Seismic Event Classification of Low-Magnitude Events (M1–4)

Overview

This project benchmarks classical machine learning algorithms for the automatic classification of low-magnitude seismic events (M1–4) using metadata from the USGS Earthquake Catalogue. Low-magnitude events exhibit overlapping characteristics that make manual catalogue maintenance time-consuming and prone to bias.

The goal is to identify robust, interpretable, and operationally reliable machine learning models capable of handling severe class imbalance in real seismic catalogues.

---

Motivation

Accurate seismic catalogues are essential for:

• seismic hazard assessment
• early-warning systems
• large-scale geophysical research

As monitoring systems improve, the volume of detected low-magnitude events increases, making manual classification unsustainable. Automated, explainable ML solutions are therefore critical for maintaining catalogue consistency.

---

Dataset (High-Level)

• Source: USGS Earthquake Catalogue
• Time window: 1 Aug 2025 – 20 Nov 2025
• Magnitude range: 1.0 – 4.0
• Samples: 19,958 events
• Classes: Earthquake, Quarry Blast, Explosion, Mining Explosion, Icequake, Other

⚠️ Raw data is not included in this repository. See data/README.md for the full schema, preprocessing decisions, and the USGS query to reproduce the dataset.

---

Methodology

• Feature engineering from spatial, temporal, and uncertainty attributes
• One-hot encoding for categorical variables
• Stratified 60 / 20 / 20 train-validation-test split
• Cost-sensitive learning to address class imbalance
• Model selection via GridSearchCV using weighted F1-score

Models Evaluated

• K-Nearest Neighbours (KNN)
• Radius Nearest Neighbours (RNN)
• Logistic Regression
• Linear Discriminant Analysis (LDA)
• Support Vector Machine (SVM, RBF kernel)
• Random Forest

---

Evaluation Strategy

Due to extreme class imbalance, accuracy alone is misleading. Models are compared using:

• Weighted F1-score
• Macro F1-score
• Confusion matrices

This ensures fair evaluation of minority seismic event types.

---

Key Results

• Random Forest achieved the strongest overall performance, with the highest Macro-F1 and Weighted-F1 scores.
• SVM showed competitive performance, particularly in minority-class detection.
• Linear and instance-based models performed well on the majority of classes but struggled with rare events.

A full comparison table is available in results/results_metrics_summary.md.

---

Interpretability

Random Forest feature importance analysis highlights:

• depth
• latitude
• longitude
• magnitude
• uncertainty measures (depthError, dmin)

These findings align with known physical distinctions between natural and anthropogenic seismic sources.

---

Repository Structure

Repository Structure
├── data/ # Dataset documentation
├── notebooks/ # End-to-end ML pipeline
├── results/ # Tables, figures, and evaluation summaries
├── requirements.txt # Project requirements
└── README.md # Project overview

---

Reproducibility

All preprocessing, modelling, and evaluation steps are fully documented in: notebooks/seismic_ml_pipeline_git_clean.ipynb

---

👤 Author

Eslam Abuelella
MSc Data Science – Coventry University

• GitHub: https://github.com/eslamhussienabuelella
• Portfolio: https://eslamhussienabuelella.github.io

---

📄 License

This project is shared for academic and portfolio purposes.