Label Noise Robust Classification Of Hyperspectral Data

摘要

Supervised classification on remotely sensed data is a classical method to detect objects automatically and to update topo-graphical datasets. To train the classifier, already labeled image data are needed. The training labels are typically created manually for parts of the given data, which is time-consuming and thus costly. Another approach would be to use the labels from an old and outdated dataset, called map, to avoid manual effort. Due to changes over time some of the labels might be wrong, and thus the used classifier has to be able to deal with that. In this paper, we apply label noise robust classification on two hyperspectral datasets using a simulated outdated map for training.