Automatic Auroral Detection in Color All-Sky Camera Images
Automatic Auroral Detection in Color All-Sky Camera Images
复制标题
DOI：
10.1109/jstars.2014.2321433
复制
发表时间：
2014-05
影响因子：
5.5
通讯作者：
Jayasimha Rao;N. Partamies;O. Amariutei;M. Syrjäsuo;K. V. D. Sande
中科院分区：
文献类型：
--
作者：
Jayasimha Rao;N. Partamies;O. Amariutei;M. Syrjäsuo;K. V. D. Sande
每年冬天，奇迹网络中的全天相机(ASC)会以10-20 S的固定间隔拍摄夜空图像。这相当于数百万张图像，不仅需要修剪，还需要高效的极光活动探测技术。在本文中，我们描述了一种自动分类ASC图像的方法，该方法将ASC图像分为三个互不相容的类别：极光、非极光和多云。这不仅减少了要处理的数据量，而且有助于建立将磁起伏和极光活动联系起来的统计模型，帮助我们更接近预测极光活动。我们将不同的特征提取技术与支持向量机分类相结合进行了实验。比例尺不变特征变换(SIFT)特征的颜色变体，特别是对手SIFT特征，被发现比其他特征提取技术表现得更好。利用对手SIFT特征，我们能够构建一个交叉验证准确率为91%的分类模型，该模型利用时间信息和剔除离群值进行了进一步改进，使其足够准确，可用于操作数据剪枝目的。由于该问题本质上类似于场景检测，因此局部点描述特征比基于全局和基于纹理的特征描述符具有更好的性能。
Every winter, the all-sky cameras (ASCs) in the MIRACLE network take images of the night sky at regular intervals of 10-20 s. This amounts to millions of images that not only need to be pruned, but there is also a need for efficient auroral activity detection techniques. In this paper, we describe a method for performing automated classification of ASC images into three mutually exclusive classes: aurora, no aurora, and cloudy. This not only reduces the amount of data to be processed, but also facilitates in building statistical models linking the magnetic fluctuations and auroral activity helping us to get a step closer to forecasting auroral activity. We experimented with different feature extraction techniques coupled with Support Vector Machines classification. Color variants of Scale Invariant Feature Transform (SIFT) features, specifically Opponent SIFT features, were found to perform better than other feature extraction techniques. With Opponent SIFT features, we were able to build a classification model with a cross-validation accuracy of 91%, which was further improved using temporal information and elimination of outliers which makes it accurate enough for operational data pruning purposes. Since the problem is essentially similar to scene detection, local point description features perform better than global- and texture-based feature descriptors.