Investigation of the impact of dimensionality reduction and feature selection on the classification of hyperspectral EnMAP data

In this paper, we address the classification of hyperspectral data which is comparable to the data acquired with the Environmental Mapping and Analysis Program (EnMAP) mission, a hyperspectral satellite mission supposed to be launched into space in the near future. While simulated EnMAP data has already been released, only relatively few studies have focused on investigating the performance of approaches for classifying such EnMAP data. Hence, in a recent paper, a contest for classifying EnMAP data has been initiated to foster research about possible exploitation strategies. Based on the dataset presented therein, we present a framework involving techniques of dimensionality reduction, feature selection and classification. We involve several classifiers for pixelwise classification based on different learning principles and investigate the impact of approaches for dimensionality reduction and feature selection on the classification results. The derived results clearly reveal the potential of respective techniques and provide the basis for further improvements in different research directions.