Spatial-Spectral Hyperspectral Image Classification Using Random Multiscale Representation

This paper presents a novel spatial-spectral classification method for remotely sensed hyperspectral images. First of all, a multiscale representation technique based on random projection, referred as random multiscale representation (RMSR), is proposed to extract the spatial features from the given scene. The idea behind RMSR is to properly model the spatial characteristics comprised by each pixel vector and its neighbors by some criteria computed at all reasonable scales, and then compress the implicit high-dimensional spatial features by using a very sparse measurement matrix that approximately preserves the salient spatial information. The entire process is explicitly performed by computing simple criteria (i.e., the first two moments) at rectangular scales of random bands, according to the nonzero entries of the sparse measurement matrix. Subsequently, a composite kernel framework is utilized to balance the extracted spatial features and the original spectral features in the classifier. Our proposed method is shown to be effective for hyperspectral image classification purposes. Specifically, our experimental results with hyperspectral images collected by the airborne visible/infrared imaging spectrometer and the reflective optics spectrographic imaging system demonstrate the effectiveness of the proposed method as compared to other state-of-the-art spatial-spectral classifiers.