REMOTE SENSING SCENE CLASSIFICATION BASED ON RES-CAPSNET

Capsule Network (CapsNet) is a brand new network structure. Aiming at limitations of Convolutional Neural Networks (CNNs), it designs capsule vector and dynamic routing to represent features and perform classification. However, though CapsNet has achieved state-of-the-art performance on simple MNIST data set, its potentials on remote sensing are not widely studied and explored. In this paper, we proposed a new network structure called Res-CapsNet to achieve remote sensing scene classification based on CapsNet. By introducing double residual modules into basic CapsNet, the capsule network is able to perform well on remote sensing images with more complex textures. Experimental results on UCMerced data set validate the effectiveness of our model, which also shows the potentials of capsule layers compared to pooling.