Machine learning-based automated horizontal-to-vertical spectral ratio analysis for site characterisation and regional transferability using KiK-net data in the Kathmandu Valley

The Kathmandu Valley faces high seismic risk due to its Himalayan location and thick sediments. Traditional horizontal-to-vertical spectral ratio (HVSR) analysis relies on manual interpretation, limiting scalability. This study introduces automated HVSR, a machine learning (ML) framework automating HVSR processing for site characterisation. Using a comprehensive dataset of 2,363 earthquakes from the Kiban Kynoshin network (KiK-net), a strong-motion network in Japan, three ML models were used to predict the dominant frequency (fd) from HVSR curves, namely: random forest, support vector regression (SVR) with a linear kernel, and SVR with a radial basis function kernel. The linear SVR model demonstrated an R² value of 0.992, root mean square error of 1.54, and mean absolute error of 1.397 Hz. Bootstrap uncertainty analysis showed narrow confidence ranges for most samples, confirming stable and consistent predictions. Feature importance results indicate that the fd and mean frequency carry nearly all predictive weights, while amplitude-based parameters contribute very little. A feature space overlap assessment further demonstrated that the fd and amplification ranges in the KiK-net dataset align well with expected site response characteristics of the Kathmandu Valley. The workflow offers a practical and scalable solution for data scarce regions and provides a strong foundation for future microzonation and site response studies.