Hyperspectral images mapping with group sparse representations

In this paper, we propose a new framework for remote sensing images classification based on group sparse representations. It is well known that it is very difficult to find a suitable sparse representation for remote sensing images because of complicated ground features. Here a remote sensing image is deemed a combination of sub-images of smooth, edges and point like components. Since each domain transformation method is only capable of representing a particular kind of ground objects or textures, a group of domain transformations are combined to sparsely represent the whole image. By applying the group sparse representations as a prior in Maximum a Posterior (MAP) for ill-conditioned problems, each channel of an input hyperspectral data cube can be separated into sub-images of the same size as the input image by using the iterative soft-thresholding algorithm (ISTA). For a particular channel, smooth areas will most likely show in the sub-image with similar amplitude. These smooth areas in a sub-image will work as implicit spatial constraint for that channel. Therefore, those reconstructed sub-images can be added to the original hyperspectral data cube as an augmented data cube with spatial constraint embedded for solving classification problem. This new framework brings hopes for all the classification methods without spatial constraints for the purpose of improving classification accuracy. The improved classification results can be achieved without any changes to those classification methods but using the augmented data cube.