I. Introduction
Landcover classification is one of the most common tasks for remote sensing applications, and the development of adequate classification strategies is an ongoing research field. In this context, hyperspectral imagery is probably the most valuable as well as challenging single data source. Hyperspectral sensors provide detailed and spectrally continuous spatial information, enabling the discrimination between spectrally similar landcover classes (e.g., [1] and [2]). However, it is well known that data dimensionality and high redundancy between individual spectral bands cause challenges during data analysis, e.g., the performance of standard supervised classifiers is often limited in terms of classification accuracy. Therefore, alternative methods such as support vector machines (SVMs, e.g., [3] and [4]), ensemble-based learning (e.g., [5] and [6]), or classifiers based on multinomial logistic regression (e.g., [7]–[9]) have been successfully used for hyperspectral image classification.