UAV-Based Multispectral Imagery in Crop Breeding: Cicer Milkvetch forage yield prediction as an example

Manual data collection in forage crop breeding is highly labor-intensive and becomes an operational bottleneck. Application of unmanned aerial vehicles (UAVs) offers a solution to modernize field data collection. This study evaluates the effectiveness of predicting individual plant biomass yield in cicer milkvetch (Astragalus cicer L.) from multispectral imagery from UAVs with field agronomic data. A space-planted nursery containing over 200 half-sib families, including commercial cultivars AC Oxley II and AC Veldt, was established in 2023 at the Livestock and Forage Centre of Excellence (LFCE) of the University of Saskatchewan near Saskatoon, SK, Canada. The experimental layout followed a randomized complete block design with three replications and a 1-meter plant spacing. UAV flights were conducted 3 times from 1 May through 30 June 2024 to collect vegetation indices, including the Normalized Difference Vegetation Index (NDVI) and Excess Green Index (ExG), and canopy area measurements. Forage dry matter yield (DMY), spring vigor score and plant height were utilized as ground-truth data to validate these UAV-derived datasets. A Random Forest (RF) model was used to predict DMY. Hyperparameter tuning significantly optimized model prediction, and increased the coefficient of determination from R² = 0.50 to R² = 0.71. Canopy area and vegetation indices (NDVI and ExG) collected immediately prior to biomass harvest were the most accurate indicators. In contrast, initial NDVI was insignificant in predicting DMY. These outcomes demonstrate that UAV-based phenotyping provided fast and acceptable predictive capabilities for forage crop breeding; however, the timing of flights seemed significant in improving predicting accuracy.