Hierarchical Multi-category Support Vector Machines Based on Inter-class Separability in Feature Space

According to the characteristics of support vector machines (SVMs), a separability measure of feature space based on support vector data description is proposed. By using this measure, two hierarchical multi-category SVMs, binary-Tree SVMs and single layer clustering SVMs, are presented. Experimental results with multi-spectral remote sensing image show that the proposed separability measure is very effective. The experiments also indicate that the single layer clustering SVMs is substantially faster than 1-v-1 SVMs, 1-v-r SVMs and DAG SVMs in classification, while maintaining comparable accuracy to these algorithms.