“Plant Root Analysis with Multiple Instance Learning”
In order to understand how to increase crop yields, breed drought tolerant plants, understand the relationship between root architecture and soil organic matter, and understand the role roots can play in greenhouse gas mitigation by enhancing the sequestration of carbon in soil, we need to be able to study plant root systems effectively. However, we are lacking high-throughput, high-quality sensors, instruments and techniques for plant root analysis. Techniques available for analyzing root systems in field conditions are generally very labor intensive, allow for the collection of only a limited amount of data and are often destructive to the plant. Once root data and imagery have been collected using current root imaging technology, analysis is often further hampered by the challenges associated with generating accurate training data.
Most supervised machine learning algorithms assume that each training data point is paired with an accurate training label. Obtaining accurate training label information is often time consuming and expensive, making it infeasible for large plant root image data sets. Furthermore, human annotators may be inconsistent when labeling a data set, providing inherently imprecise label information. Given this, often one has access only to inaccurately labeled training data. To overcome the lack of accurately labeled training, an approach that can learn from uncertain training labels, such as Multiple Instance Learning (MIL) methods, is required. In this talk, I will discuss our team’s approaches to characterizing and understanding plant roots using methods that focus on alleviating the labor intensive, expensive and time consuming aspects of algorithm training and testing.
Alina Zare is an Associate Professor and conducts research and teaches in the area of pattern recognition and machine learning in the Electrical and Computer Engineering Department at the University of Florida. Dr. Zare develops algorithms for automated analysis of large data sets from a variety of (usually, non-visual) sensors including multi- and hyperspectral imagery, LiDAR, ground penetrating radar, mini-rhizotron imagery, synthetic aperture sonar, wide band electromagnetic induction data, synthetic aperture radar, and others. Her current research work includes applications in plant phenotyping, plot root imaging and analysis, forestry, remote sensing, landmine and explosive hazard detection, sub-pixel target characterization and detection, and underwater scene understanding. Website: https://faculty.eng.ufl.edu/alina-zare/