This thesis investigates the role of preprocessing, rejection classes and ample relabelling in the classification of convective storm cells. This problem is representative of pattern recognition problems displaying high data dimensionality, small sample sets, and imperfect sample labelling. A battery of standard classifiers are compared using preprocessing strategies such as interquartile membership, principal and independent components. Rejection classes initiate the trade-off between improvement of performance and exhaustive classification; this is accomplished by refusing to assign class labels to samples 'near' class boundaries. Classifier specific values are used to define these boundaries. Sample relabelling is based on robust reclassification and median average deviation, fuzzy logic and probabilistic learning. This thesis uses meteorological volumetric radar data to analyse the effectiveness of these concepts. It is determined that the number of independent components to consider should not be basedon a cumulative variance in principal components and that interquartile membership is mot effective with real variables; rejection classes pay a high price in terms of the number of unlabelled samples although they improve classifier performance; robust reclassification consistently improves classifier performance over a broad range of classifiers. Future validation of the number of event prototypes will confirm the application of robust reclassification to this problem.