Semi-supervised classification of hyperspectral images based on two branch autoencoder

Abstract

In this paper, we propose a novel semi-supervised method based on two branch autoencoder (TBAE) for hyperspectral images. A branch is classification function, and another branch is decoder function. Both of functions promote each other in training process. The encoder is fit for both classifier and decoder. The encoder and the decoder guide feature extraction from unlabeled samples. The encoder and the classifier guide classification for all labeled samples. We can get more information from unlabeled samples for classification. Finally, we use convolutional autoencoder (CAE) to extend TBAE and obtain TBCAE. Compared with ANN and CNN, TBAE and TBCAE have a better performance in the case of a small number of labeled samples. Use different datasets to verify these methods and print visualization of hidden layer and the reconstructed data. The results demonstrate that the proposed framework obtains credible results with a small number of labeled samples.