Retraining maximum likelihood classifiers using a low-rank model


In this paper we propose a method for retraining a maximum likelihood classifier such that it may be applied to cases when the data distribution of the test data is different from the training data distributions. The proposed approach for retraining the classifier to the test data distribution is based on a constrained low-rank modeling of the unknown parameters, and may be designed such that the class structure is (to a larger degree) maintained after retraining. The proposed methodology is evaluated on two different applications; (1) cloud detection in Quickbird and WorldView-2 images and (2) tree cover mapping of tropical forest. The results show that the retrained classifiers clearly outperform their non-retrained counterpart.