Fully Group Convolutional Neural Networks for Robust Spectral–Spatial Feature Learning

Convolutional neural network (CNN) has been widely applied in hyperspectral image (HSI) classification exhibiting excellent performance. Weak generalization of CNN models to different datasets is a common issue this domain largely because limited amount labeled training samples. In article, we propose fully group convolutional (FGCNN) method that integrates cascades shuffled convolutions tailored stages. To our knowledge, the first reported full-group model general, and design it particular for robust spectral–spatial HSI. primary feature extraction stage, develop an original multiscale spectral approach based on novel concept multikernel depthwise convolution define terms importance-weighted convolution. subsequent introduce discriminative with competition block capture informative features relatively few parameters. The final fusion stage defined as lightweight sharply reduces weights compared traditional methods fully connected layers. Experimental results three show proposed FGCNN yields accuracy under same hyperparameter settings current state-of-the-art.