Premium
Automatic Aurora Image Classification Framework Based on Deep Learning for Occurrence Distribution Analysis: A Case Study of All‐Sky Image Data Sets From the Yellow River Station
Author(s) -
Zhong Yanfei,
Ye Richen,
Liu Tingting,
Hu Zejun,
Zhang Liangpei
Publication year - 2020
Publication title -
journal of geophysical research: space physics
Language(s) - English
Resource type - Journals
eISSN - 2169-9402
pISSN - 2169-9380
DOI - 10.1029/2019ja027590
Subject(s) - computer science , artificial intelligence , sky , pooling , pattern recognition (psychology) , convolutional neural network , contextual image classification , deep learning , convolution (computer science) , feature (linguistics) , image (mathematics) , remote sensing , artificial neural network , geology , geography , meteorology , linguistics , philosophy
Developing a computational model for aurora image classification is an important task for polar research. However, the design of the handcrafted features in the traditional methods has insufficient descriptive ability for auroral morphology and is dependent on expert knowledge. In this paper, an automatic aurora image classification framework based on deep learning is proposed to solve the problem. In this framework, three basic deep learning networks—AlexNet, VGG, and ResNet—are used to automatically classify all‐sky aurora images via convolution and pooling processes and a back‐propagation mechanism, without requiring manual intervention. As shown in the class activation mapping (CAM), the proposed framework can extract the complex auroral feature representations automatically and discriminatively. The results obtained with all‐sky aurora images from the Yellow River Station demonstrate that the proposed framework has effective transfer ability and can achieve real‐time classification. It is also shown that the proposed method can achieve higher average classification accuracy than the AI‐MFLDA method. The statistical auroral occurrence distribution can be obtained based on the classification results. There are dominant morphological characteristics in the four magnetic local time (MLT) regions of the dayside aurora oval, which confirms the effectiveness of the proposed method.