Support tensor machine with local pixel neighborhood for hyperspectral image classification

A multiclass support tensor machine (STM) for the classification of remotely-sensed imagery is investigated in this study aiming at simultaneously exploiting the spectral and spatial information for accurate image interpretation. Spatial relationship of neighboring pixels has been taken into consideration by a local pixel neighborhood (LPN), which processes the local imagery patch as a cube, and is capable of separating land classes in both spectral and spatial domains. To deal with the tensor data and keep the data structure in high-order feature space, support vector machine has been extended to support tensor machine by the multilinear algebra. Experiments conducted on AVIRIS hyperspectral image revealed that the STM achieved much better results than the standard SVM classifier.