Empirical models to estimate herbage mass and grass quality from multispectral imagery acquired by unmanned aerial vehicles (UAVs) often poorly generalize to different grasslands. We therefore investigated whether the generalization performance can be improved by replacing the commonly used single-target regression algorithms by corresponding multi-target algorithm adaptations which can simultaneously predict herbage mass and grass quality (dry matter percentage, crude protein, and structural carbohydrates). By additionally considering the relationships between the target variables, these multi-target algorithm variants have the potential to yield better generalization performance. We found that for Partial Least Squares, K-Nearest Neighbours, and Random Forest, the multi-target variants tended to perform better than their single-target counterparts, while for Extremely Randomized Trees mostly the opposite was true. Given the usual lack of ground-truth data for the model to learn the underlying relationships, we suggest to consider the use of multi-target regression whenever several grass parameters are estimated.
