PYRAMID CONVOLUTIONAL NEURAL NETWORKS AND BOTTLENECK RESIDUAL MODULES FOR CLASSIFICATION OF MULTISPECTRAL IMAGES

The newly emerging classifier using deep network architectures and pyramid bottleneck modules exhibits stronger capability than traditional classifiers. However, they are only suitable for color images or hyperspectral images due to the structural, textural and spectral differences against multispectral images. In this paper, a new network is designed for the classification of high-resolution multispectral images. The new network follows the architecture of pyramid residual network, but the input size, filter size, and filter number of each layer are totally different. These designs make the pyramid residual network conforming to the multispectral advantages of spatial resolutions so as to improve classification performance. Experiments on the satellite multispectral data from GF-1 and RapidEye demonstrate the superiority of the new network.