Adaptive hierarchical classification with limited training data
This research focused on the development of a hierarchical approach for classification that is robust with respect to training data that are limited both in quantity and spatial extent. Many difficult classification problems involve a high dimensional input and output space (candidate labels). Due to the "curse of dimensionality," it is necessary to reduce the size of the input space when there is only a limited quantity of training data available. While a significant amount of research has focused on transforming the input space into a reduced feature space that accurately discriminates between the classes in a fixed output space, traditional approaches fail to capitalize on the domain knowledge and flexibility gained by transforming the feature space and the output space simultaneously. A new approach is proposed that utilizes domain knowledge, which is automatically discovered from the data, to combat the "small sample size" problem. Spatially limited training data can result in poor inference concerning the true populations. The detrimental impact that can result if this issue is ignored is explored and demonstrated. Transferal of information that was previously acquired is used to update the signatures with the new clusters if the hypothesis that the new clusters are indeed just deformed versions of what already exists in the spectral library is accepted. Independent of limited training data, both in terms of the spatial implications and limited quantity, different sampling subsets of the same ground truth may result in slightly different classifiers. This issue has not been addressed rigorously. The advantages gained by using an ensemble of classifiers built from sub-samples of training data are widely acknowledged but have not previously been used in the context of a hierarchical classifier for remote sensing data or for hyperspectral data in general. The ensemble of classifiers is used to identify a suitable level of the tree for situations where the resolution of the output space cannot be supported. Further decisions of how the classification structure should be adapted and at what level need to be made are explored. Furthermore, pseudolabeled data are utilized to improve classification results at that level of resolution.