RSDB - A Global Seismic Waveform Dataset from the Raspberry Shake Network for Machine Learning Applications

Seismic datasets are critical for advancing research in earthquake monitoring and machine learning (ML) applications. However,traditional datasets often rely on expensive and complex seismometers,limiting their accessibility. Through this research,we address this challenge by developing a seismic dataset based on waveforms recorded by Raspberry Shake geophones,which are cost-efficient and easy to deploy. The dataset includes detailed earthquake metadata and waveform data for about 333,000 waveforms recorded by more than 2,400 Raspberry Shake stations from January 2022 to May 2024. More than 25,000 earthquakes of magnitude greater than 3.5 are used in the development of the dataset. Each waveform lasts 120 seconds and is collected at 100 Hz,resulting in almost 11,000 hours of data. Metadata includes 35 attributes structured in a CSV file,while waveforms are stored in the HDF5 file as NumPy arrays. Metadata includes already available as well as derived features such as P and S wave arrival times,which were derived using the PhaseNet,a deep neural-network-based phase picker,originally trained on Northern California Earthquake Data Center Catalog. The Phasenet that we used for phase picking is trained on the DiTing seismological dataset. The dataset presented in this study is prepared similarly to existing seismological resources like STEAD but offers the unique advantage of leveraging affordable instruments.