and local population (Yamaoka et al., 2016).
Automatic classification of seismic signals originating from active volcanoes (Falsaperla et al., 1996; Langer et al., 2013; Ren et al., 2020; Scarpetta et al., 2005) and video data of plume rise (Witsil & Johnson, 2020) helped to characterize and elucidate eruptions’ behavior. ML‐based analysis of time series of gas emission, gravimetric, and tilting data alerted us one day before the occurrence of flank eruptions (Brancato et al., 2019). Deep learning‐based image analysis procedures allowed us to classify volcano deformations (Anantrasirichai et al., 2018; Gaddes et al., 2019) and understand how lava flows occurred (Corradino et al., 2019), and demonstrated its applicability for automatic prediction of future volcano behavior.
In recent years, muography has achieved progress in monitoring and exploring subsurface density distributions of various geological edifices, such as active volcanoes (Lesparre et al., 2012; Lo Presti et al., 2020; Oláh et al., 2019b; Tanaka et al., 2007, 2014) or underground rock structure (Cimmino et al., 2019; Gluyas et al., 2019; Guardincerri et al., 2017; Lázaro Roche et al. 2019, Lesparre et al., 2017; Oláh et al., 2012; Oláh et al., 2018a; Schouten & Lendru, 2018; Tanaka, 2015). The application of 3D image transformation techniques (Nagahara & Miyamoto, 2018; Tanaka et al., 2010) and the integration of muography with gravimetry (Barnoud et al., 2021; Davis & Oldenburg, 2012; Nishiyama et al., 2014; Okubo & Tanaka, 2012) allow us to resolve geological edifices with a voxel size of a few tens of meters. The combination of muography with seismic monitoring allows the detection of abrupt density changes in hydrothermal systems (Le Gonidec et al., 2019). ML‐based muographic image processing is under development to predict short‐term volcano eruptions (Nomura et al., 2020).
Currently, only the Sakurajima volcano shows persistent activity among the volcanoes that are continuously monitored with muography during recent years (D'Alessandro et al., 2019; Le Gonidec et al., 2019; Lo Presti et al., 2020; Oláh et al., 2019b). The Sakurajima volcano is an andesitic composite volcano formed on the Aira caldera in Kagoshima Bay, Kyushu, Japan. The last plinian eruption occurred in 1914. A recent study has shown that the magma supply rate of the Aira caldera amassed enough magma within approx. 130 years to feed a plinian type eruption (Hickey et al., 2016). The two craters, Minamidake and Showa, have erupted explosively more than 3,000 times in the last five years (Japan Meteorological Agency, 2020). The mechanism of these vulcanian type eruptions (Iguchi et al., 2008; Kazahaya et al., 2016) is reviewed in another chapter of this monograph (Oláh & Tanaka, 2021). During these eruptions, one of the two active craters ejected aerosols and gas with a bulk volume of below 107m to a height of 1,000–5,000 m above the crater rims, throwing fragments of volcanic plug and lava bombs usually within approx. 3,000 m radius. Although sometimes the injected ash cloud reached Kagoshima City and caused difficulties to the local transport, the activity of the Sakurajima volcano usually impacts the nearby area that is continuously visited by tourists. The constant threat to this area motivates the improvement and coordination of volcano monitoring techniques. The forecasting of short‐term eruptions of the Sakurajima volcano is a good candidate for using ML techniques to automate muographic image processing. The Multi‐Wire‐Proportional‐Chamber (MWPC)‐based Muography Observation System (MMOS) (Oláh et al., 2018b, 2019a; Varga et al., 2015, 2016, 2020) of Sakurajima Muography Observatory (SMO) has acquired sufficient amounts of data since 2018 to study the feasibility and limits of forecasting short‐term, vulcanian type eruptions.
In this chapter, we focus on how ML techniques can process muographic images captured through the Sakurajima volcano for the forecasting of impending vulcanian type eruptions. Our analysis is based on recently developed methods that are applied on new data sets collected by the MMOS during the eruption episodes of the Minamidake crater occurring between October 2018 and July 2020.
4.2 MUOGRAPHIC OBSERVATION OF THE SAKURAJIMA VOLCANO
In this section, we review how the muographic images are produced and prepared for ML‐based data processing. The MWPC‐based Muography Observation System operates with ten tracking systems in the SMO at a distance of approx. 2,800 m from the active craters of the Sakurajima volcano. Currently, all tracking systems are oriented towards the Showa crater at the same site to maximize the detection acceptance and thus the number of muons which are observed with a flux of 0.02 to 0.2 1/cm2/sr/day through the kilometer‐thick crater regions. Other chapters of this monograph describe the applied detector technology (Varga et al., 2021) as well as the experimental setup and data analysis methods (Oláh & Tanaka, 2021).
Fig. 4.1 shows the data flow diagram of MMOS. The data produced by MMOS are transferred to a remote server where automated track reconstruction and data quality assurance are performed based on a combinatorial algorithm (Oláh et al., 2018b). In the next step of data processing, the track count maps (muograms) are produced after the track selection, which is based on the goodness of track fits (χ2/ndf ). These muograms are sent to the database of the International Virtual Muography Institute, Global (Oláh et al., 2019a). For this study, the flux of penetrating particles were calculated for time periods of 24 hours (00:00:00 to 23:59:59). Fig. 4.2 shows three consecutive daily muograms that were captured by the MMOS with an angular bin size of Δθ x = Δθ y ≈ 23 mrad in both horizontal and vertical directions, corresponding to a spatial resolution of 60 meters. The black rectangles indicate the three crater regions that were selected for this study.
Figure 4.1 The data flow diagram of MWPC‐based Muography Observation System.
Figure 4.2 Three muograms captured by MWPC‐based Muography Observation System, corresponding to three consecutive days. The three black rectangles denote the angular regions of Minamidake (solid line), Showa crater (dashed line), and Surface (dotted line), respectively.
4.3 A CONCEPTUALIZATION OF VOLCANO ERUPTION FORECASTING WITH MUOGRAPHY
Our study applies the concept that was developed by Nomura et al. (2020) with a few modifications due to the different experimental setup (e.g., slightly different detector orientation, improved angular resolution, increase of the number of MMOS modules from five to seven in March 2019 and from seven to ten in August 2019 during two system upgrade works, etc.): the inputs of ML techniques are fed with data extracted from seven muon flux images that were recorded on seven consecutive days. The prediction of the ML model is compared to an eruption label that is “1” if there was at least one eruption on the 8th day; otherwise the eruption label is “0.” Three regions are extracted from the daily muon flux images to measure the accuracy of eruption forecasting: (i) the activated Minamidake crater with 7×6 segments in the angular region of −0.1725 ≤ tan (θx) < –0.0115 and 0.1495 ≤ tan (θy) < 0.2875 shown by the solid rectangle in Fig. 4.2; (ii) the deactivated Showa crater with 4 × 5 segments in the angular region of –0.0115 ≤ tan (θx) < 0.0805 and 0.1265 ≤ tan (θy) < 0.2415 shown by the dashed rectangle in Fig. 4.2; and (iii) a surface region of Sakurajima volcano with 7×6 segments in the angular region of 0.1265 ≤ tan (θx) < 0.2875 and 0.0805 ≤ tan (θy) < 0.2185 shown
