<p>Earthquake early warning (EEW) systems issue alerts in the brief interval between the first P-waves and the arrival of strong S-waves or surface waves. Traditional algorithms degrade under heterogeneous, noisy conditions, motivating complementary artificial intelligence (AI) approaches. However, most previous studies emphasized model accuracy alone; we provide a benchmark of modern deep-vision backbones within a standardized time–frequency pipeline to characterize the accuracy–latency trade-off and assess real-time feasibility for EEW event/noise discrimination. We fine-tune and evaluate six pretrained vision models for event/noise discrimination on 10 s vertical-component windows aligned so that the P onset occurs <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\sim\)</EquationSource> </InlineEquation>7–8 s into the window, i.e., only the first 2–3 s of post-onset P-wave signal are included. Each window is transformed to a continuous wavelet transform power scalogram and fed to pre-trained convolutional neural networks (CNNs; EfficientNet-B0, ResNet-18/50, ShuffleNetV2) and vision transformers (ViT-B/16, Swin-Tiny), fine-tuned for binary classification. Models are trained on the STanford EArthquake Dataset (STEAD, global) and externally evaluated on the Italian Seismic Dataset for Machine Learning (INSTANCE, Italy) and the Red Sísmica del Noroeste de México (RESNOM; the CICESE broadband network in northwestern Mexico), with decisions derived from either a fixed threshold or dataset-specific calibration. The models achieve near-ceiling discrimination on STEAD and strong but dataset-dependent transfer to independent networks, with external performance in the range MCC <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\approx\)</EquationSource> </InlineEquation> 0.87–0.97 and F1 <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\approx\)</EquationSource> </InlineEquation> 0.93–0.98, depending on model, training regime, and thresholding strategy. The lightweight CNNs evaluate a single trace in only 0.6 ms, a latency compatible with real-time EEW requirements. This study introduces a unified framework for benchmarking early event/noise discrimination based on short P-onset windows, explicitly incorporating end-to-end computational latency. By demonstrating cross-network generalization, millisecond-scale inference, and physically interpretable focus on the P onset, the results provide a realistic foundation for next-generation AI modules within operational early warning systems.</p>

Real-time seismic signal classification using deep vision architectures for earthquake early warning systems

期刊论文
出版模式：

开放获取
发表日期：

2026年3月18日

查看全文

Andrés Navarro-Rodríguez,
Oscar Alberto Castro-Artola,
Boris Rösler,
Enrique Efrén García-Guerrero,
Oscar Adrian Aguirre-Castro,
Jorge Galarza-Falfan,
Everardo Inzunza-Gonzalez

摘要

Earthquake early warning (EEW) systems issue alerts in the brief interval between the first P-waves and the arrival of strong S-waves or surface waves. Traditional algorithms degrade under heterogeneous, noisy conditions, motivating complementary artificial intelligence (AI) approaches. However, most previous studies emphasized model accuracy alone; we provide a benchmark of modern deep-vision backbones within a standardized time–frequency pipeline to characterize the accuracy–latency trade-off and assess real-time feasibility for EEW event/noise discrimination. We fine-tune and evaluate six pretrained vision models for event/noise discrimination on 10 s vertical-component windows aligned so that the P onset occurs \(\sim\) 7–8 s into the window, i.e., only the first 2–3 s of post-onset P-wave signal are included. Each window is transformed to a continuous wavelet transform power scalogram and fed to pre-trained convolutional neural networks (CNNs; EfficientNet-B0, ResNet-18/50, ShuffleNetV2) and vision transformers (ViT-B/16, Swin-Tiny), fine-tuned for binary classification. Models are trained on the STanford EArthquake Dataset (STEAD, global) and externally evaluated on the Italian Seismic Dataset for Machine Learning (INSTANCE, Italy) and the Red Sísmica del Noroeste de México (RESNOM; the CICESE broadband network in northwestern Mexico), with decisions derived from either a fixed threshold or dataset-specific calibration. The models achieve near-ceiling discrimination on STEAD and strong but dataset-dependent transfer to independent networks, with external performance in the range MCC \(\approx\) 0.87–0.97 and F1 \(\approx\) 0.93–0.98, depending on model, training regime, and thresholding strategy. The lightweight CNNs evaluate a single trace in only 0.6 ms, a latency compatible with real-time EEW requirements. This study introduces a unified framework for benchmarking early event/noise discrimination based on short P-onset windows, explicitly incorporating end-to-end computational latency. By demonstrating cross-network generalization, millisecond-scale inference, and physically interpretable focus on the P onset, the results provide a realistic foundation for next-generation AI modules within operational early warning systems.