Predicting herbage yield in perennial ryegrass breeding trials using UAV derived data and machine learning

Several methods have been developed to estimate biomass yield in ryegrass using remotely sensed spectral and structural features. This study builds further upon procedures developed in the breeding program of ILVO. In previous work, we focused on canopy height as the main predictor of yield. Here we investigate whether the prediction of herbage yield in perennial ryegrass can be improved using canopy height information combined with spectral bands captured using different sensors. We used six breeding trials comprising 115 diploid and 112 tetraploid varieties and populations, with a total of 468 plots. A series of UAV flights were carried out with two sensors, a 10-band multispectral and an RGB camera system. The acquired data were then used to estimate the yield of the first spring cut in May 2020. Repeated nested cross-validation allowed us to evaluate the performance of the predictive models. Three machine learning algorithms (Random Forest, Support Vector Machine and Partial Least Squares Regression) were applied, to better understand the applicability of those techniques for accurate yield assessments. This study provides new insights to ryegrass biomass estimation related to earliness and ploidy level.