Explainable Machine Learning for Earthquakes - dataset

This dataset supports the research presented in the paper "Explainable Machine Learning for Earthquakes: SHAP Interpretation of CNNs to Distinguish Seismic Spectrograms of Foreshocks and Aftershocks".

It contains seismic waveform data and corresponding RGB spectrograms derived from the 2016 Norcia seismic sequence in central Italy. The raw dataset (dset.zip) includes three-component seismic waveforms (NS, EW, Z) recorded at 100 Hz, centered around manually picked P-wave arrivals. The preprocessed dataset (dset_preprocessed.zip) contains RGB spectrograms created by stacking short-time Fourier transforms of each component into image-like representations. These RGB images are used as input for training a convolutional neural network (CNN) to classify seismic events as either foreshocks or aftershocks.

The dataset has been curated to enable both model training and explainability analysis. In particular, the RGB spectrograms are designed for compatibility with computer vision models, while accompanying code enables SHAP (SHapley Additive exPlanations) analysis of trained CNNs. The SHAP values highlight the time-frequency regions most relevant to the model’s classification decision and reveal physically interpretable patterns linked to changes in fault zone elastic properties over the seismic cycle.

This dataset may be used to:

Train deep learning models to classify foreshocks vs. aftershocks

Reproduce the preprocessing pipeline (raw → spectrogram)

Explore explainability methods (e.g., SHAP) in seismological ML

Contents:

dset.zip: Raw waveform data

dset_preprocessed.zip: RGB spectrograms ready for model input

Conference Paper (ICIAP 2025):👉 https://doi.org/10.1007/978-3-032-10185-3_25

For additional resources (trained models, SHAP tensors, code), see the associated Zenodo code repository:👉 https://doi.org/10.5281/zenodo.15747240

Or the associated Github code repository:👉 https://github.com/mich1803/CNN-Explainability-on-Earthquakes