A simple but efficient method is proposed in this paper to exploit the capability of autoencoder with Extended Morphological Profile (EMP) for hyperspectral image (HSI) classification. In our work, the extended morphological profile is employed to extract the spatial information, then we join it with the spectral feature to describe the spectral-spatial property of the hyperspectral image. The obtained features are then fed into an autoencoder as input. After pre-training, the reconstruction layer is removed, then the network is equipped with a logistic regression at the last layer, with the role of supervised fine-tuning and classification. Experiments on KSC data set indicates that the proposed scheme can indeed achieve better performance of feature learning than the primitive features.