Deep learning has achieved good results in recent times in the field of hyperspectral images (HSIs) classification. For traditional convolutional neural networks, the excessive increase in network depth will lead to overfitting and gradient disappearance. In addition, most of the methods for HSIs classification usually don’t consider the strongly complementarity and relevance between different levels or directions. In order to solve the two problems above, this paper proposes a spatial-spectral joint framework based on Gabor filtering and feature fusion network (GFFN) for HSIs classification. Firstly, Gabor filter can effectively extract spatial information including edge and texture in different directions. Moreover, Gabor filtering can effectively alleviate the overfitting problem, especially when the training sample set is small. Secondly, this paper extracts spectral features by combining convolutional neural network (CNN) and residual network (ResNet). This effectively solves the degradation problem of CNN. Thirdly, the proposed method integrates the feature outputs in different directions to further improve the classification accuracy. The experimental results obtained with three public HSI datasets show that the GFFN is superior to other competitive ones.
|