CapsNet and Triple-GANs Towards Hyperspectral Classification

Hyperspectral processing technology becomes one of the most focused issues in remote sensing field. In the hyperspectral classification, significant improvements have been achieved by various deep learning methods. In general, deep learning algorithms adopt a cascade of layers to extract the hierarchical features. However, the deep hierarchical property will cause some defects such as overfitting and gradient vanishing. In this paper, a hybrid method based on CapsNet and Triple-GANs has been explored to avoid overfitting and extract the effective features. Unlike ordinary CNN, the CapsNet is the consist of a group of capsules with vectorizing the activation output which could consider not only spectral deep features but also the relative locations of these features. The Triple-GANs is a game system with three players: a generator, a classifier and a discriminator. When the Triple-GANs converges to balance, the credible labelled samples could been obtained by the generator which boost the CapsNet in the classification task. The main contents are as follows: 1)By introducing the CapsNet in the hyperspectral feature extraction, the 2D convolution operations are replaced by 1D to adapt the pixel-wised spectral features. 2) Use the Triple-GANs and CapsNet to do the hyperspectral classification on small training dataset. Experimental results show that this algorithm can obviously improve the performance of classification compared with the traditional methods.