Using active learning to quantify how training data errors impact classification accuracy over smallholder-dominated agricultural systems