A dynamic classification scheme for mapping spectrally similar classes: Application to wetland classification

Abstract

Applying numerous features is common for complex land cover classification, which makes feature selection necessary. Although the selected feature subsets may yield high overall accuracies in the case of spectrally similar classes such as wetlands, several individual accuracies are often low. A reason is that one set of features used to separate one specific class from the rest might not be appropriate for delineating another class. An additional reason is that while the overall accuracy is applied to evaluate the potential of a feature subset, it may be influenced by a few high-accuracy classes that are spectrally distinct and for which collecting enough training data is feasible. In this article, rather than simultaneously mapping all classes, they were individually classified using a different feature selection. Spectral analysis was applied to determine both the order of the classes to be mapped and a merging scheme that was applied to the remaining classes to increase the accuracy of the target class. The proposed approach was applied to wetland mapping using five pilot sites throughout Newfoundland and Labrador, Canada. The dataset available for each pilot site differed in quality and quantity. However, the proposed method accurately classified wetlands in all pilot sites, even those with limited satellite and/or field data, and outperformed the classic method by increasing the average producer and user accuracies of wetlands by up to 22% and 25%, respectively, and yielding overall accuracies up to 93%. Among wetland classes, Shallow Water was easier to be distinguished, as a result of being spectrally less similar to the rest of the wetland classes. Based on the obtained results, the proposed method can be effectively applied for classifications involving spectrally similar classes, including sea ice and crop mapping.