Applying a dynamic subspace multiple classifier for remotely sensed hyperspectral image classification

The multiple classifier system has received remarkable attentions for improving the performance of a single classifier in recent years. The random subspace method (RSM) is one of the multiple classifier systems. In RSM, classifiers are trained by data set with randomly selected and fix-sized feature subsets and are combined using simple majority vote in the final decision rule. The feature subset size of the reduced data set and the fashion to construct the feature subset are two key issues affecting the performance of RSM. The former must be pre-assigned and the latter is randomly generated based on the former assignment. This study applies a dynamic subspace multiple classifier system to the classification of hyperspectral images, and investigates its performance on various conditions. The experimental results demonstrate that the dynamic subspace multiple classifier can achieves better classification results than RSM, and some important results are revealed as well in this study.