Hyperspectral image classification using multi-level features fusion capsule network with a dense structure

The convolution neural network (CNN) methods have achieved excellent performance in hyperspectral image (HSI) classification. However, the convolution network fails to utilize the relative position information of the image effectively. The emergence of the capsule network has solved this limitation and made significant progress in the field of HSI classification. However, most capsule-based methods stack convolution layers to extract feature information, which can only get in-depth information while losing shallow information. In this paper, we proposed a multi-level features fusion capsule network based on dense structure (MLFF-CapsNet). In this framework, we designed a dense block composed of four-layer convolutions to fully extract spectral and spatial features. Then, each extracted feature map is concatenated to form multi-level features fusion capsules, which are transported to the dynamic routing algorithm to obtain the prediction category. The model presented has strong characterization and generalization capabilities under a few labeled samples. Experimental results on three hyperspectral datasets demonstrate that the proposed method achieves superior classification performance over the advanced comparison models.