Hyperspectral Feature Selection Based on Mutual Information and Nonlinear Correlation Coefficient

Mutual information (MI) has obvious potential for feature selection, but this has not been fully exploited in the past. In order to make numerical computation easier and more accurate, the MI of the whole multi-dimensional data can be decomposed into an amount of one-dimensional MI and one-dimensional conditional MI components. This paper reveals that using one-dimensional MI components to replace the one-dimensional conditional MI components may be problematic when the features are highly correlated, and we propose a method that using nonlinear correlation coefficient (NCC) to replace some one-dimensional MI components, which also including the conditional ones. Simulations are carried out on the AVIRIS 92AV3C dataset and the results show great potential for improvement in classification accuracy.