Feature extraction plays an important role in hyperspectral data processing.Due to the complicated imaging procedure,many traditional feature extraction method s fail to manage hyperspectral imagery.Moreover,the limited number of the labeled sample s in hyperspectral images also brings adverse effects to the supervised deep learning based methods,which are commonly used in feature extraction.In order to reduce the d ependence on the labeled samples in hyperspectral feature extraction,a new unsupervised f eature extraction method is proposed for presenting the spectral information based on c onvolutional neural networks and generative adversarial networks.In order to keep the traini ng process stable and improve the feature representation ability,a gradient penalty term r egarding the discriminator in the generative adversarial network is introduced into the objec tive function, which leads the performance of the discriminator to its optimal value gradually. In the feature extraction stage,a channel max-pooling method is proposed to improve the spect ral structure of the hyperspectral image by reducing the data dimension and retaining the spec tral information of hyperspectral image as much as possible.To evaluate the performa nce of the proposed method,support vector machines (SVM) and k-nearest neighbor (KNN) met hods are used to assess the extracted features.Experimental results on two real data sets show that the proposed method is superior to the traditional feature extraction methods.