Analyzing hyperspectral and hypertemporal data by decoupling feature redundancy and feature relevance

The high information redundancy in hyperspectral and hypertemporal Earth observation data can limit the performance of supervised learning algorithms. Traditional sequential feature selection approaches start the search on the full set of correlated features, which is a computationally expensive task and impedes the search and discovery of spectral or temporal segments relevant for classification or regression tasks. We therefore propose to decouple the reduction of redundancy from the ranking of features. This is achieved by: 1) an unsupervised clustering of spectrally or temporally correlating neighboring features; 2) the definition of cluster representatives; and 3) the determination of the representatives' relevance by an support vector machine-based feature forward selection. Exemplified by two data sets for solving both, a hyperspectral and a hypertemporal classification problem, we show that our approach leads to well-interpretable spectral and temporal clusters, with comparable accuracies to more processing extensive traditional sequential feature selection.