Manifold alignment for classification of multitemporal hyperspectral data

For hyperspectral image classification, manifold learning has proven to be useful for feature extraction from high dimensional data sets. In a traditional manifold learning framework, a low dimensional manifold describing spectral characteristics is developed for classification. However, drift of class distributions in multitemporal hyperspectral data can induce unfaithful manifold representations while exploiting spectral similarities of these images. Classes in multitemporal images often exhibit similar geometries but possibly are represented in different manifold coordinates. It may be possible to utilize certain prior information of these temporally related image data to map such similarities to a common latent space. In this paper, a manifold alignment framework is proposed to leverage prior knowledge while exploiting spectral similarities in the underlying manifolds of two multitemporal hyperspectral images. The essential similar local geometric structures of classes in the temporal sequence are encoded into a common feature space where the classificationtask is naturally feasible.