Efficient Spectral Pyramid and Spectral-Spatial Feature Interactive Hyperspectral Image Classification.

Deep-learning frameworks have been widely used in the hyperspectral image (HSI) classification and have demonstrated promising performance. In this paper, we propose a novel HSI classification method with a deeper network and fewer parameters. Two novel modules named the efficient spectral pyramid (ESP), and improved spectral-spatial feature interactive (SSI) are designed to improving the SS3FCN, which is proposed in our previous work. Specifically, the ESP module composed of the dilated convolution is utilised to increase the spectral receptive field and make up the lost spectral information. In addition, the improved SSI module is leveraged to reduce trainable parameters and strengthen spectral features. Finally, the advancement of the proposed method is experimentally proved on three representative HSI data sets, and the effectiveness of these two novel modules are verified with ablation experiments.