Semi-supervised Hyperspectral Classification

Z.Wang, N. Nasrabadi, and T. Huang, "Semi-supervised Hyperspectral Class-action using Task-driven Dictionary Learning with Regularization", IEEE Transactions on Geosciences and Remote Sensing ( TGRS), vol. 56, no. 3, pp. 1161-1173, Mar. 2015.



  •  We present a semi-supervised method for single pixel classification of hyperspectral images. The proposed method is designed to address the special problematic characteristics of hyperspectral images, namely, high dimensionality of hyperspectral pixels, lack of labeled samples, and spatial variability of spectral signatures.

  • To alleviate these problems, the proposed method features the following components. First, being a semi-supervised approach, it exploits the wealth of unlabeled samples in the image by evaluating the confidence probability of the predicted labels, for each unlabeled sample. Second, we propose to jointly optimize the classifier parameters and the dictionary atoms by a task-driven formulation, in order to ensure that the learned features (sparse codes) are optimal for the trained classifier. Finally, it incorporates spatial information through adding a Laplacian smoothness regularization to the output of the classifier, rather than the sparse codes, making the spatial constraint more flexible.

  • The proposed method is compared to a few comparable methods for classification of several popular datasets, and it produces significantly better classification results. The comparison results on Indian Pie dataset are displayed as follows (Joint + Laplacian is the proposed one).

Zhangyang (Atlas) Wang​