to manual measurements to monitor plant growth and health in large-scale settings, but finding the best features to measure is a complicated question. In this thesis, we explore a data-driven approach to discovering phenotyping features by incorporating deep learning in the field of plant phenotyping. We demonstrate that a metric learning approach uses a large number of images and limited label data to automatically learn features that distinguish different cultivars. We show that these features can be used to recognize unseen cultivar varieties, identify genetic variability, and detect visible characteristics of drought stress. We also demonstrate that this approach can be extended to multiple sensor modalities and explore novel generative approaches to highlight very subtle variations in image appearance across classes. Finally, we publish a comprehensive dataset of sorghum imagery along with structured evaluation protocols to support future research in this domain.