Multi-scale Densely 3D CNN for Hyperspectral Image Classification

Yong Xiao, Qin Xu, Dongyue Wang, Jin Tang, Bin Luo

Abstract: Convolutional neural networks (CNNs) have shown good performance in hyperspectral image classification. However, there are still several problems unsolved. To address the existing problems, in this paper, we develop a multi-scale densely 3D convolutional neural network (CNN) for hyperspectral image classification. To characterize the hierarchical spatial-spectra feature, we design several branches with 3D convolutional kernels of different size. To overcome the gradient vanishing, we explore the dense connection where the input of each layer includes all the feature map produced in the previous layer. The proposed network that has five convolutional layers in total is shallow and short. The experimental results on Indian Pines and University of Pavia datasets have demonstrate the proposed method can acquire significant improvements over the state-of the-arts and require less computation and implementation time.