Single-Station CNN Location Workflow and Earthquake Catalog for the 2026 Northern Hokkaido Earthquake Swarm

This archive contains the final earthquake detection catalog, trained convolutional neural network (CNN) models, catalog metadata, and analysis scripts associated with the study “Uncatalogued Micro-Seismicity Preceding the Main Phase of the 2026 Northern Hokkaido Earthquake Swarm, Revealed by a Machine Learning Workflow.” The released catalog contains 12,048 earthquakes detected using three-component waveform data from Hi-net station N.HNBH. The single-station location workflow uses CNN models to predict earthquake back-azimuth (BAZ), incidence angle (INC), and hypocentral distance. In this version of the workflow, hypocentral distance is predicted directly by a CNN using a 10-s three-component waveform together with predicted BAZ, predicted INC, and observed S-P time. The archive includes the final earthquake catalog, trained BAZ/INC/hypocentral-distance models, training and testing scripts, prediction scripts, and associated catalog metadata. Observed NIED Hi-net waveform data are not redistributed because of NIED data-use restrictions. Users wishing to rerun the waveform-dependent training or prediction procedures should obtain the required N.HNBH waveform data directly from NIED Hi-net. The actively maintained single-station CNN location workflow is available at https://github.com/likean-seis/single-station-CNN-location-workflow. The GitHub repository serves as the maintained umbrella repository for the method, whereas this Zenodo record preserves the fixed version associated with the northern Hokkaido swarm study.