Seismic Network Detection & Analysis

• Seismic network-based approaches are methods integrating sensor arrays, coherence analysis, and machine learning to enhance earthquake detection and catalog accuracy.
• They employ techniques such as coherence imaging, statistical network modeling, and deep learning, reducing location errors and filtering out noise effectively.
• Applications span real-time early warning, microseismic monitoring, and geophysical interpretation, offering actionable improvements in event characterization.

A seismic network based approach leverages distributed sensor arrays and/or inter-event networks—often in combination with advanced computational and machine learning techniques—to enhance earthquake detection, location, catalog cleaning, phase picking, pattern analysis, and geophysical interpretation. This paradigm exploits spatiotemporal redundancy, inter-station correlation, and population-level data aggregation to outperform single-station or locally windowed algorithms. Both traditional and modern (deep learning, computer vision) frameworks are used across real-time EEW, catalog curation, microseismic monitoring, and fundamental seismotectonic research.

1. Coherence-Based Network Detection and Catalog Cleaning

Modern workflows utilize large-N arrays and dense seismic networks to systematically exploit inter-station coherence for robust event discrimination and high-resolution earthquake location. The dominant scheme involves constructing a 4D coherence volume $C(x,y,z,t)$ via waveform stacking over a hypothesized grid of spacetime origins:

$$C(x, y, z, t) = \sum_{i=1}^N \sum_{j=1}^N w_{ij}\, s_i\bigl(t + \tau_i(x, y, z)\bigr) \cdot s_j\bigl(t + \tau_j(x, y, z)\bigr)$$

Here, $s_i(t)$ are high-pass filtered traces, $\tau_i$ are theoretical P/S travel times for candidate hypocenter $(x, y, z)$, and $w_{ij}$ are optional SNR-dependent or uniform weights. Projecting $C(x,y,z,t)$ by maximization over $(z,t)$ yields the 2D spatial “coherence image” $C(x,y)$.

A lightweight ResNet-style CNN, trained entirely on synthetic data generated by stochastic Green’s functions and realistic noise realizations, is then applied to each $C(x,y)$. The resulting output is a binary event/noise classification score. Real events (single, focused maxima in coherence image) are retained and re-located by waveform stacking, yielding an automated, high-resolution seismic catalog. Application to the Hengill COSEISMIQ dataset (12,374 events) reduced location error to $\sim$200 m (from $\sim$500 m using pick-based methods), eliminated 8% spurious events, and improved catalog completeness by $\sim$0.2 magnitude units (Solda et al., 31 Aug 2025).

Metric/Operation	Value/performance (LQ–COSEISMIQ)	Comments
Total events	12,374	$\text{M}_w$ range: –0.9…4.5
Retained as “event”	11,372	s ≥ 0.1 CNN threshold
Flagged as “noise”	1,002	s < 0.1
Precision/Recall	$96\%$/ $92\%$	Against HQDD reference
Med. mislocation (h/v)	150 m / 220 m

This “coherence-matrix + deep learning” pipeline obviates hand-labeled training data, enables rapid deployment to new arrays, and is extremely robust to false triggers and overlapping events (Solda et al., 31 Aug 2025).

2. Statistical and Topological Network Models of Seismicity

Another class of seismic-network methodologies casts earthquake catalogs into discrete networks—either inter-epicenter or spatial-cell based—with links defined via temporal proximity, recurrence, or signal similarity. These models expose scale-free, small-world, and long-range correlated properties of seismicity not visible to traditional approaches.

Epicentral network construction partitions the region into grid cells (20 km side for global studies, depth $<$70 km, $M \ge 4.5$). Directed temporal windows (T=3,800 s) are used to link successive or co-occurring events, revealing scale-free degree distributions $P(k) \sim k^{-\gamma}$ (with $\gamma = 1.85 \pm 0.01$), q-exponential cumulative degree laws ($q \approx 1.95$), and small-worldness (clustering coefficient $C \approx 0.52$, mean shortest-path $L \approx 5.2$) (Ferreira et al., 2014). Such networks demonstrate spatial and temporal earthquake “hubs,” non-local memory, and clustering on scales inconsistent with a Poisson process.

Metric	Value (global catalog)	Randomized catalog
Degree exponent $\gamma$	$1.85 \pm 0.01$	—
$q$-exponential parameter	$q=1.95\pm0.01$, $\beta=5.6$	—
Clustering coefficient $C$	0.518	$3.3 \times 10^{-4}$
Average path $L$	5.24	$\sim$10

Record-breaking network methods (Japan, California) further capture rupture-length scaling and temporal self-similarity by encoding events as nodes and constructing directed edges to those breaking the previous “closest-so-far” distance record. Distance and waiting-time PDFs exhibit universal scaling, enabling physical estimations of rupture length and revealing regional heterogeneity in short-time exponents (e.g., magnitude dependence in subduction regimes) (Revathi et al., 2010).

Correlation-based cell networks build spatial lattices (e.g., 100 km grid) and link nodes with high Pearson correlation of energy-release histories. These networks are highly assortative, display strong temporal persistence, and contain a significant fraction of connections at $>$1,000 km separation—directly indicating long-range dynamic coupling and memory in crustal dynamics (Tenenbaum et al., 2011).

3. Time-Evolving Station Networks: Phase/Frequency Measures and Event Monitoring

Beyond catalog-based approaches, seismic networks can be dynamically constructed in time, with nodes as sensors and edges reflecting pairwise coherence, cross-correlation, synchronization index, or mutual information in short (e.g., 30 s) analysis windows.

Let $x_i(t)$ denote the filtered velocity at station $i$. The weighted adjacency at time $t_k$ is

• $W_{ij}^{\rm CC}(t_k):$ Max lagged cross-correlation (1s lag window)
• $W_{ij}^{\rm SI}(t_k):$ Hilbert phase synchronization
• $W_{ij}^{\rm MI}(t_k):$ Normalized mutual information of amplitude PDFs
• $W_{ij}^{\rm coh}(t_k):$ Frequency-limited coherence, max over 0.5–20 Hz

Global indicators include the leading eigenvalues $\lambda_1(t_k)$ of $W(t_k)$, the number of edges exceeding threshold $\theta$, and k-means clustering to separate “event” links. For California mainshocks, this approach attained 100% TPR for $M_w > 2.5$ (FPR $\sim$0.3 h⁻¹), with coherence + k-means consistently outperforming amplitude- or time-domain measures in signal specificity and SNR (Ashkenazy et al., 2023).

This approach provides a real-time event-detection capability, scalable to $N \sim 100$ stations, and is extensible to EEW system integration. Its computational cost—$O(N^2)$ per window—is manageable for regional arrays (Ashkenazy et al., 2023).

4. Deep-Learning Architectures for Network-Scale Detection, Phase Picking, and Location

Deep learning models have enabled true end-to-end seismic network processing, including phase picking, joint detection, and hypocenter location.

PhaseNet and related 1D U-Net models process 3-component data to yield per-sample P/S/noise probabilities with precision/recall/F1 values of 0.96/0.94/0.95 (P) and 0.93/0.90/0.91 (S), feeding results directly into network association/location modules (Zhu et al., 2018).

EQNet employs a backbone ResNet (1D over stations and time), shift-and-stack back-projection module (kinematic alignment given station geometry and $v_P,v_S$), and an event-detection CNN. The joint loss over picks and detection is optimized end-to-end. On Ridgecrest, it matched or exceeded catalog F1 scores ($> 0.75$ vs SCSN and Liu 2020), achieved 90% of event times within $\pm$1 s, and 90% of epicenters within 6 km (Zhu et al., 2021).

Fully convolutional locators (e.g., FCN) map raw $[T \times S \times C]$ data to a $[L \times W \times H]$ grid of location probabilities, obviating velocity/inversion models, and localize events with mean test errors $4.9\,\text{km}$ (epicenter)/$1.0\,\text{km}$ (depth) on Oklahoma-induced seismicity (Zhang et al., 2018).

Network-level detection enhancement (e.g., SAIPy) aggregates single-station outputs using clustering and pattern-matching, delivering a 200–700% increase in detected volcano-tectonic events over catalogs, while preserving magnitude and polarity performance (Quinteros-Cartaya et al., 1 Feb 2026).

5. Applied Case Studies: Seismic Network Data in Operations

Microseismic and induced seismicity monitoring in dense arrays (Groningen: 832 stations, SCA: $N\sim100$) utilizes simple neural networks over station-wise attributes (e.g., multi-window STA/LTA, frequency densities) to enhance sensitivity by +65% over classic detectors, achieve onset/duration picks within $\pm0.1$s, and permit real-time traffic-light system design (Paap et al., 2020).

CO$_2$-leakage and anomalous geophysical process detection applies deep, densely connected nets (Seismic-Net) to continuous time windows, capturing varied event morphologies and delivering precision/recall up to 0.89/0.92, with robust resistance to non-seismic artifacts (Wu et al., 2018).

Seismic pattern interpretation with full-volume 3D cube annotation is achieved using deconvolutional U-shaped networks trained on interpreted seismic pattern datasets (SpiNet), automating identification of faults, salt bodies, and stratigraphic patterns, and expediting high-level geophysical modeling (Di, 2018).

6. Implementation Considerations and Future Directions

• Synthetic training data: The use of physics-based waveform synthesis, noise simulation, and network geometry modeling enables robust training for deployment in data-sparse or novel environments (Solda et al., 31 Aug 2025).
• Integration of wireless IoT and LPWANs: Modern deployments employ cost-effective LPWAN technologies (LoRaWAN, NB-IoT) for distributed sensing, cloud aggregation, and trigger/ambient noise streaming with scalable duty-cycled architectures, multi-year lifetimes, and demonstrated QC over networks of $10^3$ nodes (Jamali-Rad et al., 2024).
• Computational scaling: Network measures based on pairwise statistics ($O(N^2)$) remain tractable at regional array scales ($N<500$), while deep-learning architectures can exploit GPU parallelism for both training and inference in near real time.
• Limitations and open issues: Key challenges include calibration of detection/association thresholds for variable network density, the impact of geometric or regional heterogeneity (e.g., in subduction domains), handling of overlapping or emergent low-SNR events, and continued development of generalizable, physically-informed models for cross-region transfer.

Network-based seismic approaches, ranging from physical–statistical network models and correlation/coherence-based detection to modern deep learning frameworks and IoT-enabled wireless sensing, have established new standards for catalog robustness, completeness, and real-time event characterization—fundamentally advancing both operational and research seismology (Solda et al., 31 Aug 2025, Ashkenazy et al., 2023, Ferreira et al., 2014, Zhu et al., 2018, Jamali-Rad et al., 2024).