Interpretable Machine Learning Automates Quality Control for Broadband Seismic Observation Networks

An integrated machine learning framework combining multi-domain feature engineering, gradient boosting classification, and model interpretability techniques has demonstrated the ability to automate quality control for continuous geodetic observation networks. Developed by researcher Chuannan Chen and reported in Neural Processing Letters, the automated framework pairs eXtreme Gradient Boosting, known as XGBoost, with SHapley Additive exPlanations, or SHAP, to inspect continuous broadband seismic waveforms. The study directly targets a primary operational challenge in modern observational geophysics: broadband seismographic stations generate vast volumes of continuous waveform recordings, creating an incoming data stream whose sheer scale makes comprehensive manual quality inspection completely infeasible across distributed observation networks.

High-quality waveform recordings are fundamental to the scientific reliability of geophysical products derived from geodetic observation infrastructure. However, conventional automated quality control approaches rely largely on single-indicator threshold rules, which frequently fall short in operational settings. According to the study, these traditional rule-based mechanisms suffer from limited generalizability across diverse recording environments, produce high false alarm rates, and remain incapable of capturing compound degradation patterns that manifest across multiple signal domains at the same time. These shortcomings underscore the necessity of an analytical framework capable of evaluating multidimensional features simultaneously while maintaining transparent decision-making criteria.

To overcome the limitations of isolated indicator thresholds, the newly proposed architecture implements an extensive feature engineering strategy that computes 30 descriptive attributes from each sliding window segment of waveform data. These attributes are gathered across six complementary signal categories: time-domain, frequency-domain, gradient, statistical, detection and autocorrelation, and energy metrics. By computing metrics across this broad spectrum of domains, the extraction pipeline provides a holistic characterization of the recordings. This design allows the framework to evaluate both stationary amplitude properties and transient anomalous behaviors within each analysis window, capturing subtle distortions that single-indicator checks regularly overlook.

Training an effective supervised model required addressing the absence of comprehensive manual quality labels across vast geodetic datasets. To resolve this, the methodology incorporates a multi-criteria pseudo-labeling strategy that assigns quality annotations based on the co-occurrence of at least two independent diagnostic exceedances. This labeling routine is combined with waveform-level anomaly injection, wherein synthetic anomalies—specifically spikes, gain steps, dropouts, and clipping episodes—are directly superimposed onto raw trace recordings. Injecting these synthetic disruptions onto the raw waveform ensures that any resulting feature perturbations propagate through the entire 30-feature extraction pipeline in a physically consistent manner prior to classification.

The research validated the performance of the XGBoost framework using 24 hours of continuous broadband seismic data gathered from five separate Global Seismographic Network stations: IU.ANMO, IU.COLA, II.BFO, IU.TUC, and IU.KIP. These five installations were selected to represent contrasting noise regimes and distinct sensor types, providing a diverse operational testbed totaling 172,675 analysis windows. The author reported results across three distinct validation structures: within-station evaluation, leave-one-station-out transfer testing across stations, and an independent benchmark evaluated against 1,200 expert-adjudicated windows alongside operational IRIS MUSTANG quality metrics generated by an independent pipeline wholly disjoint from the pseudo-labeling rule.

Experimental evaluations revealed that the XGBoost classifier achieved strong detection performance across all test configurations. In within-station testing, the model attained an overall test accuracy of 0.974 and an F1-score of 0.903. Under the leave-one-station-out transfer protocol, which evaluates how well the algorithm generalizes to an unseen station with a distinct noise environment, the classifier retained an F1-score of 0.812. When measured against the 1,200 independent expert-adjudicated labels and operational IRIS MUSTANG metrics, the framework achieved an F1-score of 0.774. This independent performance notably surpassed several conventional and baseline benchmarks, including a tuned single-indicator STA/LTA threshold with an F1-score of 0.511, an amplitude-clipping detector at 0.542, an unsupervised isolation forest at 0.629, and a logistic-regression baseline restricted to the five labeling features at 0.701.

To confirm that the machine learning model was identifying genuine multi-domain degradation rather than simply memorizing the criteria used during pseudo-labeling, the author conducted a feature-ablation experiment. When all five features defining the labeling rule were withheld from the model, the classifier still achieved an F1-score of 0.731 on the independent expert-labeled dataset. Furthermore, global interpretability analysis using SHAP values uncovered the underlying physical indicators driving model predictions. Peak-to-peak amplitude and standard deviation emerged as the dominant anomaly predictors, recording mean absolute SHAP values of 1.897 and 1.896, respectively. These were followed by maximum absolute gradient, kurtosis, and the STA/LTA ratio, aligning with established physical signatures of seismic signal degradation. Local waterfall decompositions and multi-sample heatmaps confirmed that the classifier identifies multiple distinct degradation modes rather than relying on a single anomaly archetype.

In addition to classification accuracy and physical interpretability, the study evaluated the computational feasibility of deploying the system in operational environments. The inference process required just 0.42 milliseconds per sliding window on a single central processing unit core, equating to a processing throughput of approximately 2,380 windows per second. This high throughput confirms that the automated system can operate in real time at network scale across mission-critical geodetic infrastructure. By providing accurate per-window quality flags alongside sample-level explanations, the framework bridges high-performance machine learning classification with operational transparency, offering domain experts the necessary verifiable evidence to audit automated quality decisions.

Subject of Research: Interpretable machine learning for seismic observation quality control

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

Article References: Chen, C. (2026). Towards Trustworthy Geodetic Data: Integrating XGBoost and SHAP for Interpretable Seismic Observation Quality Control. Neural Processing Letters. https://doi.org/10.1007/s11063-026-11897-4

Image Credits: AI Generated

DOI: 10.1007/s11063-026-11897-4

Keywords: geodetic data, seismic observation, broadband seismograph, quality control, XGBoost, SHAP, machine learning interpretability, feature engineering, Global Seismographic Network, waveform degradation, anomaly injection, pseudo-labeling, IRIS MUSTANG, STA/LTA ratio, peak-to-peak amplitude, real-time seismic monitoring, feature ablation, geophysical networks