Miao YU
··· views
← Publications

Resolving Kinematic Characteristics of Anthropogenic Seismic Sources using Dense Nodal Arrays and Spatiotemporal Deep Learning

Yu, M., Chai, M., Meng, H., et al. — Journal of Applied Geophysics, 2026

Under Review

Array SeismologyBeamformingConvLSTMSource LocalizationMulti-Target TrackingDense Seismic Array

Abstract

Continuous monitoring of shallow anthropogenic seismic sources — pedestrians, vehicles, construction, and industrial machinery — is challenging in urban environments due to complex wavefield scattering and high ambient noise. We present a systematic framework for tracking single and multiple moving sources using small-aperture dense nodal arrays. For isolated sources, multi-array delay-and-sum beamforming combined with geometric cross-localization and an empirical, field-calibrated slowness correction achieves an 87.48% localization accuracy with a 4.03 m spatial uncertainty (RMSE) on field data from four 4x4 nodal arrays (3 m spacing) in Huizhou, China. For concurrent multiple sources, geometric cross-localization produces a highly aliased objective function with a combinatorial explosion of spurious intersections. We address this by reformulating multi-target tracking as an end-to-end spatiotemporal sequence-prediction problem, introducing a dual-input Convolutional LSTM (ConvLSTM) architecture that fuses dynamic multi-array beam-energy sequences with static array-position encoding. Optimized via Hungarian matching and a joint position-confidence loss, the network resolves severe multi-source ambiguities — including crossing and closely parallel trajectories — achieving an F1-score of 0.927 and a mean localization error of 2.76 m on independent field data.

Status: Submitted to Journal of Applied Geophysics in July 2026; currently under review.

Contribution: Co-first author. Developed the multi-array beamforming and cross-localization pipeline, derived the field-calibrated slowness correction map, and co-designed the dual-input ConvLSTM tracking architecture and its Hungarian-matching training objective.

Code & Data: GitHub repository · Zenodo archive: code · field data

Significance: This work bridges classical array seismology with geometry-aware deep learning for urban intelligence applications — resolving ambiguities that defeat purely geometric source tracking and directly demonstrating the translational impact of geophysical methods on city-scale monitoring solutions.

Figures

Study site and dense nodal seismic array layout at Bailu Lake Kite Square, Huizhou, China — four 4×4 arrays (3 m station spacing) used for beamforming-based source tracking.
Figure 1. Study site and dense nodal seismic array layout at Bailu Lake Kite Square, Huizhou, China — four 4×4 arrays (3 m station spacing) used for beamforming-based source tracking.
Dual-input ConvLSTM architecture: multi-array beam-energy sequences and static station-coordinate encodings are fused to predict source location and detection confidence.
Figure 2. Dual-input ConvLSTM architecture: multi-array beam-energy sequences and static station-coordinate encodings are fused to predict source location and detection confidence.
Real-time tracking of a walking pedestrian: beam-energy distributions from two arrays (a-c) rotate with the pedestrian's motion, the waveform (d) shows individual footstep arrivals, and (e) shows the reconstructed trajectory relative to the arrays.
Figure 3. Real-time tracking of a walking pedestrian: beam-energy distributions from two arrays (a-c) rotate with the pedestrian's motion, the waveform (d) shows individual footstep arrivals, and (e) shows the reconstructed trajectory relative to the arrays.
Precision, recall, and F1-score across position-loss weighting choices; the best configuration (weight = 4) achieves an F1-score of 0.927.
Figure 4. Precision, recall, and F1-score across position-loss weighting choices; the best configuration (weight = 4) achieves an F1-score of 0.927.
Export BibTeX ⇩