AUTOMATIC HYPERSPECTRAL IMAGE CLASSIFICATION BASED ONDEEP FEATURE FUSION NETWORK

The traditional machine learning algorithm always pays attention to spectral features on automatic hyperspectral image (HSI) classification, and there exists insufficient feature extraction under the condition of small samples. In addition, the generalization ability of the model is not strong. In this paper, a novel method named specific two-dimensional-three-dimensional fusion strategy is proposed, which uses a spatial-spectral feature fusion network based on two-dimensional convolution and three-dimensional convolution to extract the rich features, so as to keep the spatial and spectral information intact. The validity of this method is verified by comparing different classification algorithms. Experiments were carried out on three widely used HSI data sets (i.e. Indian Pines, Salinas and Pavia University). In case of small training sets, the experimental results show that the proposed method outperforms the existing methods.

Automated summary

A novel specific two-dimensional-three-dimensional fusion strategy is proposed in this paper, using a spatial-spectral feature fusion network based on two-dimensional convolution and three-dimensional convolution to extract rich features while keeping spatial and spectral information intact. Experimental results show that the proposed method outperforms existing methods in cases of small training sets.