Robust classification of hyperspectral data

High dimensionality and highly correlated features are two important characteristics of hyperspectral data that leads to poor performance of conventional classification methods. Furthermore, hyperspectral sensors usually provide relatively low optical resolution, which implies that pixels are bound to cover a mixture of objects with different reflective properties. Since it is common to define sharp labels on pixels, classes might not be adequately described with a single mode Gaussian as it is done in many conventional and contemporary classification methods for hyperspectral data. We study a framework that facilitates a penalized classification, making the classifier robust for overfitting. This framework also allows the classes to be modeled as a mixture of subclasses, giving the model more flexibility.