 This paper examined the effect of mislabeled training data on classification performance for land cover mapping. It was found that both support vector machines, SVM, and random forests, RF, were relatively insensitive to low levels of random noise up to 25 to 30 percent. However, when the level of noise increased, the accuracy of both algorithms dropped significantly. Additionally, the RF algorithm proved to be more robust than the SVM algorithm, regardless of the type of noise present. Furthermore, the cross-validation procedure was affected by the presence of class-label noise. This article was authored by Charlotte Pelletier, Sylvia Valero, Jordy Inglada, and others.