Motivation
The Mt. Meager Distributed Acoustic Sensing (DAS) experiment generated a massive dataset capturing microseismic activity in a geologically complex and hazardous region with high geothermal potential. While seismic events had been manually located using traditional epicenter-based methods, the volume and complexity of the DAS data demanded scalable and automated solutions. Events were suspected to cluster geographically, especially around glacial and slope areas—making spatially-aware event classification essential. The project was motivated by the need to design a labeling and clustering strategy that could enable machine learning to assist or even automate the seismic event location process in future studies.
Our Solution
The project developed a machine learning-based pipeline to classify and cluster seismic events based on features extracted from the DAS data. Using k-means clustering on variables like spatial coordinates (x, y) and event power, multiple cluster configurations (2–5 clusters) were explored and visualized in 2D and 3D plots. The Elbow Method and silhouette scores were used to identify optimal clustering schemes. These clusters were then compared to known geophysical zones (e.g., Glacier North West, Slope South East) to validate their geographical relevance. The final labeled datasets and clustering models were exported, setting the groundwork for training supervised learning models to predict event locations based on raw event features.
My Contributions
I independently implemented a full data science workflow in Python to preprocess, cluster, and visualize microseismic events from the Mt. Meager DAS dataset. I handled data cleaning (including imputation of missing values), standardized key numerical features, and executed k-means clustering with different configurations. I developed visualizations to explore relationships between spatial features and clustering results and calculated silhouette scores to evaluate clustering quality. I also contributed to the development of a labeling methodology by mapping clusters to known geographic features and exported labeled datasets for future use.
Project Outcomes
Presentation: MoMacMo Presentation
Looking to discuss further? Contact me at research@mkmaharana.com
