Optimization of UAS‐based high‐throughput phenotyping to estimate plant health and grain yield in sorghum

• Published in: Plant phenome journal Vol. 3; no. 1
• Main Authors: Galli, Giovanni, Horne, David W., Collins, S. Delroy, Jung, Jinha, Chang, Anjin, Fritsche‐Neto, Roberto, Rooney, William L.
• Format: Journal Article
• Language: English
• Published: 2020
• ISSN: 2578-2703; 2578-2703
• DOI: 10.1002/ppj2.20010

High‐throughput phenotyping (HTP) has enabled the acquisition of vast amounts of data. Therefore, finding the most informative phenological stage(s) and high‐throughput traits could lead to significant optimization of HTP‐assisted selection. An investigation as to when phenotypic data should be collected and how it should be processed from unmanned aerial system (UAS) imagery for the optimization and assessment of two primary traits in grain sorghum [Sorghum bicolor (L). Moench], namely, grain yield and plant health (based on anthracnose scores) was conducted. By evaluating multiple flight dates across the growing season via multispectral UAS‐based imagery, a set of scenarios composed of combinations of flight dates and vegetation indices were constructed for analysis. In this sense, results showed no increase in predictive ability when combining multiple vegetation indices. Hence, using only an index with a higher predictive ability (e.g., normalized difference vegetation index (NDVI) or modified simple ratio (MSR) for plant health with 0.75; and any tested index but chlorophyll index (CIg) for grain yield with ∼0.55) is recommended. Likewise, the combining of multiple flights did not result in a significant increase in predictive ability for either primary trait. Thus, we observed that a single flight for each trait (e.g., 121 d after sowing with 0.81 for plant health; 104 d after sowing with 0.59 for grain yield) was optimal. Concerning, the predictive algorithms examined, partial least squares regression (PLSR) and neural network, results were similar, with PLSR generally outperforming. In addition, we discuss our findings from an application standpoint of a field‐based breeding program and suggest additional optimization options.