[S-CG46] Uncovering stress accumulation and fault strengthening of megathrust earthquakes

[SCG46-P05] Distinguishing Seismic and Landslide-Induced MTDs: Imaging and Machine Learning

*M Rene Castillo¹、Brendan Crowell¹、Jill Leonard-Pingel¹、Christine Regalla²、Amy Gough³、Uisdean Nicholson³、Rina Fukuchi⁴、Marianne Conin⁵、Patrick Fulton⁶、Jamie Kirkpatrick⁷、Shuichi Kodiara⁸、Kohtaro Ujie⁹、Lena Madea⁸、Natsumi Okutsu⁸、Nobu Eguchi⁸、IODP Expedition 405 Scientists (1.The Ohio State University, School of Earth Sciences, USA、2.Northern Arizona University, School of Earth and Sustainability, USA、3.Heriot-Watt University, Institute of GeoEnergy Engineering, UK 、4.Naruto University of Education, Graduate School of Education, Japan 、5.University of Lorraine, GeoResources Laboratory, France 、6.Cornell University, USA 、7.University of Nevada Reno, Geological Sciences and Engineering, USA、8.Japan Agency for Marine-Earth Science and Technology (JAMSTEC), Japan、9.University of Tsukuba, Graduate School of Science and Technology, Japan)

キーワード：Machine Learning, Sedimentology

The Japan Trench, a dynamic subduction zone off eastern Japan, serves as a natural laboratory for exploring submarine geohazards and their far-reaching effects. As exemplified by the catastrophic 2011 Tohoku-Oki Mw 9.1 earthquake, this region is prone to powerful megathrust events that trigger a complex cascade of secondary hazards, including tsunamis, landslides, and widespread sediment remobilization. A major objective for the International Ocean Discovery Program, IODP Expedition 405 (JTRACK) is to identify sedimentologic deposits that record past hazardous events. However, one major challenge in sedimentology and geohazard analysis is distinguishing between seismically-induced and landslide-induced mass transport deposits (MTDs). This task typically involves processing large volumes of sediment samples, which is time-consuming and subject to observer bias. However, advances in ML now enable a more detailed examination of core imagery, revealing subtle differences in sediment layers that might otherwise go unnoticed. Incorporating physical properties from logging/discrete core samples and geochemical data will enhance accuracy. This enhanced level of analysis significantly improves our ability to characterize and interpret the cores with greater accuracy and confidence.

Here, machine learning (ML) is applied to Exp 405 sediments to develop quantitative criteria from core imagery to distinguish seismically induced and landslide-induced deposits. Seismically-induced MTDs often display pervasive sediment deformation – such as convolute bedding, load casts, and micro-faulting. They also typically contain finely layered, fluidized, or homogenized intervals due to cyclic liquefaction, along with low-cohesion deformation features like contorted laminations and dewatering structures. These deposits generally have less erosive basal contacts and can be correlated across multiple core sites, aligning with known earthquake records. In contrast, landslide-induced MTDs exhibit discrete shear zones, abrupt stratigraphic contacts, and grading patterns. They also tend to be more chaotic and heterogeneous, often containing rotated or disaggregated clasts. Unlike seismic MTDs, landslide-derived deposits are often more spatially isolated and lack clear regional correlation. Emphasizing these contrasting features within core samples, we develop robust, quantitative criteria for distinguishing between these two defined MTD types using ML. The workflow involves: (1) Core imaging and artifact correction, (2) Digital feature extraction, (3) Supervised ML classification model development, and (4) Model validation against known regional event stratigraphy. A convolutional neural network (CNN) is trained on imagery comprised of TSCL and XCT datasets, classifying MTDs probabilistically. Blind test validation ensures accuracy, reducing observer subjectivity. Additionally, the CNN results will be compared with the preliminary expedition 405 results. Understanding these distinctions enhances our knowledge of mass flux within the frontal prism and sediment stability, with implications for earthquake hazard analysis and basin evolution. Through machine learning and imaging, we propose a framework for improving earthquake event reconstruction and refining predictive models for sediment stability in tectonically active regions.