[S-VC29] Monitoring and assessment of volcanic activities

[SVC29-09] Automatic identification of volcanic ash components with AI technology for promoting a real-time monitoring of eruption

*Nobuo Geshi¹, Keiko Matsumoto¹, Ayumu Miyakawa², Takuya Itaki² (1.Research Institute of Earthquake and Volcano Geology, The National Institute of Advanced Industrial Science and Technology, 2.Research Institute of Geology and Geoinformation, The National Institute of Advanced Industrial Science and Technology)

Keywords:Volcano Ash, Eruption Monitoring, Automatic Image Identification

Volcanic ash particles provide information about the physical properties and mechanical behavior of the magma driving explosive eruption. Volcanic ash particles are produced in explosive eruptions such as magmatic and phreatomagmatic eruptions, and disperse and fall relatively quickly to distant areas, they can be collected at points far from the vent and are suitable samples for tracing changes in eruption mechanisms along the eruption process. However, it is difficult to process a large number of samples quickly along the eruption because it is necessary to identify the constituent particles and their characteristics with the naked eye using an optical microscope after collecting and processing the ejecta to make them observable. In addition, the information obtained is biased by the experience and skill of the observer, making the objectivity of the information problematic. To solve this problem, we applied the recently developed automatic image identification technology to the feature analysis of volcanic ash particles, and developed a technique for rapid analysis of a large volume of samples. We attempted to classify volcanic ash particles that had been washed and sieved using an optical microscope by using a transition learning model based on deep learning. This enabled us to appraise much larger number of particle speedy than that of conventional manual method. We applied the identification model to the Aso 2014-15 and Sakurajima 2017-2022 ejecta, and succeeded in extracting changes in the ash component particle ratios that reflect volcanic activity transitions captured by surface phenomena. Furthermore, we developed a system for pre-processing of samples in the field to transfer the obtained microscope images for automatic grain analysis. Although particle analysis by automatic image recognition still has issues such as the setting of teacher data and the method of appraising ambiguous particles, it is a promising method for screening the minimum information necessary for eruption monitoring and for detecting anomalies.