Exploration of Machine Learning Methods to Seismic Event Discrimination in the Pacific Northwest

Akash Kharita \orcid 0000-0003-0612-7734, Marine Denolle \orcid 0000-0002-1610-2250, Alexander R. Hutko \orcid 0000-0003-3561-2747, J. Renate Hartog \orcid 0000-0002-4116-7806, Stephen D. Malone \orcid 0000-0002-5143-1850 · 2025

Read Exploration of Machine Learning Methods to Seismic Event Discrimination in the Pacific Northwest by Akash Kharita \orcid 0000-0003-0612-7734, Marine Denolle \orcid 0000-0002-1610-2250, Alexander R. Hutko \orcid 0000-0003-3561-2747, J. Renate Hartog \orcid 0000-0002-4116-7806, Stephen D. Malone \orcid 0000-0002-5143-1850 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-23795.txt.

Introduction

Akash Kharita and four co-authors compare machine learning methods for telling seismic sources apart in the Pacific Northwest: earthquakes, explosions, surface events, and noise. Networks keep densifying and signals overlap, so four-way sorting matters for real monitoring. Earlier studies mostly tried binary tasks and rarely pitted classic feature-based learning against deep learning fairly.

The team used about two hundred thousand three-component waveforms from over seventy thousand events. Random forests ran on three feature sets, while convolutional networks took raw time series or spectrograms. Spectrogram networks won, topping ninety-two percent accuracy both inside the region and on outside data, against eighty-nine percent for the best forest. One winning network is tiny — about seventy thousand parameters — and scans a full day of data in roughly nine seconds on ordinary hardware. It ships inside the SeisBench toolkit with released checkpoints, ready for real-time operations.

Catalog record

CREATORAkash Kharita \orcid 0000-0003-0612-7734, Marine Denolle \orcid 0000-0002-1610-2250, Alexander R. Hutko \orcid 0000-0003-3561-2747, J. Renate Hartog \orcid 0000-0002-4116-7806, Stephen D. Malone \orcid 0000-0002-5143-1850
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE88 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-23795

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Exploration of Machine Learning Methods to Seismic Event Discrimination in the Pacific Northwest Akash Kharita \orcid 0000-0003-0612-7734, Marine Denolle \orcid 0000-0002-1610-2250, Alexander R. Hutko \orcid 0000-0003-3561-2747, J. Renate Hartog \orcid 0000-0002-4116-7806, Stephen D. Malone \orcid 0000-0002-5143-1850 Abstract Seismic event discrimination in the PNW \credit ConceptualizationAK,MD \credit MethodologyAK, MD \credit SoftwareAK \credit ValidationAK,AH,SM \credit Formal AnalysisAK,MD \credit InvestigationAK,MD,AH \credit ResourcesRH \credit Writing - Original draftAK \credit Writing - Review & EditingAK,MD,AH,SM \credit VisualizationAK \credit SupervisionAK,MD,RH \credit Project administrationMD,RH \credit Funding acquisitionMD,RH {summary} Abstract Accurately separating tectonic, anthropogenic, and geomorphologic seismic sources is essential for Pacific Northwest (PNW) monitoring but remains difficult as networks densify and signals overlap. Prior work largely treats binary discrimination and seldom compares classic ML (feature-engineered) and deep learning (end-to-end) approaches under a common, multi-class setting with operational constraints. We evaluate methods and features for four-way source discrimination—earthquakes, explosions, surface events, and noise—and identify models that are both accurate and deployable. Using \sim 200k three-component waveforms from > 70k events in an AI-curated PNW dataset, we test random-forest classifiers on TSFEL, physics-informed, and scattering features, and CNNs that ingest time series (1D) or spectrograms (2D); we benchmark on a balanced common test set, a 10k-event network dataset, and out-of-domain data (global surface events; near-field blasts). CNNs taking spectrograms lead with accuracy performance over 92% for within-domain (as a short-and-fat CNN SeismicCNN 2D) and out-of-domain (as a long and skinny CNN QuakeXNet 2D), versus 89% for the best random forest; performance remains strong at low SNR and longer distances, and generalizes to independent network and global datasets. QuakeXNet-2D is lightweight ( \approx 70k parameters; \sim 1.2 MB), implemented into seisbench , scans a full day of 100 Hz, three-component data in 9̃ s on commodity hardware, with released checkpoints. These results show spectrogram-based CNNs provide state-of-the-art accuracy, efficiency, and robustness for real-time PNW operations and transferable surface-event monitoring. 1 Introduction The Pacific Northwest (PNW) region of the United States, situated at the dynamic boundary between the North American continental plate and the Juan de Fuca oceanic plate, presents unique challenges and opportunities in seismic monitoring in a multi-geohazard-prone landscape with a subduction-zone plate boundary. The PNW experiences diverse seismic sources (Fig. 1 ), including large megathrust earthquakes ( Witter et al. 2003 , e.g.,) , intraslab ( Ichinose et al. 2004 , e.g.,) and crustal earthquakes ( Gomberg and Bodin 2021 , e.g.,) , slow repeating earthquakes ( Bartlow 2020 ; Rogers and Dragert 2003 ; Wech and Bartlow 2014 , e.g.,) , tectonic tremors ( Wech et al. 2010 , e.g.,) , and low-frequency earthquakes ( Royer and Bostock 2014 , e.g.,) . Beyond earthquakes, over twenty active and glaciated volcanoes, as well as extensive mountain ranges, experience frequent landslides and debris flows ( Luna and Korup 2022 , e.g.,) . Anthropogenic activities, such as quarry blasts (REF), generate ground motion intensities comparable to those of small-magnitude earthquakes, further complicating the source of this seismicity (REF) ( Kramer et al. 2024 ) . Such a variety of seismic sources necessitates robust classification methods to accurately label and catalog these events . The PNW Seismic Network (PNSN) ( Hellweg et al. 2020 ) , a key component of the Advanced National Seismic System (ANSS), has been operating since 1969 and currently manages over 600 seismic stations in the states of Washington and Oregon, providing essential data for seismic event analysis. Current event detection relies on traditional techniques such as the Short-Time Average to Long-Time Average (STA/LTA) ratio algorithm ( Allen 1982 , e.g.,) . While effective for basic event detection, this approach has limited accuracy when discriminating between visually similar waveforms from different event types, such as earthquakes, controlled explosions, and glacier-related surface events. These limitations have become more pressing with the expansion of seismic networks and the increasing volume of data ( Carniel et al. 2021 ; Kong et al. 2019 , e.g.,) , particularly in regions like the PNW, where the simultaneous occurrence of multiple seismic sources adds to the complexity of waveform interpretation. Traditional discrimination techniques, such as analyzing the spectral ratios of P and S waves or differences in local and coda magnitudes, which were developed for binary classification between earthquakes and explosions ( Koper et al. 2016 ; Koper et al. 2024 , e.g.,) , also face limitations. For example, these parameters are often unavailable during preliminary analyses, which prevents accurate classification in near real-time. While calculating duration (coda) magnitude and local magnitude alongside their spectral ratios could provide valuable insights into potential misclassifications ( Koper et al. 2016 ; Koper et al. 2024 , e.g.,) , such comprehensive analyses are rarely performed in real-time workflows. Small magnitude earthquakes (M L < 3) have been difficult to distinguish from mining and single-shot explosions recorded at local distances (< 50–150 km) due to similarities in their seismic waveforms and spectral characteristics ( Fig. 2 ) and (Fig. 3 ) particularly at frequencies below 10 Hz ( Wang et al. 2020 ; Koper et al. 2021 , e.g.,) . Similarly, seismic signals generated by mass movements on volcanoes, such as landslides, debris flows, and lahars, closely resemble those associated with low-frequency volcanic seismicity and shallow volcano-tectonic earthquakes. These events are characterized by emergent waveforms, indistinguishable P and S phases, and dominant frequencies below 5 Hz ( Wassermann 2012 ; Allstadt et al. 2014 ; Allstadt et al. 2018 , e.g.,) . Furthermore, seismic signals from rockfalls with a significant free-fall component can exhibit similarities to small earthquakes when recorded locally ( Hibert et al. 2011 ; Hibert et al. 2014 ) . In volcanic regions near populated areas, volcano-seismic signals, including those related to low-frequency volcanic events or shallow volcano-tectonic earthquakes, can also be misinterpreted as anthropogenic noise ( Wassermann 2012 ) . Many studies have focused on the binary classification of seismic events, particularly distinguishing between earthquakes and explosions ( Koper et al. 2021 ; Kong et al. 2022 , e.g.,) , or other sources such as slope failures ( Wenner et al. 2020 ; Chmiel et al. 2021 ; Hibert et al. 2017 , e.g.,) or icequakes ( Pirot et al. 2023 ; Kharita et al. 2024 , e.g.,) . Binary classification is generally easier because it involves separating only two classes with more distinct waveform features and requires a solution for a single decision boundary. As a result, performance in binary problems is typically very high, with many studies reporting classification accuracies exceeding 95% ( Wang et al. 2020 ; Koper et al. 2021 , e.g.,) . By contrast, in complex environments such as volcanic regions, where seismic signals from landslides, pyroclastic flows, and low-frequency volcanic events share similar waveforms, multi-class classification is necessary and considerably more difficult ( Wassermann 2012 ; Allstadt et al. 2014 ; Allstadt et al. 2018 , e.g.,) . Discriminating among multiple event types is more challenging because the boundaries between classes are less distinct, and performance is typically lower, often in the range of 75–90% ( Hibert et al. 2014 ; Hibert et al. 2017 ; Hibert et al. 2019 , e.g.,) . Classic Machine Learning (CML) and Deep Learning (DL) have shown promise in seismic event classification by enabling automated feature extraction and effective discrimination between

Read the complete text