Hyperspectral Image Classification Based on the Fusion of Superpixels and Deformable Features

Convolutional neural network is widely used in hyperspectral image classification for its excellent spatial-spectral feature extraction ability. However, the distribution of objects in hyperspectral images is highly complex, and it makes it difficult for conventional convolutional kernels to extract deep spatial features, which limit the classification performance. Therefore, a novel deformation feature fusion convolution network (DFFCN) is proposed in this work. Firstly, deformable convolution based on superpixels is used to extract the features of complex ground objects. At the same time, traditional convolution is used to extract neighborhood features of ground objects. Then, the above features are fused based on a channel attention mechanism. Finally, deep feature extraction and classification are performed on the fused features. The features based on the distribution of ground objects are fused while the neighborhood features of ground objects are retained. Experimental results on three real HSI data sets demonstrate that DFFCN can obtain better classification performances than several well-known classification methods.