Multi-classifiers and decision fusion for robust statistical pattern recognition with applications to hyperspectral classification.

摘要

In this dissertation, a multi-classifier, decision fusion framework is proposed for robust classification of high dimensional data in small-sample-size conditions. Such datasets present two key challenges. (1) The high dimensional feature spaces compromise the classifiers' generalization ability in that the classifier tends to over-fit decision boundaries to the training data. This phenomenon is commonly known as the Hughes phenomenon in the pattern classification community. (2) The small-sample-size of the training data results in ill-conditioned estimates of its statistics. Most classifiers rely on accurate estimation of these statistics for modeling training data and labeling test data, and hence ill-conditioned statistical estimates result in poorer classification performance.This dissertation tests the efficacy of the proposed algorithms to classify primarily remotely sensed hyperspectral data and secondarily diagnostic digital mammograms, since these applications naturally result in very high dimensional feature spaces and often do not have sufficiently large training datasets to support the dimensionality of the feature space. Conventional approaches, such as Stepwise LDA (S-LDA) are sub-optimal, in that they utilize a small subset of the rich spectral information provided by hyperspectral data for classification. In contrast, the approach proposed in this dissertation utilizes the entire high dimensional feature space for classification by identifying a suitable partition of this space, employing a bank-of-classifiers to perform "local" classification over this partition, and then merging these local decisions using an appropriate decision fusion mechanism. Adaptive classifier weight assignment and nonlinear pre-processing (in kernel induced spaces) are also proposed within this framework to improve its robustness over a wide range of fidelity conditions. Experimental results demonstrate that the proposed framework results in significant improvements in classification accuracies (as high as a 12% increase) over conventional approaches.