Clustering of Ultraviolet Auroral Oval Images Based on Deep Representation Learning
-
摘要: 极光受太阳风驱动的地磁亚暴等大尺度动力学影响，其形态及演化因不同的太阳风-磁层-电离层耦合作用可能表现不同。目前，极光卵及其形态的归类大多依据极光演化理论作主观定性分析，没有明确的分类标准，故难以借助统计分析方法和有监督分类模型开展客观定量研究。建立了基于深度表征学习的紫外极光卵图像聚类模型（MoCo-GMM），并利用空间环境参数设计了评估模型物理合理性的方法，在大规模POLAR卫星紫外极光卵图像数据上进行了实验，聚类结果不仅具有良好的簇内凝聚性和簇间分散性，且具备一定的物理可解释性，有效实现了基于图像的极光卵及其形态的客观归类。
-
关键词:
- 极光卵及形态归类 /
- 紫外极光卵图像聚类 /
- MoCo-GMM模型
Abstract: Aurora is affected by large-scale dynamics such as geomagnetic substorm driven by solar wind, due to varies solar wind-magnetosphere-ionosphere coupling effects, its morphology and evolution can be different. Currently, categorization of aurora oval and its morphology is mostly based on auroral evolution theory to do subjective qualitative analysis and has no clear classification standard, which makes it challenging to conduct objective quantitative research with statistical analysis method and supervised classification models. Ultraviolet (UV) auroral oval image clustering model (MoCo-GMM) was established based on deep representation learning, also a method was designed to evaluate physical rationality of the model by using space environment parameters. Additionally, experiments on large-scale POLAR UV auroral oval image data were carried out. Clustering results of MoCo-GMM obtained not only delightful intra-cluster cohesion and inter-cluster separation, but a certain degree of physical interpretability, which means we effectively realized objective categorization of aurora oval and its morphology based on images. -
图 2 MoCo对比学习紫外极光卵图像表征
Figure 2. Contrastive learning of UV auroral oval image representations with MoCo
图 3 VggNet-16迭代式分类的准确率和损失值变化曲线
Figure 3. Accuracy and loss curves of iterative classification with VggNet-16
图 4 不同模型对日晖干扰图像和极光卵边界模糊图像的形态提取掩膜对比
Figure 4. Comparison of morphology extraction masks obtained by different models from images with dayglow effect and images with blurred boundary
图 5 紫外极光卵图像数据预处理管线及效果
Figure 5. Pipeline and results of UV auroral oval image preprocess
图 6 不同编码器结构的MoCo表征学习的准确率变化曲线
Figure 6. Accuracy curve of MoCo representation learning with different encoders
图 9 POLAR紫外极光卵图像MoCo-GMM 10-聚类结果
Figure 9. MoCo-GMM 10-clustering results of POLAR UV auroral oval images
图 10 不同模型提取同一幅紫外极光卵图像特征的10-最近邻对应图像
Figure 10. UV auroral oval images corresponding to 10-nearest neighbors of feature extracted from the same image with different models
图 11 不同空间环境参数组合与紫外极光卵图像10-聚类结果的关联度
Figure 11. Relation degree between different combinations of space environment parameters and UV auroral oval image 10-clustering results
表 1 各算法紫外极光卵图像特征提取结果对比
Table 1. Comparison of UV auroral oval image feature extraction results with different algorithms
Algorithm Result Representation Degree ORB No unified dimensions \ LBP No unified dimensions \ PCA 15 dimensions 0.865 HOG 3600 dimensions 0.854 MoCo 512 dimensions 0.892 
下载: 导出CSV
表 2 各特征聚类算法效果对比
Table 2. Comparison of six feature clustering algorithms.
Model clusters CH SC GMM 10 183.14 0.057 K-means++ 10 182.79 0.054 DBSCAN 5 132.77 –0.028 11 67.06 –0.031 Spectral clustering 10 152.37 0.044 BIRCH 10 151.63 0.035 AHC 10 151.05 0.034
下载: 导出CSV
-
[1] KEIKA K, NAKAMURA R, BAUMJOHANN W, et al. Substorm expansion triggered by a sudden impulse front propagating from the dayside magnetopause[J]. Journal of Geophysical Research: Space Physics, 2009, 114(52): A00C24 [2] BOUDOURIDIS A, ZESTA E, LYONS R, et al. Effect of solar wind pressure pulses on the size and strength of the auroral oval[J]. Journal of Geophysical Research: Space Physics, 2003, 108(A4): 8012 doi: 10.1029/2002JA009373 [3] YANG Q J, WANG Y Y, REN J. Auroral image classification with very limited labeled data using few-shot learning[J]. IEEE Geoscience and Remote Sensing Letters, 2022, 19: 6506805 [4] KVAMMEN A, WICKSTRØM K, MCKAY D, et al. Auroral image classification with deep neural networks[J]. Journal of Geophysical Research: Space Physics, 2020, 125(10): e2020JA027808 [5] YANG Q J, ZHOU P H. Representation and classification of auroral images based on convolutional neural networks[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020, 13: 523-534 doi: 10.1109/JSTARS.2020.2969245 [6] AKASOFU S I. Auroral morphology: a historical account and major auroral features during auroral substorms[M]//KEILING A, DONOVAN E, BAGENAL F, et al. Auroral Phenomenology and Magnetospheric Processes: Earth and Other Planets. Washington DC: American Geophysical Union, 2012: 29-38 [7] HENDERSON M G. Auroral substorms, poleward boundary activations, auroral streamers, omega bands, and onset precursor activity[M]//KEILING A, DONOVAN E, BAGENAL F, et al. Auroral Phenomenology and Magnetospheric Processes: Earth and Other Planets. Washington DC: American Geophysical Union, 2012: 39-54 [8] GRODENT D, BONFOND B, YAO Z, et al. Jupiter’s aurora observed with HST during Juno orbits 3 to 7[J]. Journal of Geophysical Research: Space Physics, 2018, 123(5): 3299-3319 doi: 10.1002/2017JA025046 [9] NICHOLS J D, KAMRAN A, MILAN S E. Machine learning analysis of Jupiter’s far‐ultraviolet auroral morphology[J]. Journal of Geophysical Research: Space Physics, 2019, 124(11): 8884-8892 doi: 10.1029/2019JA027120 [10] HE K M, FAN H Q, WU Y X, et al. Momentum contrast for unsupervised visual representation learning[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle: IEEE, 2020 [11] HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[C]//IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas: IEEE, 2016 [12] CHEN X L, FAN H Q, GIRSHICK R, et al. Improved baselines with momentum contrastive learning[OL]. arXiv preprint arXiv: 2003.04297, 2020 [13] MCLACHLAN G J, BASFORD K E. Mixture models: inference and applications to clustering[M]. New York: Marcel Dekker, 1988 [14] TIBSHIRANI R, WALTHER G, HASTIE T. Estimating the number of clusters in a data set via the gap statistic[J]. Journal of the Royal Statistical Society Series B (Statistical Methodology) , 2001, 63(2): 411-423 doi: 10.1111/1467-9868.00293 [15] 胡泽骏, 韩冰, 连慧芳. 基于广义回归神经网络的行星际/太阳风参数和地磁指数的紫外极光强度建模[J]. 地球物理学报, 2020, 63(5): 1738-1750HU Zejun, HAN Bing, LIAN Huifang. Modeling of ultraviolet auroral intensity based on Generalized Regression Neural Network associated with IMF/solar wind and geomagnetic parameters[J]. Chinese Journal of Geophysics, 2020, 63(5): 1738-1750 [16] 韩冰, 连慧芳, 胡泽骏. 基于神经网络模型的紫外极光卵边界建模[J]. 中国科学: 技术科学, 2019, 49(5): 531-542 doi: 10.1360/N092018-00227HAN Bing, LIAN Huifang, HU Zejun. Modeling of ultraviolet auroral oval boundaries based on neural network technology[J]. Scientia Sinica Technologica, 2019, 49(5): 531-542 doi: 10.1360/N092018-00227 [17] 王倩, 孟庆虎, 胡泽骏, 等. 紫外极光图像极光卵提取方法及其评估[J]. 极地研究, 2011, 23(3): 168-177WANG Qian, MENG Qinghu, HU Zejun, et al. A method for extracting auroral ovals in UVI images and its evaluation[J]. Chinese Journal of Polar Research, 2011, 23(3): 168-177 [18] SIMONYAN K, ZISSERMAN A. Very deep convolutional networks for large-scale image recognition[C]//3 rd International Conference on Learning Representations. San Diego: ICLR, 2014 [19] BOYKOV Y Y, JOLLY M P. Interactive graph cuts for optimal boundary & region segmentation of objects in N-D images[C]//Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001. Vancouver: IEEE, 2001 [20] 王梓涵, 佟继周, 邹自明, 等. 基于U-net的紫外极光观测极光卵形态提取[J]. 空间科学学报, 2021, 41(4): 667-675 doi: 10.11728/cjss2021.04.667WANG Zihan, TONG Jizhou, ZOU Ziming, et al. Auroral oval morphology extraction based on u-net from ultraviolet aurora observation[J]. Chinese Journal of Space Science, 2021, 41(4): 667-675 doi: 10.11728/cjss2021.04.667 [21] VAN DER MAATEN L, HINTON G. Visualizing data using t-SNE[J]. Journal of Machine Learning Research, 2008, 9(11): 2579-2605 [22] RONEN M, FINDER S E, FREIFELD O. DeepDPM: Deep clustering with an unknown number of clusters[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New Orleans: IEEE, 2022: 9851-9860 -
下载:
点击查看大图
计量
- 文章访问数: 21
- HTML全文浏览量: 12
- PDF下载量: 3
- 被引次数: 0
