Hyperspectral image classification using distance metric based 1-dimensional manifold embedding

Hyperspectral remotely sensed image provides very informative information for a wide range of applications that relate to landcover classification. Many studies have shown that the spectral-spatial information is well effective for hyperspectral image (HSI) classification. However, for the spatial based methods, it may sometimes encounter many difficulties in obtaining the spatial prior of different landcovers. Moreover, the spatial prior has to be carefully tuned during each experiment. In this paper, we propose a distance metric learning based 1-dimensional manifold embedding (1DME) for hyperspectral image classification. In our approach, the Mahalanobis matric is first employed to learn an similarity metric of pairwise pixels. The measurement can well indicate proximity of different classes. Then, according to the piecewise affinity, we adopt the developed 1-dimensional manifold embedding to sort the entire data points so that pixels with similar property stay close. Since the entire data points are ordered, several regressors are applied to the ordered sequence, and the averaged results are treated as the prediction. Experiment is conducted on the well acknowledged Indian Pines benchmark data set, and the results validate the efficiency of the proposed method.