5:15 PM - 6:45 PM
[SCG50-P05] Attempt to detect tsunami magnetic field using machine learning
Keywords:Tsunami electromagnetic fields, Machine learning
Tsunami magnetic fields arise when conductive sea water moves within the Earth's main magnetic field as a tsunami wave (e.g., Tyler, 2005). It has been revealed that tsunami electromagnetic fields are observed prior to the arrival of tsunami waves, providing valuable information about tsunami wave heights and the direction of tsunami propagation(e.g., Lin et al. 2021).Therefore, this phenomenon is expected to be applicable to tsunami early warning systems. However, there are some challenges; the observation of tsunami electromagnetic fields is restricted to significant tsunami events because the signal-to-noise ratio must be sufficiently large to detect signals. Visual inspections, either in the time or frequency domain, have been conventional methods to identify tsunami electromagnetic signals(e.g., Schnepf et al. 2016), but they typically require much time and effort.
To address these challenges, we employ machine learning to detect tsunami electromagnetic fields.
As a first step, we focus on building machine learning models to determine whether time series magnetic data includes tsunami electromagnetic fields or not. We select supervised learning, which is one of the machine learning methods and prepare a large amount of training data. The training data consists of two types: one includes tsunami magnetic components, and the other does not. We calculate the tsunami magnetic components by a numerical simulation method for tsunami electromagnetic fields (Minami et al. 2017), while we use seafloor magnetic data not involved in any tsunami events as background noise for simulated signal or the data not including tsunami signals. We associate these two types of data with their corresponding answer labels, and then we set them as training datasets. The answer labels are scalar values of 0 or 1; we set 1 for tsunami magnetic data and 0 for non-tsunami magnetic data. We feed these datasets into the machine learning model and conduct training iterations. In the 1D-CNN model, a high accuracy of 85% was recorded on the test datasets, generated separately from the training datasets to validate the model's performance. Accuracy, in this context, represents the agreement rate between the input data and the answer labels. In this research, we input the real observation data into our machine learning models and assess their performance. Here, we develop two types of machine learning models for two data sampling rates: 1-minute for observations in the Philippines Sea and 2-minute for the observation at the northwest Pacific(NWP). The real observation data contains tsunami electromagnetic fields associated with the 2006 and 2007 Kuril Islands earthquakes. As a result, our model successfully detected tsunami magnetic signals within the data at NWP corresponding to the tsunami arrival. In this presentation, we provide a more detailed report of our model's performance, focusing on the signal-to-noise ratio.
To address these challenges, we employ machine learning to detect tsunami electromagnetic fields.
As a first step, we focus on building machine learning models to determine whether time series magnetic data includes tsunami electromagnetic fields or not. We select supervised learning, which is one of the machine learning methods and prepare a large amount of training data. The training data consists of two types: one includes tsunami magnetic components, and the other does not. We calculate the tsunami magnetic components by a numerical simulation method for tsunami electromagnetic fields (Minami et al. 2017), while we use seafloor magnetic data not involved in any tsunami events as background noise for simulated signal or the data not including tsunami signals. We associate these two types of data with their corresponding answer labels, and then we set them as training datasets. The answer labels are scalar values of 0 or 1; we set 1 for tsunami magnetic data and 0 for non-tsunami magnetic data. We feed these datasets into the machine learning model and conduct training iterations. In the 1D-CNN model, a high accuracy of 85% was recorded on the test datasets, generated separately from the training datasets to validate the model's performance. Accuracy, in this context, represents the agreement rate between the input data and the answer labels. In this research, we input the real observation data into our machine learning models and assess their performance. Here, we develop two types of machine learning models for two data sampling rates: 1-minute for observations in the Philippines Sea and 2-minute for the observation at the northwest Pacific(NWP). The real observation data contains tsunami electromagnetic fields associated with the 2006 and 2007 Kuril Islands earthquakes. As a result, our model successfully detected tsunami magnetic signals within the data at NWP corresponding to the tsunami arrival. In this presentation, we provide a more detailed report of our model's performance, focusing on the signal-to-noise ratio.