Benjamin Mack

Can I Trust My One-Class Classification?

Contrary to binary and multi-class classifiers, the purpose of a one-class classifier for remote sensing applications is to map only one specific land use/land cover class of interest. Training these classifiers exclusively requires reference data for the class of interest, while training data for other classes is not required. Thus, the acquisition of reference data can be significantly reduced. However, one-class classification is fraught with uncertainty and full automatization is difficult, due to the limited reference information that is available for classifier training. Thus, a user-oriented one-class classification strategy is proposed, which is based among others on the visualization and interpretation of the one-class classifier outcomes during the data processing. Careful interpretation of the diagnostic plots fosters the understanding of the classification outcome, e.g., the class separability and suitability of a particular threshold. In the absence of complete and representative validation data, which is the fact in the context of a real one-class classification application, such information is valuable for evaluation and improving the classification. The potential of the proposed strategy is demonstrated by classifying different crop types with hyperspectral data from Hyperion.

In-depth comparisons of MaxEnt, biased SVM and one-class SVM for one-class classification of remote sensing data

ABSTRACT One-class classification is an increasingly used classification approach in the remote sensing community. It can be used to classify one specific class and requires only labelled training data of this class of interest. Several one-class classifiers have been introduced and many comparative studies evaluate their performance. In most, if not all, of these comparisons each classifier is analysed with a specific (hyper-) parameter and threshold selection approach. In this letter, we present an ‘in-depth’ comparison of the frequently used one-class classifiers one-class Support Vector Machine (SVM), biased SVM (bSVM) and MaxEnt, that is, a frequently used maximum entropy-based technique. We create various classification tasks with eight classes of interest and three feature sets (multi-temporal RapidEye, TerraSAR-X and fused data set) and evaluate the overall performance of the two classifiers with different parameter and threshold selection approaches. Our results show that over all classification approaches bSVM outperforms one-class SVM and MaxEnt in terms of discriminative power. However, the state-of-the-art implementation of the bSVM performed relatively poor and the best results were obtained with an alternative threshold selection approach. We found that even the best overall approach still performs poor in a significant amount of tasks. Therefore, we conclude that not only more sophisticated model selection approaches should be developed but also diagnostic tools that support the user in the evaluation of a classification result in the absence of a complete and representative test set.