SeismicX-continous Dataset

Overview

This dataset provides an AI-ready seismic waveform index constructed from continuous recordings of the Southern California Seismic Network (SCEDC), United States.

The dataset consists of 14 consecutive days of continuous seismic waveform data and is explicitly designed for model evaluation and benchmarking, rather than for model training.

To enable controlled and interpretable testing of earthquake detection and phase-picking algorithms, the dataset is intentionally composed of two contrasting seismicity regimes:

• High-activity period: 7 consecutive days containing approximately 4,000 cataloged earthquakes
• Low-activity period: 7 consecutive days containing approximately 300 cataloged earthquakes

This high/low seismicity contrast allows systematic assessment of model behavior under fundamentally different event-density conditions using the same station network and instrumentation, making the dataset particularly suitable for benchmarking detection robustness.

The dataset is primarily intended for evaluating:

• False-positive rates in earthquake detection on continuous data,
• Missed-detection rates under sparse seismicity,
• Sensitivity of detection and phase-picking models to background noise levels,
• Stability of model performance across contrasting seismic activity regimes.

This dataset is not intended for model training, as its limited temporal coverage (14 days) and intentionally biased event distribution are optimized for controlled testing and comparative benchmarking rather than for learning generalizable representations.

Data Structure

Each row in the dataset corresponds to a fixed-length waveform window extracted from continuous recordings and associated with an event–station–phase pair. Waveform windows are indexed together with event metadata, station information, and quality-control attributes to support reproducible and quantitative evaluation.

Data Splits

The dataset is organized for evaluation purposes only. Any subdivision (e.g., by time or seismicity regime) is intended to support comparative testing and sensitivity analysis, rather than training–validation workflows.

Waveform Access

Waveforms are not stored directly in the Parquet index files. Instead, each sample contains a URI and an internal path pointing to externally stored continuous waveform data (e.g., miniSEED- or TAR-based shards). This design supports scalable access to continuous data while keeping the dataset lightweight and suitable for benchmarking workflows.

Quality Control

Quality-control (QC) metrics are provided for each waveform window, including signal-to-noise ratio (SNR), clipping indicators, and phase-pick residuals. These metrics facilitate stratified evaluation and allow users to explicitly analyze detection errors and false alarms as a function of data quality.

Known Limitations

• The dataset represents a short, fixed time window (14 days) and is not suitable for training large-scale or general-purpose models.
• Earthquake occurrence is intentionally imbalanced between high- and low-activity periods, reflecting the benchmark-oriented design.
• Results obtained on this dataset should be interpreted as controlled performance tests, not as indicators of long-term operational performance.

Citation

• Title: A Deep Learning Framework for Pg/Sg/Pn/Sn Phase Picking and Its Nationwide Implementation in Mainland China
• Authors: Yuqi Cai, Ziye Yu, et al. (yuziye@cea-igp.ac.cn)
• DOI: https://doi.org/10.1029/2025JH000944