Classification of IR Images of Small Eruptions at the Erebus Volcano, Antarctica

Infrared images of the lava lake on top of the Erebus volcano show that exist both small (i) non-explosive small and (ii) explosive large eruptions. Large eruptions have detected and classified in the past through several means such as average IR luminosity and Neural Networks. We propose here a technique to detect small events using Support Vector Machine (SVM) and report the detection of hundreds of events across the year 2015. Each is dimensionally reduced to 631 Zernike moments, used as feature descriptors in the SVM classifier. Small bubbles are categorized with a performance score (Ration of corrected prediction over all small bubble samples) in the range (7392) % for different classification tasks. By identifying the position of the small eruptions, our results contribute to resolve the controversy of whether they originate from a deep source or from the convecting lake. We also investigate the time frequency of these small events and observe several stable regimes with sharp transitions from one to another.