Hindcasting the Occurrence Time of Major Earthquakes using Machine Learning and Time Series Analysis
Author ORCID Identifier
https://orcid.org/0000-0001-9074-8984
Date of Award
17-8-2025
Document Type
Thesis
School
School of Computing
Programme
Ph.D.-Doctoral of Philosophy
First Advisor
Dr.N.Venkatanathan
Keywords
Earthquake Forecasting, Outgoing Longwave Radiation, Relative Humidity, Time Series Analysis, Machine Learning
Abstract
An earthquake is an intense shaking of the ground that typically occurs when tectonic plates move beneath the surface of the Earth. Scientists analyze historical seismic records and geophysical and atmospheric signs to build models estimating the probability of major earthquakes by detecting patterns and anomalies in data such as ground deformation and seismic waves. This study would help in predicting earthquakes to minimize risks of people and buildings.
The present study investigates the devastating earthquakes along the Chilean subduction zone in South America, linking non-seismic data to machine learning predictions of Outgoing Longwave Radiation (OLR) and Relative Humidity (RH) anomalies. Tectonic activity is highly variable in space, therefore the study region must be defined. The first stage which is getting the clusters done by the proposed method Local Maxima-based Spatio Cluster Analysis Network (LMSCAN). In this clustering method, the main quake are taken into consideration to classify the data into the seismology parameters of interest (magnitude and latitude) and the grouping of microshocks. To assess the efficacy of the clustering process, the effectiveness of their proposed technique will be measured against of conventional clustering algorithms such as K-means, Agglomerative Hierarchical Clustering, and Density-Based Spatial Clustering of Applications with Noise (DBSCAN).
By providing detection of Outgoing Longwave Radiation (OLR) as non-seismic precursor, the Singular Spectrum Analysis - Percentile-median Absolute Deviation Method (SSA-PADM) has been proposed, proving a new technique, for predicting the anomaly activity up to 6 months prior to the occurrence of major earthquakes.
The existing techniques which include Isolation Forest, 2 Sigma, Median Absolute Deviation and Percentile algorithms are compared for performance against this technique on detecting anomalies preceding a major earthquake. Moreover, the correlation of atmospheric parameters, OLR and RH are also implicated as predictors of the estimated seismic events through the Proposed Atmospheric and Radiative Anomaly Detection (ARAD) approach. The analysis explores the relationship between the drop of RH flux index value related to the raise of the flux index of OLR near the epicenter as a possible precursor of major earthquakes, with advance times from 3 to 40 days.
Accuracy improvements are found compared to wellknown methods like One-Class SVM (Support Vector Machine), Elliptic Envelope and Isolation Forest. With OLR and RH considered reliable predictors, the atmospheric variables are being forecasted with a hybrid machine learning model called Multi-Layer Perceptron with Expanded Window Cross-Validation (MLP-EWCV). The established methods, including Extreme Gradient Boosting (XGBoost), Random Forest Regressor, and Support Vector Regression (SVR), were used for comparison in order to evaluate the improvement in hindcasting accuracy and efficiency offered by the MLP-EWCV model.
The study of anomalous behaviour of OLR and RH may help to detect anomalies before earthquakes, thereby possibly functioning as an early warning for disaster management systems. The current research study attempts to understand Chile (South America) possible micro shocks with respect to tectonic model as a whole because it falls in the inter plate region where usually micro shocks are observed just before the major earthquakes.
Recommended Citation
D, Rubidha Devi Ms, "Hindcasting the Occurrence Time of Major Earthquakes using Machine Learning and Time Series Analysis" (2025). Theses and Dissertations. 11.
https://knowledgeconnect.sastra.edu/theses/11