A comprehensive machine learning pipeline for predicting earthquake magnitudes from early-stage 3-component seismic waveforms using 1D CNNs with data augmentation and weighted loss strategies.

🌟 Key Features

• End-to-end pipeline from raw seismic data to magnitude prediction
• Multi-component analysis using Z, N, E seismic components
• Physics-informed data augmentation for rare high-magnitude events
• Weighted loss function to handle severe class imbalance
• Comprehensive evaluation across different sampling rates and configurations

📊 Performance Highlights

Configuration	Sampling Rate	Augmentation	MAE	R²	Accuracy (±0.2 mag)
Best Model	100 Hz	✅ Yes	0.276	0.166	41.6%
Baseline	50 Hz	❌ No	0.313	-0.007	32.7%

🚀 Quick Start

Prerequisites

Core Dependencies

pip install torch==2.2.2 torchvision==0.17.2
pip install numpy==1.26.4 pandas==2.2.3 scikit-learn==1.7.1
pip install obspy==1.4.2 matplotlib==3.10.0
pip install scipy==1.16.0 tqdm==4.67.1

Full Environment

pip install -r requirements.txt

Usage

from seismic_magnitude_prediction import run_experiment

# Run best configuration
results = run_experiment(
    seconds=5,
    sample_rate=100,
    data_file="seismic_data_5_seconds_sampling_rate_100.pkl",
    target_length=500,
    batch_size=32,
    num_epochs=100,
    learning_rate=0.001,
    is_augmented=True
)

📁 Project Structure

├── seismic_magnitude_prediction.py  # Main ML pipeline
├── data_preprocessing.py            # Data extraction & validation
├── requirements.txt
├── README.md
└── data/
   ├── Turkey_data/
   │   ├── Catalog/                 # Text catalog files
   │   └── DATA/                    # MiniSEED waveform files
   └── processed/
       └── *.pkl                    # Processed datasets

🛠️ Methodology

Data Pipeline

Catalog Parsing: Extract event metadata from Turkish seismic catalogs Waveform Extraction: Extract 5-second snippets around P-wave arrivals Quality Control: 7-level validation system for data integrity Standardization: Resample to 100 Hz and normalize signals

Model Architecture

Input: 3-channel waveforms (Z, N, E components) × 500 time samples Architecture: 4-layer 1D CNN with batch normalization and global pooling Output: Single regression value (earthquake magnitude)

Class Imbalance Solutions

Data Augmentation: Physics-informed noise, time shift, amplitude scaling Weighted Loss: Higher penalties for rare high-magnitude events (10x weight for mag 6.5+)

📈 Dataset Information

Source: Turkish Seismic Network
Total Samples: 17,773 (50 Hz) / 12,671 (100 Hz)
Magnitude Range: 3.5 - 7.5
Components: 3-channel seismic recordings (Z/N/E)
Duration: 5 seconds post P-wave arrival

Class Distribution

Magnitude 3.5-4.5: ~90% of samples (dominant)
Magnitude 4.5-5.5: ~8% of samples  
Magnitude 5.5-6.5: ~1% of samples
Magnitude 6.5-7.5: <0.1% of samples (critical but rare)

🔬 Experimental Results

Key Findings

100 Hz + Augmentation significantly outperforms other configurations
Data augmentation is crucial for high-resolution data (prevents overfitting)
Class imbalance remains the primary challenge for high-magnitude prediction

Performance by Magnitude Range

Mag 3.5-4.5: MAE = 0.224 ✅ (Excellent)
Mag 4.5-5.5: MAE = 0.766 ⚠️ (Acceptable)
Mag 5.5+: MAE > 1.3 ❌ (Poor - insufficient data)

📋 Requirements System Requirements

Python 3.11+
CUDA 12.1+ (optional, for GPU acceleration)
8GB+ RAM (for large seismic datasets)

Tested Environment
This project was developed and tested with:

Python 3.11.13
PyTorch 2.2.2 with CUDA 12.1 support
ObsPy 1.4.2 for seismic data processing
NumPy 1.26.4 (ObsPy compatibility requirement)

Core Dependencies

# Core ML & Data Processing
torch==2.2.2
torchvision==0.17.2
numpy==1.26.4
pandas==2.2.3
scikit-learn==1.7.1

# Seismic Data Processing
obspy==1.4.2

# Visualization
matplotlib==3.10.0

# Data Utilities
scipy==1.16.0
tqdm==4.67.1

# Development Environment (optional)
jupyter==1.1.1
ipython==9.4.0

📄 License

MIT License - see LICENSE file for details.

📚 Citation

 @misc{seismic2025,
 title={Seismic Magnitude Prediction using Deep Learning},
 author={Sam84723},
 year={2025},
 url={https://github.com/sam84723/seismic_magnitude_predictor}
}

🙏 Acknowledgments

Dr. Itzhak Lior from HUJI university
Turkish Seismic Network for providing the dataset
ObsPy community for seismic data processing tools
PyTorch team for the deep learning framework

⚠️ Note: This model is for research purposes. For real earthquake early warning systems, additional validation and safety measures are required.