Modern Architecture

DEEP LEARNING-BASED EARTHQUAKE EARLY WARNING SYSTEM WITH DOMAIN ADAPTATION

Master Research

Received : October 2026

Accepted : October 2026

Program: MSc Computer Science

Category: 7MG001 The Master Research Project

Authors

Zin Yin Minn

University of Wolverhampton

0 Views0 Downloads

Abstract

Earthquake Early Warning (EEW) systems play a critical role in reducing casualties and infra structure damage by providing advance warning before the arrival of strong ground shaking. Conventional EEW approaches rely on handcrafted signal processing, explicit seismic phase picking, and region-specific calibration. While effective in well-instrumented regions, such methods often lack adaptability when deployed in new tectonic environments with different noise characteristics, station geometries, and limited historical data. These limitations are par ticularly pronounced in developing regions with sparse seismic infrastructure.
This dissertation addresses these challenges by designing, implementing, and evaluating a deep learning-based EEW prototype that incorporates domain adaptation. The proposed system em ploys an end-to-end convolutional neural network (CNN) architecture capable of processing raw seismic waveforms in near real time. A multi-task learning framework is adopted to jointly perform earthquake detection and key parameter estimation, including magnitude, epicentral distance, and S-P time. Domain adaptation is systematically evaluated to assess cross-regional generalization performance.
The methodology follows a two-stage training strategy. Models are first pre-trained on the large-scale Stanford Earthquake Dataset (STEAD) to learn general seismic representations, and subsequently fine-tuned using region-specific datasets from the United States Geological Sur vey (USGS) and a regional Myanmar seismic network. Performance is evaluated using standard classification metrics and regression error measures.
The proposed system achieves near state-of-the-art performance on the STEAD benchmark, with strong generalization to USGS data and meaningful performance on the Myanmar dataset despite limited data availability. While distance estimation degrades under domain shift, mag nitude and S-P time estimates remain comparatively stable, supporting magnitude-driven alert logic. A real-time demonstration system confirms operational feasibility with low-latency in ference. Overall, the results demonstrate that CNN-based EEW systems with domain adaptation provide a practical and flexible solution for region-specific deployment, particularly in data scarce environments.

Keywords

7MG001 The Master Research Project

Introduction

Earthquake Early Warning (EEW) systems aim to provide rapid alerts immediately after the detection of an earthquake, offering critical seconds to tens of seconds for protective actions before strong ground shaking arrives. These systems have demonstrated societal benefits in reducing casualties, infrastructure damage, and secondary hazards when effectively deployed (Allen and Melgar, 2019; Böse et al., 2018).
Conventional EEW methodologies typically rely on handcrafted signal processing, explicit seismic phase picking, and region-specific empirical relationships derived from physical mod els of wave propagation. While these approaches have achieved operational success in several regions, they often require extensive manual calibration, dense station coverage, and detailed regional velocity models. As a result, their scalability and adaptability to new tectonic environ ments remain limited, particularly in data-scarce or resource-constrained regions.
Recent advances in deep learning have introduced an alternative paradigm for EEW, ena bling end-to-end learning directly from seismic waveforms. Convolutional neural networks (CNNs) and related architectures have shown strong performance in earthquake detection, phase picking, and source parameter estimation by learning discriminative representations from large volumes of waveform data (Mousavi et al., 2019; Zhu and Beroza, 2019). These data driven approaches reduce reliance on handcrafted features and offer the potential for more flex ible and extensible EEW systems.
However, despite promising results on benchmark datasets, several challenges remain unre solved. Many existing studies focus on isolated subtasks such as detection or magnitude estimation rather than evaluating integrated, system-level EEW pipelines. Furthermore, models trained on large curated datasets often exhibit degraded performance when applied to new re gions with different station geometries, noise characteristics, and instrumentation. Addressing this gap requires explicit consideration of domain adaptation, where models are adapted to new regions using limited local data rather than relying on unrealistic assumptions of universal gen eralization.

Related Papers