Deep 2D Convolutional Neural Network with Deconvolution Layer for Hyperspectral Image Classification

Feature extraction and classification technology based on hyperspectral data have been a hot issue. Recently, the convolutional neural network (CNN) has attracted more attention in the field of hyperspectral image classification. To enhance the feature extracted from the hidden layers, in this paper a deconvolution layer is introduced in the deep 2DCNN model. Analyzing the function of convolution and pooling to determine the structure of the convolutional neural network, deconvolution is used to map low-dimensional features into high-dimensional input; the target pixel and its pixels in a certain neighborhood are input into the network as input data. Experiments on two public available hyperspectral data sets show that the deconvolution layer can better generalize features for the hyperspectral image and the proposed 2DCNN classification method can effectively improve the classification accuracy in comparison with other feature extraction methods.