5:15 PM - 6:45 PM
[U16-P16] Estimation of aftershock extent just after a large earthquake using Kernel Density Estimation: Application to the 2024 Noto Peninsula earthquake
Keywords:Aftershock extent, Kernel Density Estimation, The 2024 Noto Peninsula earthquake
When a large earthquake occurs, aftershocks that become active around the mainshock cause further shaking in the area that has already been damaged by strong motions due to the mainshock. In the initial phase immediately after the mainshock, when disaster mitigation activities such as evacuation, rescue, and emergency response are carried out, there is a high need for information on the spatial extent of aftershocks. This study proposes a method to estimate the aftershock extent within a few hours after the mainshock occurrence by applying Kernel Density Estimation (KDE) to earthquake catalog data. Immediately after the mainshock, the quality and quantity of earthquake catalog data become incomplete due to the effect of the mainshock and active aftershocks. Furthermore, the source information available immediately after a large earthquake is that obtained by automatic processing, but its accuracy is generally less than that obtained by manual processing. Thus, there is a need to estimate the extent of the aftershock stably, even under such difficult circumstances.
As input, we use a 2D point cloud distribution composed of the horizontal positions (latitude and longitude) of earthquakes. As a preprocessing step, the latitude and longitude of the earthquake catalog are converted to a Cartesian coordinate system with the origin at the position of the mainshock's epicenter and a unit of 100 km. Then, we apply KDE to the 2D point cloud distribution to estimate a 2D probability density distribution nonparametrically. We use the SciPy library for the KDE algorithm, utilize a Gaussian function as the kernel function, and adopt Scott's rule for selecting the kernel bandwidth. Finally, areas with a probability density of 0.5 or higher in the obtained 2D probability density distribution are extracted as the aftershock extent. Because the probability density distribution may have multiple peaks or the area with a probability density of 0.5 or higher may be too wide, we also considered a method in which a 2D Gaussian distribution with rotation is fitted to the 2D probability density distribution obtained by KDE, and then the area with a probability density of 0.5 or higher is extracted from the 2D Gaussian distribution as the aftershock extent.
This proposed method was applied to the Hi-net automatic processing earthquake catalog for aftershocks of the 2024 Noto Peninsula earthquake. The aftershock area of this earthquake is known to extend from the midpoint between the Noto Peninsula and Sado Island along the northeast direction to the west coast of the Noto Peninsula along the southwest direction, with a length of approximately 150 km. When applying the proposed method to the earthquake catalog 12 hours after the mainshock (with 242 events), the extracted aftershock extent corresponds to the known aftershock area. When applied to the earthquake catalog 6 hours after the mainshock (with 98 events), the obtained aftershock extent is close to the result at 12 hours but with less northeastward extension. When applied to the earthquake catalog 3 hours after the mainshock (with 38 events), the extracted aftershock extent was too spatially widespread due to the small number of events. However, the estimated aftershock extent covered the entire Noto Peninsula, indicating that the occurrence of seismic activation over the entire peninsula can be detected by the proposed method as early as 3 hours after the mainshock. We also found that the aftershock extent based on the 2D Gaussian distribution was close to the aftershock extent based on the probability density distribution obtained by KDE, and that the probability density distribution obtained by KDE contained information on the spatial variability of the aftershock activity.
As input, we use a 2D point cloud distribution composed of the horizontal positions (latitude and longitude) of earthquakes. As a preprocessing step, the latitude and longitude of the earthquake catalog are converted to a Cartesian coordinate system with the origin at the position of the mainshock's epicenter and a unit of 100 km. Then, we apply KDE to the 2D point cloud distribution to estimate a 2D probability density distribution nonparametrically. We use the SciPy library for the KDE algorithm, utilize a Gaussian function as the kernel function, and adopt Scott's rule for selecting the kernel bandwidth. Finally, areas with a probability density of 0.5 or higher in the obtained 2D probability density distribution are extracted as the aftershock extent. Because the probability density distribution may have multiple peaks or the area with a probability density of 0.5 or higher may be too wide, we also considered a method in which a 2D Gaussian distribution with rotation is fitted to the 2D probability density distribution obtained by KDE, and then the area with a probability density of 0.5 or higher is extracted from the 2D Gaussian distribution as the aftershock extent.
This proposed method was applied to the Hi-net automatic processing earthquake catalog for aftershocks of the 2024 Noto Peninsula earthquake. The aftershock area of this earthquake is known to extend from the midpoint between the Noto Peninsula and Sado Island along the northeast direction to the west coast of the Noto Peninsula along the southwest direction, with a length of approximately 150 km. When applying the proposed method to the earthquake catalog 12 hours after the mainshock (with 242 events), the extracted aftershock extent corresponds to the known aftershock area. When applied to the earthquake catalog 6 hours after the mainshock (with 98 events), the obtained aftershock extent is close to the result at 12 hours but with less northeastward extension. When applied to the earthquake catalog 3 hours after the mainshock (with 38 events), the extracted aftershock extent was too spatially widespread due to the small number of events. However, the estimated aftershock extent covered the entire Noto Peninsula, indicating that the occurrence of seismic activation over the entire peninsula can be detected by the proposed method as early as 3 hours after the mainshock. We also found that the aftershock extent based on the 2D Gaussian distribution was close to the aftershock extent based on the probability density distribution obtained by KDE, and that the probability density distribution obtained by KDE contained information on the spatial variability of the aftershock activity.