2DPCANet: Dayside aurora classification based on deep learning

Abstract

The mysterious and beautiful aurora represents various physical meaning, thus the classification of aurora images have significant scientific value to human beings. Principal component analysis network (PCANet) has achieved good results in classification. But when using PCANet to extract the image features, it transform original image into a vector, so that the structure information of the image are missing. Compared with PCA, 2DPCA is based on 2D image matrices rather than 1D vectors so that 2DPCA can use the structure information of original image more efficiently and reduce the computational complexity. Based on PCANet, we develop a classification method of aurora images, 2-dimension PCANet (2DPCANet). To evaluate 2DPCANet performance, a series of experiments were performed on two different aurora databases. The classification rate across all experiments was higher using 2DPCANet than PCANet. The experiment results also indicated that the classification time is shorter using 2DPCANet than PCANet.