A label noise tolerant random forest for the classification of remote sensing data based on outdated maps for training

摘要

Supervised classification of remotely sensed images is a classical method for change detection. The task requires training data containing images with known class labels. If these labels are acquired from an outdated map, the classifier must cope with errors in the training labels. Several papers state that the random forest classifier is robust with respect to these errors (label noise) by design, but for a large amount of label noise or for noise affecting different classes differently this assumption does not necessarily hold. In this paper we suggest an adaptation of the random forest classifier by integrating a model for label noise based on the idea that a training sample should not be assigned to one class only, but to all classes, each with a certain probability. The adapted random forest is embedded in an iterative scheme for the context-based classification of remote sensing data using the outdated map not only to provide the labels of the training samples, but also to support the classification process in unchanged areas. Our experiments are based on five test areas and the results show a higher accuracy using the suggested new method than using the standard random forest.