A Hybrid CNN-LSTM Architecture for Seismic Event Detection Using High-Rate GNSS Velocity Time Series

Highlights: This study evaluates a hybrid CNN–LSTM model for seismic event detection from 5 Hz GNSS velocity time series, trained with pseudo-synthetic data and tested on an inde-pendent real GNSS setting. What are the main findings? Hybrid CNN–LSTM model outputs were evaluated both at frame-level and event–station-level, and multiple multi-channel decision schemes (vote-based, any-channel, weighted fusion) were reported under both fixed and validation-calibrated operating points. On the independent real GNSS test setting, the horizontal components (H0/H1) show substantially stronger discrimination than the vertical (UP) component, with the largest degradation observed on UP under real ambient noise. What are the implications of the main findings? Event–station aggregation and fusion/operating-point choice govern missed-detection vs false-alarm trade-off on real data. The observed real-test performance drop is consistent with a measurable training–test distribution shift; MMD- and SHAP-based diagnostics help explain changing decision patterns. Global Navigation Satellite Systems (GNSS) have become essential tools in geomatics engineering for precise positioning, cadastral surveys, topographic mapping, and deformation monitoring. Recent advances integrate GNSS with emerging technologies such as artificial intelligence (AI), machine learning (ML), cloud computing, and unmanned aerial systems (UAS), which have greatly improved accuracy, efficiency, and analytical capabilities in managing geospatial big data. In this study, we propose a hybrid Convolutional Neural Network–Long Short Term Memory (CNN-LSTM) architecture for seismic detection using high-rate (5 Hz) GNSS velocity time series. The model is trained on a large synthetic dataset generated by and real high-rate GNSS non-event data. Model performance was evaluated using real event and non-event data through an event-based approach. The results demonstrate that a hybrid deep-learning architecture can provide a reliable framework for seismic detection with high-rate GNSS velocity time series.