Rapid Earthquake Magnitude Prediction

Overview

Developed a real-time earthquake magnitude prediction system using supervised learning techniques on seismic data from Kandilli Observatory. Focused on early P-wave signals for rapid and accurate magnitude estimation to support early warning systems.

Tools

• Python, Pandas, NumPy, scikit-learn, PyTorch

Dataset

• Earthquake waveform data (2004–2024) from Kandilli
• Broadband vertical component (BHZ)
• Preprocessing: SNR filtering, demeaning, P-wave picking with STA/LTA

Features

• Waveform statistics (mean, std, kurtosis, skewness)
• FFT coefficients
• Peaks and troughs
• Autocorrelation signals

Modeling

• Tested Linear Regression, Decision Tree, Gradient Boosting, Random Forest
• Random Forest outperformed others with better generalization and non-linear pattern capture

Data Imbalance

• High-magnitude events were rare
• Used oversampling (SMOTE) and undersampling
• Evaluated using F1, Precision-Recall, AUPRC

Window Size

• Compared 2s, 3s, 4s windows
• 2–3s were better for real-time; 4s slightly improved accuracy

Conclusion

ML can enhance earthquake early warning systems. Random Forest models with carefully crafted features from P-waves provide fast, accurate magnitude predictions.