Multiscale Densely Connected Attention Network for Hyperspectral Image Classification

Abstract

Hyperspectral image classification (HSIC) based on deep learning has always been a research hot spot in the field of remote sensing. However, most of the classification models extract relevant features based on fixed-scales convolution kernels, which ignores the complex features of hyperspectral images (HSIs) at different scales and impairs the classification accuracy. To solve this problem, a multiscale densely connected attention network (MSDAN) is proposed for HSIC. First, the model adopts three different scales modules with dense connection to enhance classification performance, strengthen feature reuse, prevent overfitting and gradient disappearance. Besides, in order to reduce the model parameters and strengthen the extraction of spatial–spectral features, the traditional three-dimensional convolution is replaced by three-dimensional spectral convolution block and three-dimensional spatial convolution block. Furthermore, the spectral–spatial–channel attention is embedded into the end of each scale to enhance the favorable features for classification and further extract the discriminant features of the corresponding scale. Finally, the key feature extraction module is developed to extract multiscale fusion features to further enhance the classification performance of the network. The experimental results carried out on real HSIs show that the proposed MSDAN architecture has significant advantages compared with other most advanced methods.