Fuzzy Entropy based feature selection for classification of hyperspectral data

This paper proposes to use a fuzzy entropy based feature selection approach to reduce the dimensionality of DAIS hyperspectral data. To compare its performance, three other filter based feature selection approaches were used. A support vector machine was used as a classification algorithm. In order to compare various feature selection approaches with full dataset, McNemar’s test based test for non-inferiority was used. The results of the non-inferiority testing between total features and selected features by different feature selection approaches suggests the suitability of fuzzy entropy based approach in terms of classification accuracy and small number of features used to derive same level of accuracy as achieved by the use of total number of features. Results suggest inferior performance by Relief based feature selection approach with the used dataset. The classification accuracy achieved by entropy and signal to noise ration based feature selection approach was comparable to that of fuzzy entropy based approach but requires more number of selected features.