A Practical Application of Unsupervised Machine Learning for Analyzing Plant Image Data Collected Using Unmanned Aircraft Systems

Abstract

Unmanned aircraft systems are increasingly used in data-gathering operations for precision agriculture, with compounding benefits. Analytical processing of image data remains a limitation for applications. We implement an unsupervised machine learning technique to efficiently analyze aerial image data, resulting in a robust method for estimating plant phenotypes. We test this implementation in three settings: rice fields, a plant nursery, and row crops of grain sorghum and soybeans. We find that unsupervised subpopulation description facilitates accurate plant phenotype estimation without requiring supervised classification approaches such as construction of reference data subsets using geographic positioning systems. Specifically, we apply finite mixture modeling to discern component probability distributions within mixtures, where components correspond to spatial references (for example, the ground) and measurement targets (plants). Major benefits of this approach are its robustness against ground elevational variations at either large or small scale and its proficiency in efficiently returning estimates without requiring in-field operations other than the vehicle overflight. Applications in plant pathosystems where metrics of interest are spectral instead of spatial are a promising future direction.

Keywords

• LARS
• UAS
• UAV
• unsupervised learning
• high-throughput phenotyping