Evaluating the performance of hyperspectral feature selection algorithm using Tsallis entropy redundancy

Feature selection is a widely used technique in hyperspectral data processing, but there is little work concerning the evaluation of the performances with respect to different feature selection methods especially when the ground truth map is absent. This paper analyzes the selection criterion from probabilistic statistics and correlation analysis, and points out the disadvantage of directly using them for evaluation application, and then proposes the evaluation index based on the Tsallis entropy redundancy. This index has direct proportion in the relevant information among multi variables, and can be easily extended for the classification performance evaluation of hyperspectral feature selection. The AVIRIS data has been applied to the proposed method and the results show that when the overall classification difference is no less than 2%, the correct rate of evaluation is greater than 75%, furthermore, when the difference is no less than 8%, the correct rate is greater than 90%.