[Paper Review] A CNN-BiLSTM Model with Attention Mechanism for Earthquake Prediction

Parisa Kavianpour, Mohammadreza Kavianpour+2 more|arXiv (Cornell University)|Dec 26, 2021

Earthquake Detection and Analysis25 citations

TL;DR

This paper proposes a CNN-BiLSTM model with an attention mechanism to predict the number and maximum magnitude of earthquakes monthly in Mainland China, using ZOH preprocessing and nine regional divisions, and compares it against several baseline methods.

ABSTRACT

Earthquakes, as natural phenomena, have continuously caused damage and loss of human life historically. Earthquake prediction is an essential aspect of any society's plans and can increase public preparedness and reduce damage to a great extent. Nevertheless, due to the stochastic character of earthquakes and the challenge of achieving an efficient and dependable model for earthquake prediction, efforts have been insufficient thus far, and new methods are required to solve this problem. Aware of these issues, this paper proposes a novel prediction method based on attention mechanism (AM), convolution neural network (CNN), and bi-directional long short-term memory (BiLSTM) models, which can predict the number and maximum magnitude of earthquakes in each area of mainland China-based on the earthquake catalog of the region. This model takes advantage of LSTM and CNN with an attention mechanism to better focus on effective earthquake characteristics and produce more accurate predictions. Firstly, the zero-order hold technique is applied as pre-processing on earthquake data, making the model's input data more proper. Secondly, to effectively use spatial information and reduce dimensions of input data, the CNN is used to capture the spatial dependencies between earthquake data. Thirdly, the Bi-LSTM layer is employed to capture the temporal dependencies. Fourthly, the AM layer is introduced to highlight its important features to achieve better prediction performance. The results show that the proposed method has better performance and generalize ability than other prediction methods.

Motivation & Objective

• Motivate the need for accurate, timely earthquake predictions to reduce damage and casualties.
• Develop a hybrid CNN-BiLSTM-AM architecture to capture spatial and temporal dependencies in regional seismic data.
• Incorporate an attention mechanism to emphasize influential features for improved prediction.
• Preprocess data with zero-order hold to enhance model robustness and generalization.
• Evaluate the model on monthly earthquake data for Mainland China divided into nine sub-regions and compare with baseline methods.

Proposed method

• Use a one-dimensional CNN to extract spatial features from monthly earthquake data.
• Apply BiLSTM layers to learn temporal dependencies in the extracted features.
• Incorporate an attention mechanism to weight influential features for prediction.
• Preprocess data with zero-order hold to mitigate zero-value effects.
• Use fully connected layers to generate final predictions for monthly counts and magnitudes.
• Train with Adam optimizer using MSE loss, 150 epochs, and 32 batch size, with data normalized by Min-Max scaling.

Experimental results

Research questions

• RQ1 Can the CNN-BiLSTM-AM model simultaneously predict the monthly number of earthquakes and the maximum monthly magnitude?
• RQ2 Does the attention mechanism improve prediction accuracy by highlighting important features?
• RQ3 What is the impact of zero-order hold preprocessing on model performance and generalization?
• RQ4 How does partitioning Mainland China into nine regions affect prediction capability and localization of future events?
• RQ5 How does the proposed method compare to shallow ML and other DL models in predicting earthquake statistics?

Key findings

• The proposed CNN-BiLSTM-AM method yields better performance and generalization ability than the compared methods.
• The model architecture effectively extracts spatial-temporal features through CNN and BiLSTM components and emphasizes important features via attention.
• ZOH preprocessing improves the training robustness and prediction accuracy.
• Nine regional divisions are used to assess regional prediction ranges for the next earthquake event.
• Compared models include SVM, DT, RF, MLP, CNN, LSTM, and CNN-BiLSTM, with the proposed method evaluated against these baselines.