Combining Contextual Information for Subspace Based Hyperspectral Image Classification

Hyperspectral image classification is a difficult task in remote sensing community due to the challenges caused by high dimensionality and limited training samples. The traditional classification approaches used to exploit the rich spectral information only, while in the past decades the contextual information has been considered extensively to promote the performance of hyperspectral classification. In this paper, we attempt to exploit the contextual information in the form of superpixel to improve the spectral based classification approach. Superpixels at multiple scales are used to learn spatial features and then a subspace based multinominal logistic regression is conducted to estimate the probability estimation at different scales. Finally, the classification result is achieved by a decision fusion process. Experimental results show that by combining the contextual information in the form of superpixels, our method works effectively and produce satisfiable results compared to the related spectral-spatial methods.