Abstract Mount Erebus, Antarctica, has a persistent lava lake with Strombolian eruptions. Volcanic eruptions can be automatically detected with multiple methods such as cross-correlation of seismic recordings and identifying anomalies in gas emissions. We demonstrate a new method of detecting Strombolian eruptions by training a convolutional neural network to automatically categorize eruptions in infrared images obtained from the rim of the crater above the Ray lava lake atop Mount Erebus. Over 9 million images were obtained from previous research ( Peters et al., 2014a ). The infrared images were shot during a span of over 2 years; one image every 2 s when weather conditions did not hinder the electrical supply. Training was performed with infrared images of the eruptions detected through seismic cross-correlation with a stacked waveform. Eruptions detected using machine learning on infrared images from December 2013 through December 2014 correctly categorized 84% of the detections as eruptions. The remaining 16% were caused by effects from the plume confounding the neural network. Nearly all of the eruptions detected utilizing seismic cross-correlation were also categorized correctly by the neural network during periods for which image data was available. We concluded machine learning is an effective method for classifying the characteristics of Strombolian eruptions which further improves the ability to study their origins while assessing the hazards posed by volcanic eruptions.

中文翻译：

---

机器学习作为南极埃里伯斯山红外图像中斯特龙博利火山喷发的检测方法

摘要 南极洲的埃里伯斯山有一个持续不断的熔岩湖，有斯特龙博利安火山喷发。可以使用多种方法自动检测火山喷发，例如地震记录的互相关和识别气体排放异常。我们展示了一种检测 Strombolian 火山喷发的新方法，通过训练卷积神经网络自动对从埃里伯斯山山顶雷熔岩湖上方的火山口边缘获得的红外图像中的火山喷发进行分类。从之前的研究中获得了超过 900 万张图像（Peters 等，2014a）。红外图像是在两年多的时间内拍摄的；当天气条件不妨碍供电时，每 2 秒拍摄一张图像。训练是使用通过叠加波形的地震互相关检测到的火山喷发的红外图像进行的。使用机器学习对 2013 年 12 月至 2014 年 12 月的红外图像检测到的喷发正确地将 84% 的检测归类为喷发。其余 16% 是由羽流混淆神经网络的影响引起的。在图像数据可用期间，神经网络也对几乎所有利用地震互相关检测到的火山喷发进行了正确分类。我们得出结论，机器学习是对斯特龙博利火山喷发特征进行分类的有效方法，可进一步提高研究其起源的能力，同时评估火山喷发造成的危害。其余 16% 是由羽流混淆神经网络的影响引起的。在图像数据可用期间，神经网络也对几乎所有利用地震互相关检测到的火山喷发进行了正确分类。我们得出结论，机器学习是对斯特龙博利火山喷发特征进行分类的有效方法，可进一步提高研究其起源的能力，同时评估火山喷发造成的危害。其余 16% 是由羽流混淆神经网络的影响引起的。在图像数据可用期间，神经网络也对几乎所有利用地震互相关检测到的火山喷发进行了正确分类。我们得出结论，机器学习是对斯特龙博利火山喷发特征进行分类的有效方法，可进一步提高研究其起源的能力，同时评估火山喷发造成的危害。