Estimating sugarcane yield and its components using unoccupied aerial systems (UAS)-based high throughput phenotyping (HTP)

Yield and its components are the important traits for plant breeders to select the best genotypes in the breeding programs. However, traditional measurements of these traits across genotypes and environments are labor-intensive and time-consuming, as hundreds or even thousands of plots need to be estimated. A yield trial was carried out using seven sugarcane cultivars planted in a randomized complete block design with four replications for two ratoon crops to estimate sugarcane yield and its components using unoccupied aerial systems (UAS)-based high throughput phenotyping (HTP) and to compare the traditional method with UAS-based yield components in discriminating ability to assess sugarcane yield via a path coefficient analysis. UAS platforms mounted with sensors were flown over the trial. The result shows that UAS-derived plant height (PH) showed a strong relationship with the ground measured PH (R² = 0.89, RMSE = 0.15 m). Likewise, an accurate millable stalk height (MSH) estimation, using UAS-derived PH as a predictor, was observed (R² = 0.54, RMSE = 0.15 m). Canopy height model (CHM)-derived canopy cover (CC) appeared to be a promising feature to indirectly select or to predict for stalk number (SN) (R² = 0.69, RMSE = 10,975 stalks ha⁻¹). Based on a path coefficient analysis, UAS-based yield components performed equally to or slightly underperformed the traditional method. Traditionally, SN was the largest contributor to cane yield. Similarly, CC and CHM were the important components for UAS-based yield components. Additionally, the yield prediction model using UAS-derived canopy features with five cross validation schemes (CVs) revealed that model accuracy increased as association between predictor variables with a responding variable increased. The present study shows that random forest outperformed (higher r and lower RMSE) the linear regression models (stepwise, lasso, and ridge) in all CVs. The linear regressions were off when they were used to predict the performance of cultivars in untested crop/environments (CVs2 and CVs5), while a higher accuracy was observed when using random forest in those CVs. More importantly, the accuracy of all models reduced when they were tested in untested crop/environments (CVs2 and CVs5), indicating the challenge of using a prediction model applied to new environments.