Automated Strong-Motion Record Processing via Deep Learning-Based Simultaneous P -Wave Identification and High-Pass Corner-Frequency Selection

Strong-motion record processing provides the necessary inputs for seismic hazard analysis, structural design, and risk assessment; however, low-frequency noise and baseline offsets can introduce drift in the displacement obtained through integrating acceleration. Therefore, selecting an appropriate high-pass corner frequency (fcHP) is critical. Existing automated routines can be efficient; however, they remain sensitive to the pre-event noise window, which depends on accurate identification of the P-wave arrival time (tP). This study reformulates the joint estimation of tP and fcHP as a single segmentation-based localization task. For each component, the records were processed over candidate fcHP values from 0.005 to 1.28 Hz, resampled to 1024 points, and stacked into a 256 × 1024 displacement–frequency matrix. The dataset is composed of 2380 three-component records from the Kyoshin network and Kiban–Kyoshin network and uses an earthquake-level split. Reference tP and fcHP values were obtained by visual review following a fixed protocol using a web-based tool and displacement stability criteria in the pre-event and tail windows. DeepLabv3+ with an EfficientNet-B3 encoder achieved the lowest mean localization error (5.17 pixels). For independent events, tP predictions achieved R2=0.985 for the vertical component, and fcHP predictions achieved R2=0.825 across the components. This approach has been applied to recent earthquakes in Japan.