高时空分辨率的地面全天空光学成像是极光物理机制研究的重要手段，每年产生大量的观测数据，迫切需要自动化的极光图像分析技术，然而极光形态和动态演变过程极为复杂，至今尚无明确的形态特征描述和合理的分类方法。我们以往的研究工作表明，极光图像特征描述是对其进行形态分类、聚类和动态事件分析的关键，传统的特征工程方法在性能上已经遇到瓶颈。本项目开展基于多级深度学习的极光特征提取、分类和检测方法研究。利用十万数量级的标记数据集，研究有效的数据扩增、网络结构优化和多级网络训练方法，实现弧状/冕状极光的自动分类；针对原始观测数据的自动预处理问题，研究基于深度网络特征提取和支撑矢量数据描述的分立极光图像检测方法；在深度分类网络的基础上构建深度回归网络，研究深度回归网络的多级优化和多尺度搜索方法，实现局部特殊结构极光图像的自动检测和准确定位。本项目研究成果将为极光物理机制研究提供有效的海量图像自动分析方法。

Ground-based all-sky optical imaging of aurora with high spatial and temporal resolutions forms a central element in studying the physical mechanism of the near-Earth space phenomena. Automatic methods are urgently needed to process the hundreds of millions of images produced annually. However, due to its complex morphology and dynamic characteristics, up to now the number of aurora categories and the characteristics of each category are not well defined. Our previous research demonstrated that feature description plays an important role in auroral image classification, clustering analysis and dynamic event analysis. The traditional feature engineering approach has met its performance bottlenecks. This project will investigate methods for feature extraction, classification and detection of auroral image based on multi-stage deep learning. By utilizing a labeled dataset with one hundred thousand auroral images, we will study effective methods for data augmentation, network structure optimization and multi-stage training, to implement the automatic classification of arc and corona auroras. Targeting the problem of automatic raw observation pre-processing, we will study methods for discrete aurora detection based on feature extraction using deep network and detection using support vector data description. Furthermore, we will convert the deep classification network to a deep regression network. Multi-stage optimization of the regression network and multi-scale searching methods will be investigated to detect local special auroral structures. The achievements of this project are expected to provide new auroral image analysis methods for research on the physical mechanism of aurora.