A deep learning framework that predicts both the maximum magnitude and the number of earthquakes in the coming month, designed for the nonlinear, chaotic nature of seismic data. It pairs a CNN-BiLSTM-attention network with a novel Zero-Order-Hold (ZOH) preprocessing step, validated across nine seismic regions of China.
📄 Paper: A CNN-BiLSTM model with attention mechanism for earthquake prediction The Journal of Supercomputing 79 (2023) 19194–19226 · doi:10.1007/s11227-023-05369-y
👥 Parisa Kavianpour, Mohammadreza Kavianpour, Ehsan Jahani, Amin Ramezani
🌟 Highly cited — this work has accumulated 200+ citations, making it a well-recognized reference for deep-learning-based earthquake prediction.
Earthquakes are stochastic and chaotic, which makes them notoriously hard to forecast — and seismic time series are full of "empty" months that confuse a model during training. This work tackles both problems:
- Spatio-temporal feature learning. A CNN extracts spatial features, feeding a bi-directional LSTM (BiLSTM) that captures temporal patterns in both directions, so the model learns the shape of seismic history rather than treating months in isolation.
- Attention mechanism (AM). An attention layer re-weights the BiLSTM's hidden states so the most informative time steps dominate the prediction.
- Zero-Order-Hold (ZOH) preprocessing. A novel idea borrowed from missing-data imputation: the many "zero" (no-earthquake) months that destabilize training are replaced by holding the last nonzero value — reducing the effect of seismic complexity and sharpening the forecasts.
The model predicts two targets a month ahead — maximum magnitude and earthquake count — and is tested on a 50-year catalog across nine regions of China, beating shallow ML (SVM, DT, MLP, RF) and deep baselines (CNN, LSTM, CNN-BiLSTM).
⚠️ Status — documentation & resources only. This repository currently hosts the method overview, problem framing, figures, and dataset pointers. The training code is not yet public. See Roadmap.
Five blocks. Input applies ZOH preprocessing → a feature-extraction CNN (4 conv layers + flatten) → a sequence-learning block (2 BiLSTM layers with dropout) → an attention block → a prediction block (3 FC layers) that outputs the forecast.
A detailed walkthrough lives in docs/method.md.
| Challenge | What goes wrong | How this model responds |
|---|---|---|
| Chaotic, nonlinear signal | Earthquakes are random and hard to pattern; shallow ML extracts only surface features. | CNN + BiLSTM learn deep spatial and bidirectional temporal structure. |
| Many "empty" months | Zero-valued months mislead training and inflate error. | ZOH preprocessing holds the last nonzero value, easing training. |
| Not all features matter equally | Plain sequence models weight every step the same and miss critical moments. | Attention assigns higher weight to the most influential hidden states. |
Full discussion in docs/challenges.md.
Seismic months with no qualifying earthquake produce zeros that aren't part of the underlying behavior, yet they dominate training and hurt accuracy. Inspired by imputation in missing-data problems, ZOH replaces each zero with the last nonzero value from the prior month — applied only to training data, never the test set. The ablations show ZOH contributes more than attention to the final accuracy, and clearly beats mean- or max-substitution baselines.
Most prior work forecasts only time / location / magnitude. This study also predicts the number of earthquakes per month — a signal that helps paint a fuller picture of a region's seismicity — as a separate case study alongside maximum-magnitude prediction.
Mainland China (one of the world's most seismically active areas) split into nine geologically distinct regions, each trained separately. Points are real catalog earthquakes (1966–2021, magnitude ≥ 3.5, 11,442 events).
Ablation on the R² metric across all nine regions. The full model (yellow) dominates CNN-BiLSTM and its partial variants everywhere — showing ZOH and attention each add value, region by region.
- Maximum-magnitude prediction: averaged over nine regions, the model reaches R² ≈ 0.906 with RMSE/MAE of 0.074 / 0.076, versus the best baseline (CNN-BiLSTM) at R² ≈ 0.445 — and far ahead of shallow ML (SVM lowest).
- Earthquake-count prediction: improves on the CNN-BiLSTM baseline across RMSE, MAE, and R², driven by the ZOH + attention additions.
- Generalization: holds up even on a shrunken sub-region with a shifted data distribution, where baselines degrade — evidence of robustness.
- Best vs. worst region: on a tough region the model still hits R² ≈ 0.801 where a decision-tree baseline collapses to ≈ 0.111.
See docs/method.md for the metrics and training setup.
A 50-year seismic catalog of mainland China; see docs/datasets.md.
No data is redistributed here.
| Property | Value |
|---|---|
| Sources | USGS + National Seismological Center (NSC) |
| Span | Jan 1966 – May 2021, magnitude ≥ 3.5 |
| Size | 11,442 events → 665 monthly samples |
| Regions | 9 (lat 23–45°, lon 75–119°, split 3×3) |
| Targets | max magnitude/month; earthquake count/month |
.
├── README.md # you are here
├── assets/ # figures extracted from the paper
├── docs/
│ ├── method.md # CNN, BiLSTM, attention, ZOH, training setup
│ ├── challenges.md # the three challenges, in depth
│ └── datasets.md # catalog, regions, preprocessing, two case studies
├── CITATION.cff # machine-readable citation
└── LICENSE # docs/figures license
- Method overview, problem framing, and figures
- Dataset documentation and region/preprocessing design
- Reference implementation (Keras/PyTorch) of CNN-BiLSTM-AM
- ZOH preprocessing utility
- Catalog-download + region-splitting scripts
- Reproduction scripts for both case studies
If you find this work useful, please cite the paper (see CITATION.cff):
@article{kavianpour2023cnnbilstm,
title = {A CNN-BiLSTM model with attention mechanism for earthquake prediction},
author = {Kavianpour, Parisa and Kavianpour, Mohammadreza and
Jahani, Ehsan and Ramezani, Amin},
journal = {The Journal of Supercomputing},
volume = {79},
pages = {19194--19226},
year = {2023},
doi = {10.1007/s11227-023-05369-y}
}Text and documentation in this repository are released under the terms in
LICENSE. The figures in assets/ are reproduced from the published
article (© 2023 The Author(s), under licence to Springer Nature) and are included
here solely to document and explain the authors' own work; all rights remain with
the copyright holder.