UAV-derived plot lengths to enhance the accuracy of sugarcane breeding trials

Improving accuracy of field trials is fundamental for improving the selection accuracy of target traits and accelerating genetic gain rates. Sugarcane breeding field trials are commonly affected by inconsistencies, due to factors such as gaps in experimental plots caused by planting and management issues. These inconsistencies are difficult to quantify from the ground, and if left unaccounted, they can cause inflated error variances and reduced selection accuracy of target traits. Imagery from Unmanned Aerial Vehicles (UAV) and machine-learning algorithms were investigated for the accurate and high throughput detection of gaps and measurement of exact plot length. An analysis pipeline for streamlined detection of plot gaps and length from UAV images was developed using open-source packages and implemented as a browser-based application. UAV imagery was obtained from a single-row (early stage) field trial in the Burdekin region. Machine-learning algorithms were developed for image segmentation and extraction of plot length. The length of each experimental plot was used as a covariate for analysis of cane yield. Best fit of modelling, differences in estimated cane yield broad-sense heritability, and correlation with corresponding four-row (late-stage) cane yield were compared based on the results from statistical models with and without plot length as a covariate. In all scenarios, the model with plot length as a covariate produced a better fit for cane yield than the model without a covariate. Estimates of broad-sense heritability for cane yield improved when plot length was used as a covariate in most scenarios. Furthermore, cane yield corrected with plot length as covariate showed a slightly stronger correlation with cane yield from the corresponding four-row field trial. Incorporating UAV-acquired spatial covariates such as plot length into statistical models can improve trial precision and selection accuracy of target traits.