An Improved Random Forest-Based Method for Automatically Selecting the High-Pass Filter Cutoff Frequency of Strong-Motion Records

ABSTRACT Strong-motion records are foundational to earthquake engineering and seismology, yet pervasive low-frequency noise degrades their accuracy and usability. High-pass filtering is a standard denoising step, and its effectiveness hinges on reliably selecting the high-pass cutoff frequency (fHP). This study presents an automatic fHP selection method based on an improved random forest (RF) model that incorporates tree-wise weighted voting and a filtering-completion criterion to enhance accuracy and robustness. Trained on 10,636 strong-motion records from the China flatfile and benchmarked against standard machine learning baseline models (RF, support vector machine, light gradient boosting machine, K-nearest neighbors, and extreme gradient boosting), the proposed method achieved superior performance. Further comparisons with Guo et al. (2025) on overlapping China datasets showed that the proposed method exhibited closer agreement with manual selections. Validation with the 2021 Ms 6.4 Yangbi earthquake sequence confirmed that the proposed method is applicable and robust across different events. Comparisons with the methods of Ramos-Sepúlveda et al. (2025) and Xu et al. (2018) further showed that all three automatic methods produce generally reasonable cutoff-frequency estimates, although displacement-sensitive and long-period measures such as peak ground displacement and pseudospectral acceleration at 10 s are more affected by differences in cutoff-selection preferences. The method was further evaluated on the Engineering Strong-Motion dataset using 7326 records under two data-splitting strategies (record-wise and event-wise). The record-wise split yields slightly higher scores due to event information leakage, whereas the event-wise split provides a stricter and more realistic test of generalization across unseen events. Despite this stricter setting, the model still achieves consistently high performance with residuals narrowly centered around zero, demonstrating good applicability across events and datasets. Overall, the proposed method provides an efficient, accurate, and scalable solution for automated preprocessing of strong-motion data.