An unsupervised classification scheme for seismic facies mapping of sediment input in Nankai Trough using reflection amplitudes from 2D profiles

Slip behavior on the shallowest part of the subduction zone is heavily influenced by the lithostratigraphy of the incoming plate. Coseismic slip that propagates toward the seafloor is a major tsunami hazard as seen in the 2011 Tohoku-Oki earthquake. In Nankai Trough, scientific drilling combined with seismic facies mapping has allowed scientists to characterize the lithostratigraphy of the Shikoku Basin which in turn have been linked with the geotechnical properties, structural architecture, and seismicity. Seismic reflection surveys have been conducted here for more than 25 years covering over 730 x 150 km and mapping the sediment input using the traditional visual interpretation at this scale is inefficient. Taking advantage of this large seismic dataset combined with scientific drilling to map the fault properties can help improve hazard assessments. This study presents a framework for a semi-automatic classification using 2D seismic reflection profiles acquired by different surveys. Three trench-parallel lines within the accretionary wedge acquired by three different surveys were used for the case study. The amplitudes of the western and eastern profiles were first scaled based on the seafloor reflection of the central profile. Features were extracted using a rectangular window measuring 500 m above the top of the oceanic basement with varying widths measuring 1 km, 3 km, and 5 km. A total of 121 features divided into three groups were extracted. The statistical group (13) describes the strength of reflections, the spectral group (30) describes the presence or absence of internal reflections, and textural group (78) describes the continuity of reflections. The three principal components of each group were extracted and altogether subjected to K-Means clustering with 6 clusters. The 5 km window showed the most comparable classification with visual interpretation and the consistent classification in the overlap between profiles indicate a satisfactory performance of our method. Comparing the classification with previous drilling, Cluster 0 located in the overlap between the central and eastern profile is associated with turbidites occurring in basement lows. Cluster 1 is classified as noise. Cluster 2 in the western and central profile are likely siliciclastic turbidites from the Kyushu Fan. Cluster 3 and 4 in the eastern profile appears to have no drilling analog. Cluster 5 in the central profile is associated with hemipelagic mudstones. These initial results appear promising and will be tested in larger datasets and other subduction zones in the future.