Classification of IR Images of Small Eruptions at the Erebus Volcano, Antarctica, With Zernike Moments and Support Vector Machine.

To investigate the explosive nature of the Erebus volcano, Antarctica, millions of infrared images of the Ray lava lake were analyzed to identify time and position of non‐explosive small eruptions (NESEs). We developed a new technique based on two steps: (a) feature recognition using Zernike moments of the IR images and (b) classification using Support Vector Machine (SVM) applied to Zernike moments. Measures of the performance score were at least higher than 97% via different metrics for different categorization tasks. We observed distinctly different Zernike spectrums between NESEs and no‐eruption images. In the three months with the best quality images (December 2013, December 2014, and January 2015), out of about one million images per month, 654, 405, and 3,650 NESEs were detected, respectively. Using k‐means clustering three activity regimes emerged: low (≤4 NESEs per hour), intermediate (4 <NESEs per hour <14), and intense (≥14 NESEs per hour). December 2013 and December 2014 were associated with low and medium activity regimes only, while January 2015 had mainly medium and intense activity. The frequency of large eruptions and of NESEs do not show a correlation, supporting the hypothesis of the existence of multiple deep processes at the origin of the gas release. Space analysis of NESEs shows that they emerge over the surface primarily above the conduit, but not isotropically, pointing to complex interaction with the magma lake convection. Plain Language Summary: Strombolian eruptions are well‐known explosive features of the Erebus volcano, Antarctica, which has been continuously active for at least the past 50 years. To better investigate them, we developed a new machine learning algorithm to analyze millions of infrared images of the lava lake on top of the volcano. Focusing on the months with clearer images, December 2013, December 2014, and January 2015, in each month we detected tens explosive eruptions and hundreds to thousands small non‐explosive eruption (NESEs). Time analysis showed that the periods of more intense NESEs activity were a few days long and uncorrelated with the large eruption. Most NESEs appear to emerge above the conduit, but their distribution is not homogeneously distributed around it, pointing to an interaction between the rising gas and the convecting lake. NESEs' frequency vastly changes in time, which is consistent with the hypothesis that they are formed deep into the conduit by primarily CO2 and H2O, which are dissolved in different quantities depending on temperature and depth. Our hypothesis will need to be tested against measures of the chemical signature of the NESEs. Key Points: We developed of a new machine learning approach to classify eruptive activity at Erebus based on thermal images of the lava lakeWe identified thousands of previously undetected non‐explosive small eruptions (NESEs) among three million imagesThe NESEs activity vastly changes through time and their position is non‐isotropically distributed around the position of the conduit [ABSTRACT FROM AUTHOR]