Attention-based Hyperspectral Classification Networks with Spectral Separability Analysis

For hyperspectral data analysis, the vector-based classification methods treat the multidimensional hyperspectral data as orderless set to extract discriminant features, in addition, the hyperspectral pixels also can be analyzed as sequence-based data to decide land-cover categories.In this paper, we introduce the attention mechanism into the hyperspectral sequence modeling to enhance spectral differences and suppress its similarity, and we further use the differences of attention weights between categories to analyze the spectral separability of different classes.First, we use CNN to extract spatial features and obtain the spectral embeddings.Second, the GRU networks are utilized to express the sequence features of ordered hyperspectral data.Afterwards, the spectral attention layer uses a query module to retrieve "what is the important information" over the whole sequence bands, and we get the attention weights on each spectral band.Finally, the weighted features are fed into a feedforward layer and a softmax layer for the final classification.Experimental results on two well-known hyperspectral datasets show that the attention module can not only help to improve the classification accuracies of categories with similar spectral profiles, but also perform quantitative analysis of spectral separability in a certain sense, especially for the types with small differences in spectral profiles.