Can Eruptions Be Predicted? Short-Term Prediction of Volcanic Eruptions via Attention-Based Long Short-Term Memory

Hiep V. Le, Tsuyoshi Murata, Masato IguchiAAAI 2020 · DOI

Abstract

Predicting volcanic eruptions remains a significant challenge in volcanology due to the non-linear nature of geological signals and the rarity of eruption events. This paper introduces VepNet, a deep learning framework utilizing an attention-based Long Short-Term Memory (LSTM) network to forecast explosive eruptions at Japan’s Sakurajima volcano. By processing high-frequency seismic and strain data, VepNet identifies critical "trigger" signals within a two-hour window to predict eruptions one hour in advance. The model achieves an Area Under the Curve (AUC) of up to 0.8665 and supports a three-level warning system designed to mitigate disaster risks while minimizing false alarms.

Introduction

Volcanic eruption prediction is often considered the "holy grail" of geosciences. While active volcanoes like Sakurajima are heavily monitored, traditional forecasting relies heavily on manual feature engineering and expert intuition to interpret surface-level signals such as ground deformation (strain) and seismic activity. These signals are complex, non-linear, and often obscured by background noise. Furthermore, the extreme class imbalance—where eruptions are rare compared to periods of dormancy—poses a significant hurdle for standard predictive models.

Overview of the VepNet prediction pipeline, from sensor data to explosive eruption prediction.

This study seeks to automate and improve short-term (one-hour) forecasting by applying modern deep learning architectures to eight years of high-frequency sensor data from Sakurajima.

Methods

The researchers developed VepNet, an end-to-end neural network architecture optimized for multivariate time-series prediction. The model employs a dual-layer LSTM stack to capture temporal dependencies and "memory" of trends over a 120-minute input window.

The VepNet architecture featuring stacked LSTM layers and a concatenation-based attention mechanism.

A key innovation is the integration of a concatenation-based attention layer. This mechanism allows the network to dynamically weight specific time steps, effectively "focusing" on the most relevant precursory signals. To address the data imbalance, the model utilizes a weighted binary cross-entropy loss function, penalizing false negatives (missed eruptions) four times more heavily than false positives to prioritize public safety.

Data and Experimental Setup

The study leverages an extensive dataset from Sakurajima volcano spanning 2009 to 2016. The input features include:

  • Strain data: Tangential and radial strain measured by strainmeters.
  • Seismic data: Seismic energy and maximum amplitude measured by seismometers.

The dataset reflects the real-world challenge of sparsity; for instance, in 2016, the ratio of eruption to non-eruption sequences was as low as 1:184. The model uses a sliding 120-minute window to predict the status of the following hour.

Results and Discussion

VepNet demonstrated superior performance over traditional machine learning models (e.g., SVM, Random Forest) and standard CNN/LSTM architectures, reaching a peak AUC of 0.8665 in 2014.

Visualization of attention scores revealed that the model successfully identified 10-minute "trigger" windows characterized by sharp strain drops and seismic peaks. These patterns were confirmed by volcano experts as well-known precursory signals, providing interpretability to the deep learning approach.

The model's output is integrated into a Three-Level Warning System (3LWS). In high-risk scenarios, the system achieved a 68.97% eruption rate (precision) in 2014. Conversely, the low-risk category maintained an extremely low eruption rate (e.g., 0.3% in 2016), effectively identifying safe periods and reducing "warning fatigue" for local populations.

Conclusion

VepNet represents a significant step toward automated, reliable volcanic forecasting. By leveraging attention mechanisms to identify subtle precursors in complex time-series data, the model provides a robust framework for short-term disaster prevention. Future work will focus on improving performance during periods of extreme data sparsity and exploring the generalizability of the architecture to other volcanic systems.

Summary updated

BibTeX
@inproceedings{le2020can,
  author = {Hiep V. Le and Tsuyoshi Murata and Masato Iguchi},
  title = {{Can Eruptions Be Predicted? Short-Term Prediction of Volcanic Eruptions via Attention-Based Long Short-Term Memory}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i08.7043},
}
Figure