Manifold alignment for classification of multitemporal hyperspectral image data

Analyzing remotely sensed images to obtain land cover classification maps is an effective approach for acquiring information over landscapes that can be accomplished over extended areas with limited ground surveys. Further, with advances in remote sensing technology, spaceborne hyperspectral sensors provide the capability to acquire a set of images that have both high spectral and temporal resolution. These images are suitable for monitoring and analyzing environmental changes with subtle spectral characteristics. However, inherent characteristics of multitemporal hyperspectral images, including high dimensionality, nonlinearity, and nonstationarity phenomena over time and across large areas, pose several challenges for classification. This research addresses the issues of classification tasks in the presence of spectral shifts within multitemporal hyperspectral images by leveraging the concept of the data manifold. Although manifold learning has been applied successfully in single image hyperspectral data classification to address high dimensionality and nonlinear spectral responses, research related to manifold learning for multitemporal classification studies is limited. The proposed approaches utilize spectral signatures and spatial proximity to construct similar local geometries of temporal images. By aligning these underlying manifolds optimally, the impacts of nonstationary effects are mitigated and classification is accomplished in a representative temporal data manifold. Global manifolds learned from temporal hyperspectral images have a major advantage in faithful representation of the data in an image, such as retaining relationships between different classes. Local manifolds are favored in discriminating difficult classes and for computation efficiency. A new hybrid global-local manifold alignment method that combines the advantages of global and local manifolds for effective multitemporal image classification is also proposed. Results illustrate the effectiveness of utilizing common geometries of successive images in terms of classification accuracy. The proposed manifold alignment methods are also demonstrated to be successful in some practical cases where the targeted geographical region may only have training samples for one time period, yet exploration of other temporal images is desired. The proposed approaches are also demonstrated to be feasible domain adaptation methods that can handle classification spatially disjoint data sets, where training data are only available in one of the area