B1000
Title: Functional variables selection in hyperspectral image classification
Authors: Manuel Oviedo de la Fuente - University of Oviedo (Spain) [presenting]
Manuel Febrero-Bande - University of Santiago de Compostela (Spain)
Wenceslao Gonzalez-Manteiga - University of Santiago de Compostela (Spain)
Abstract: Different models classification algorithms are reviewed for the prediction of the future class of pixel in a hyperspectral image that have in common that make use of Functional Data Analysis (FDA). The advantage of FDA over classical model is that it is able to exploit this continuous nature of the information of spectral curves in a better way. We used a multiclass one-versus-one (majority voting) and one-versus-rest functional GAM model. In addition, functional non-parametric classification by mean of proximity measures (kNN and kernel classifiers) and by means of depth measures (depth-based classification) can also help to discriminate the pixel class through the shape of the spectral curve. The second part of the communication is devoted to the problem of variable selection. A selection method that is designed to mixed covariates of different nature: scalar, multivariate, functional, etc, is proposed. The proposal begins with a simple null model and sequentially selects a new variable (using distance correlation) to be incorporated into the final prediction model. The algorithm has shown quite promising results in the regression framework and its extension to the classification problem is attempted.
