Effect of Denoising in Band Selection for Regression Tasks in Hyperspectral Datasets

This paper presents a comparative analysis of six band selection methods applied to hyperspectral datasets for biophysical variable estimation problems, where the effect of denoising on band selection performance has also been analyzed. In particular, we consider four hyperspectral datasets and three regressors of different nature (ε-SVR, Regression Trees, and Kernel Ridge Regression). Results show that the denoising approach improves the band selection quality of all the tested methods. We show that noise filtering is more beneficial for the selection methods that use an estimator based on the whole dataset for the prediction of the output than for methods that use strategies based on local information (neighboring points).