Aftershock occurrence is the dominant phenomenon observed in seismicity, offering a wealth of data for the statistical learning of the earthquake nucleation process. However, physical interpretability can be biased by a number of well-known pitfalls. In this comment, we show that the study of Terakawa et al. (2020), who concluded that the spatial distribution of aftershocks is best explained by a combination of elastic strain energy and pore-fluid pressure change (ΔEFS), is beset by both confirmation bias and data leakage. Using the same data (1992 Landers mainshock), procedure (binary classification), and metrics (Area Under the Curve, AUC), we find that the simple empirical model based on aftershock-distance-to-mainshock-rupture (Mignan and Broccardo, 2019) performs to a level (AUC = 0.76) similar to the one obtained by ΔEFS (AUC = 0.75 [0.70-0.77]). Including the Big Bear rupture improves the distance-based model performance up to AUC = 0.86. From a statistical perspective, we further show that including pore-fluid pressure change, being calculated from earthquake focal mechanisms, is equivalent to implicitly propagating aftershock location information into the model. Such type of data leakage clearly improves any classifier performance.
Comment on “Elastic strain energy and pore-fluid pressure control of aftershocks” by Terakawa et al. [Earth Planet. Sci. Lett. 535 (2020) 116103] / Mignan, A.; Broccardo, M.. - In: EARTH AND PLANETARY SCIENCE LETTERS. - ISSN 0012-821X. - 544:(2020), p. 116402. [10.1016/j.epsl.2020.116402]
Comment on “Elastic strain energy and pore-fluid pressure control of aftershocks” by Terakawa et al. [Earth Planet. Sci. Lett. 535 (2020) 116103]
Broccardo M.
2020-01-01
Abstract
Aftershock occurrence is the dominant phenomenon observed in seismicity, offering a wealth of data for the statistical learning of the earthquake nucleation process. However, physical interpretability can be biased by a number of well-known pitfalls. In this comment, we show that the study of Terakawa et al. (2020), who concluded that the spatial distribution of aftershocks is best explained by a combination of elastic strain energy and pore-fluid pressure change (ΔEFS), is beset by both confirmation bias and data leakage. Using the same data (1992 Landers mainshock), procedure (binary classification), and metrics (Area Under the Curve, AUC), we find that the simple empirical model based on aftershock-distance-to-mainshock-rupture (Mignan and Broccardo, 2019) performs to a level (AUC = 0.76) similar to the one obtained by ΔEFS (AUC = 0.75 [0.70-0.77]). Including the Big Bear rupture improves the distance-based model performance up to AUC = 0.86. From a statistical perspective, we further show that including pore-fluid pressure change, being calculated from earthquake focal mechanisms, is equivalent to implicitly propagating aftershock location information into the model. Such type of data leakage clearly improves any classifier performance.File | Dimensione | Formato | |
---|---|---|---|
J19_2020_EPSL.pdf
Solo gestori archivio
Tipologia:
Versione editoriale (Publisher’s layout)
Licenza:
Tutti i diritti riservati (All rights reserved)
Dimensione
742.26 kB
Formato
Adobe PDF
|
742.26 kB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione