Introducing training and parameter tuning for KOSP classification of hyperspectral images

Kernel-based Orthogonal Subspace Projection (KOSP) provides good results in the field of classification of hyperspectral images.1 However, an open-problem is the evaluation from the ground-truth samples of the prototypes that best represent the classes. In the original formulation of KOSP, 1 this preliminary (training) stage is very simple since for each class the prototype is computed as the centroid of the ground-truth samples. In order to improve KOSP performances, in this paper we introduce a minimization problem to evaluate the best prototypes from a given ground truth of a specific classification problem. K-fold cross-validation is used to avoid overfitting. The performance of the proposed methodology is tested by classifying the widely used 'Indian Pine' hyperspectral dataset collected by the AVIRIS spectrometer.