Towards Trustworthy Geodetic Data: Integrating XGBoost and SHAP for Interpretable Seismic Observation Quality Control

Abstract The scientific reliability of geophysical products derived from continuous geodetic observation networks depends critically on the quality of the underlying waveform recordings, yet the rapidly growing volume of data generated by modern broadband seismographic stations has rendered manual quality inspection infeasible at network scale. Traditional quality control methods based on single-indicator threshold rules suffer from limited generalizability, high false alarm rates, and an inability to capture compound degradation patterns that manifest simultaneously across multiple signal domains. This study proposes an integrated framework that combines multi-domain feature engineering, eXtreme Gradient Boosting (XGBoost) classification, and SHapley Additive exPlanations (SHAP)-based model interpretability for trustworthy, automated geodetic observation quality control. A comprehensive feature engineering strategy extracts 30 descriptive attributes from each sliding window segment across six complementary signal categories—time-domain, frequency-domain, gradient, statistical, detection and autocorrelation, and energy—providing a holistic characterization of both stationary amplitude properties and transient anomalous behaviors. A multi-criteria pseudo-labeling strategy assigns quality annotations based on the co-occurrence of at least two independent diagnostic exceedances, combined with waveform-level anomaly injection in which synthetic spikes, gain steps, dropouts, and clipping episodes are superimposed on the raw trace so that the resulting feature perturbations propagate through the extraction pipeline in a physically consistent manner. The framework is validated on 24 hours of continuous broadband data from each of five Global Seismographic Network stations spanning contrasting noise regimes and sensor types (IU.ANMO, IU.COLA, II.BFO, IU.TUC, and IU.KIP), totalling 172,675 analysis windows. Three forms of validation are reported: (i) within-station evaluation, (ii) leave-one-station-out transfer, and (iii) an independent test against 1,200 expert-adjudicated windows and the operational IRIS MUSTANG quality metrics, which are generated by a pipeline wholly disjoint from the pseudo-labeling rule. Experimental results demonstrate that the XGBoost classifier attains a within-station test accuracy of 0.974 and an F1-score of 0.903, retains an F1-score of 0.812 under leave-one-station-out transfer, and achieves an F1-score of 0.774 against the independent expert labels, versus 0.511 for a tuned single-indicator STA/LTA threshold, 0.542 for an amplitude-clipping detector, 0.629 for an isolation forest, and 0.701 for a logistic-regression baseline restricted to the same five labeling features. A feature-ablation experiment in which all five label-defining features are withheld still yields an F1-score of 0.731 on the independent labels, demonstrating that the model exploits genuine multi-domain quality information rather than merely reproducing its own labeling rule. SHAP-based global interpretability analysis reveals that peak-to-peak amplitude and standard deviation are the dominant anomaly predictors (mean $$|\textrm{SHAP}|$$ | SHAP | of 1.897 and 1.896, respectively), followed by maximum absolute gradient, kurtosis, and STA/LTA ratio, all of which align with established physical signatures of seismic data quality degradation. Local waterfall decompositions and multi-sample heatmaps further confirm that the model recognizes multiple distinct degradation modes rather than relying on a single anomaly archetype. Inference requires 0.42 ms per window on a single CPU core, corresponding to a throughput of roughly 2,380 windows per second and confirming the feasibility of real-time deployment at network scale. The proposed framework bridges the gap between high-performance machine learning classification and the operational transparency demanded by mission-critical geodetic infrastructure, delivering accurate per-window quality flags accompanied by physically meaningful, sample-level explanations that enable domain experts to audit and verify every automated decision.

Authors

Institutions

Publication Details

Journal
Neural Processing Letters
Published
2026-11-14
DOI
https://doi.org/10.1007/s11063-026-11897-4
Primary Topic
Seismology and Earthquake Studies
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Towards Trustworthy Geodetic Data: Integrating XGBoost and SHAP for Interpretable Seismic Observation Quality Control

Chuannan Chen
Neural Processing Letters
Seismology and Earthquake Studies
article

Towards Trustworthy Geodetic Data: Integrating XGBoost and SHAP for Interpretable Seismic Observation Quality Control

Chuannan Chen
article en

Abstract

Abstract The scientific reliability of geophysical products derived from continuous geodetic observation networks depends critically on the quality of the underlying waveform recordings, yet the rapidly growing volume of data generated by modern broadband seismographic stations has rendered manual quality inspection infeasible at network scale. Traditional quality control methods based on single-indicator threshold rules suffer from limited generalizability, high false alarm rates, and an inability to capture compound degradation patterns that manifest simultaneously across multiple signal domains. This study proposes an integrated framework that combines multi-domain feature engineering, eXtreme Gradient Boosting (XGBoost) classification, and SHapley Additive exPlanations (SHAP)-based model interpretability for trustworthy, automated geodetic observation quality control. A comprehensive feature engineering strategy extracts 30 descriptive attributes from each sliding window segment across six complementary signal categories—time-domain, frequency-domain, gradient, statistical, detection and autocorrelation, and energy—providing a holistic characterization of both stationary amplitude properties and transient anomalous behaviors. A multi-criteria pseudo-labeling strategy assigns quality annotations based on the co-occurrence of at least two independent diagnostic exceedances, combined with waveform-level anomaly injection in which synthetic spikes, gain steps, dropouts, and clipping episodes are superimposed on the raw trace so that the resulting feature perturbations propagate through the extraction pipeline in a physically consistent manner. The framework is validated on 24 hours of continuous broadband data from each of five Global Seismographic Network stations spanning contrasting noise regimes and sensor types (IU.ANMO, IU.COLA, II.BFO, IU.TUC, and IU.KIP), totalling 172,675 analysis windows. Three forms of validation are reported: (i) within-station evaluation, (ii) leave-one-station-out transfer, and (iii) an independent test against 1,200 expert-adjudicated windows and the operational IRIS MUSTANG quality metrics, which are generated by a pipeline wholly disjoint from the pseudo-labeling rule. Experimental results demonstrate that the XGBoost classifier attains a within-station test accuracy of 0.974 and an F1-score of 0.903, retains an F1-score of 0.812 under leave-one-station-out transfer, and achieves an F1-score of 0.774 against the independent expert labels, versus 0.511 for a tuned single-indicator STA/LTA threshold, 0.542 for an amplitude-clipping detector, 0.629 for an isolation forest, and 0.701 for a logistic-regression baseline restricted to the same five labeling features. A feature-ablation experiment in which all five label-defining features are withheld still yields an F1-score of 0.731 on the independent labels, demonstrating that the model exploits genuine multi-domain quality information rather than merely reproducing its own labeling rule. SHAP-based global interpretability analysis reveals that peak-to-peak amplitude and standard deviation are the dominant anomaly predictors (mean $$|\textrm{SHAP}|$$ | SHAP | of 1.897 and 1.896, respectively), followed by maximum absolute gradient, kurtosis, and STA/LTA ratio, all of which align with established physical signatures of seismic data quality degradation. Local waterfall decompositions and multi-sample heatmaps further confirm that the model recognizes multiple distinct degradation modes rather than relying on a single anomaly archetype. Inference requires 0.42 ms per window on a single CPU core, corresponding to a throughput of roughly 2,380 windows per second and confirming the feasibility of real-time deployment at network scale. The proposed framework bridges the gap between high-performance machine learning classification and the operational transparency demanded by mission-critical geodetic infrastructure, delivering accurate per-window quality flags accompanied by physically meaningful, sample-level explanations that enable domain experts to audit and verify every automated decision.

Neural Processing Letters
China University of Geosciences (Beijing) (CN)
Seismology and Earthquake Studies
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon