Semi-supervised Hyperspectral image classification using spatial-spectral features and superpixel-based sparse codes

Classification of Hyperspectral images is an important issue in geoscience and remote sensing and attracts many attentions. To increase classification accuracy it is effective to use both spectral and spatial characteristics of pixels. An appropriate method for acquiring more discriminant features is using sparse coding and dictionary learning. To avoid large size of dictionary and to reach low computational cost, a preprocessing method is used to create superpixels. This preprocessing phase makes groups of more meaningful pixels and reduces computational cost together with the dictionary size. Moreover, these images have a large dataset and high-dimensional features therefore, the process of labeling pixels by an expert for supervised classification is very costly and time consuming. Semi-supervised methods are suitable, since they use a few labeled data to learn the structure of whole dataset. In this paper, a semi-supervised support vector machine named TSVM classifier is trained by a few numbers of samples to classify the remaining unlabeled instances. For measuring the accuracy of classification the Aviris Indian Pine dataset was used which contains a 145
_x
145
_x
200 hyper-spectral image. Simulation results showed acceptable performance in terms of overall accuracy, because of using the TSVM classifier, superpixel preprocessing, shared sparse codes and dictionary learning.