5:15 PM - 7:15 PM
[MIS11-P10] Inverse analysis of tsunami deposits using deep learning: validation of estimation accuracy for moderate-scale tsunamis
Keywords:tsunami deposit, inverse analysis, machine learning, the 2024 Noto Peninsula earthquake
The inverse analysis method for tsunami deposits aims to reconstruct hydraulic conditions based on deposit thickness and grain size distribution. This method is particularly effective for estimating the size of tsunamis with limited observational and historical records. Recently, a deep learning-based method, FITTNUSS-DNN (Mitra et al., 2020), has been proposed for the quantitative estimation of tsunami hydraulic conditions. This method has been successfully validated for large-scale tsunamis, such as the 2011 Tohoku-oki tsunami and the 2004 Indian Ocean tsunami, through comparisons with observational records (Mitra et al., 2020, 2021, 2024). However, its application to moderate-scale tsunami has been limited to a single case study of the 2006 Java tsunami (Batubo et al., 2024). Compared to large-scale tsunamis, moderate-scale tsunamis have smaller wave heights and shorter inundation distances and are more susceptible to topographic effects. The Java tsunami case study by Batubo et al. (2024) confirmed a 9.5% underestimation of inundation distance. Therefore, particularly in areas with significant topographic relief or dense infrastructure, FITTNUSS-DNN's reconstructed values, which assume flat topography, may deviate from actual hydraulic conditions. This study compares the result of the inverse analysis of tsunami deposits formed by the 2024 Noto Peninsula earthquake with the tsunami traces observed in the field surveys, aiming at evaluating the accuracy of the method for the moderate-scale tsunamis.
In the tsunami deposit survey, to obtain tsunami deposit data for inverse analysis, we established two transects perpendicular to the shoreline in the Nunoura area of Noto Town, Ishikawa Prefecture: Transect A and Transect B. Along each transect, we conducted topographic surveys and performed excavation and sampling of tsunami deposits. Grain size analysis of samples from each sampling site was conducted using a Retsch Technolog CAMSIZER. Tsunami deposits with thicknesses ranging from 5.6–1.9 cm were observed along Transect A, and from 7.8–0.7 cm along Transect B. Both transects exhibited subunit separated by accumulated plant fragments, suggesting multiple wave pulses or separate inundation events.
The inverse analysis involved training the FITTNUSS-DNN model with artificially generated training data covering the expected range of hydraulic conditions in the study area. The model's loss function evaluation showed comparable performance to previous studies. (e.g., Mitra et al., 2020; Batubo et al., 2024). The analysis reconstructed the following conditions: for transect A, inundation distance of 293±1.95 m, flow velocity of 1.70±0.02 m/s, and inundation depth of 1.27±0.00 m; for transect B, 326±16.2 m, 3.31±0.13 m/s, and 1.29±0.02 m, respectively.
We compared the accuracy of inundation distance with previous studies as a reliability assessment of the inverse analysis method. For large-scale tsunamis, previous studies of the 2011 Tohoku-oki tsunami and the 2004 Indian Ocean tsunami showed underestimation or overestimation of reconstructed values by approximately 3–15% compared to observed values (Mitra et al., 2020, 2021, 2024). The discrepancy in inundation distances reconstructed in this study showed underestimations of 11.2% and 10.7% for survey transects A and B respectively, which is similar to the results of previous studies. This suggests that our inverse analysis method maintains a comparable level of reconstruction accuracy for moderate-scale tsunamis as it does for large-scale tsunamis. Future research should focus on identifying the factors contributing to discrepancies in reconstructed values, which could enable more accurate estimation of hydraulic conditions and tsunami size.
In the tsunami deposit survey, to obtain tsunami deposit data for inverse analysis, we established two transects perpendicular to the shoreline in the Nunoura area of Noto Town, Ishikawa Prefecture: Transect A and Transect B. Along each transect, we conducted topographic surveys and performed excavation and sampling of tsunami deposits. Grain size analysis of samples from each sampling site was conducted using a Retsch Technolog CAMSIZER. Tsunami deposits with thicknesses ranging from 5.6–1.9 cm were observed along Transect A, and from 7.8–0.7 cm along Transect B. Both transects exhibited subunit separated by accumulated plant fragments, suggesting multiple wave pulses or separate inundation events.
The inverse analysis involved training the FITTNUSS-DNN model with artificially generated training data covering the expected range of hydraulic conditions in the study area. The model's loss function evaluation showed comparable performance to previous studies. (e.g., Mitra et al., 2020; Batubo et al., 2024). The analysis reconstructed the following conditions: for transect A, inundation distance of 293±1.95 m, flow velocity of 1.70±0.02 m/s, and inundation depth of 1.27±0.00 m; for transect B, 326±16.2 m, 3.31±0.13 m/s, and 1.29±0.02 m, respectively.
We compared the accuracy of inundation distance with previous studies as a reliability assessment of the inverse analysis method. For large-scale tsunamis, previous studies of the 2011 Tohoku-oki tsunami and the 2004 Indian Ocean tsunami showed underestimation or overestimation of reconstructed values by approximately 3–15% compared to observed values (Mitra et al., 2020, 2021, 2024). The discrepancy in inundation distances reconstructed in this study showed underestimations of 11.2% and 10.7% for survey transects A and B respectively, which is similar to the results of previous studies. This suggests that our inverse analysis method maintains a comparable level of reconstruction accuracy for moderate-scale tsunamis as it does for large-scale tsunamis. Future research should focus on identifying the factors contributing to discrepancies in reconstructed values, which could enable more accurate estimation of hydraulic conditions and tsunami size.