SeisMamba: Low-Latency Single-Station Seismic Magnitude Estimation for Spatially Distributed Earthquake Early Warning

Motivation

• Dense regional networks, historical catalogues, and region-specific calibration create a spatial coverage barrier for high-risk regions with sparse monitoring infrastructure.This motivates low-cost magnitude estimation from three-component waveforms at a single station.Source: §1 · p. 1

Challenges

• Single-station EEW requires spatially distributed sensing, low-latency local inference, and robustness to regional distribution shift.Models must remain useful when deployed in regions with waveform characteristics differing from training data.Source: §1 · p. 1

Research gap

• The potential of low-latency single-station magnitude estimation under regional deployment shift remains underexplored.Source: §1 · p. 2

Contribution

• SeisMamba is proposed as a compact Mamba-based architecture for low-cost single-station earthquake magnitude estimation.Its design targets spatially distributed EEW requirements rather than treating magnitude estimation as a standalone waveform regression task.Source: §1 · p. 2
• SeisMamba combines convolutional waveform encoding with sparsely placed selective state-space blocks and multi-scale feature fusion.Source: §1 · p. 2

Single-station seismic magnitude estimation

• Single-station EEW estimates earthquake properties from local waveform observations, reducing reliance on dense regional networks.Early learning-based methods used engineered P-wave features, transfer learning, or auxiliary source information.Source: §2 · p. 2
• MagNet demonstrated that neural models can estimate earthquake magnitude directly from raw waveform data.This followed earlier approaches based on engineered features, transfer learning, or auxiliary source information.Source: §2 · p. 2

Efficient sequence modelling for spatially distributed seismic sensing

• Transformers model long-range sequences effectively, but attention has quadratic complexity with sequence length.This complexity motivates alternatives for efficient processing of long seismic waveforms.Source: §2 · p. 2
• Mamba offers a selective state-space alternative with linear-time scaling and input-dependent state transitions.In seismology, PhaseMamba indicates that Mamba-style models can capture useful waveform structure for phase picking.Source: §2 · p. 2

Task formulation

• Given a 30-second ENZ waveform sampled at 100 Hz, SeisMamba predicts a scalar earthquake magnitude from a single station.The deployment setting assumes inference near one sensing device.Source: §3.1 · p. 2

Design requirements

• The model targets accurate magnitude regression, low inference latency, and robustness across geographically different seismic regions.STEAD measures average predictive performance, while regional hold-out testing evaluates representation usefulness in unseen regions.Source: §3.1 · p. 2

Architecture

• Its encoder uses hierarchical one-dimensional convolutions, sparsely placed Mamba blocks, multi-scale feature fusion, and two prediction heads.Convolutions extract local waveform features, while sparse Mamba blocks model longer temporal context.Source: §3.2 · p. 2, Figure 1 · p. 3

Efficient sequence modelling

• Mamba blocks operate only at selected encoder resolutions after convolutions shorten the sequences, with bypass paths avoiding unnecessary long-context processing.This sparse placement is motivated by efficiency and preserves direct propagation of local convolutional features where long-context modelling is unnecessary.Source: §3.2 · p. 2

Prediction heads

• The scalar head concatenates pooled descriptors across encoder depths, while an auxiliary temporal head predicts a P-wave-aligned magnitude trajectory.The scalar output is used for benchmark comparison; temporal targets are zero before P-wave arrival and equal to event magnitude afterwards.Source: §3.2 · p. 2

STEAD benchmark

• On STEAD, SeisMamba achieves the best MSE, RMSE, and R^2 among tested models while remaining close to the best MAE.The benchmark uses 30-second three-component waveforms with only a 1–40 Hz fourth-order zerophase Butterworth band-pass filter and no hand-crafted P-wave features, source-location inputs, or station-specific calibration.Source: §4.2 · p. 3, §4.1 · p. 3
• 0.55 ms per batch is SeisMamba’s STEAD inference time on an NVIDIA T4 GPU, slightly slower than PhaseNet but substantially faster than EQTransformer, AMAG, and U-Mamba.The comparison includes MagNet, PhaseNet, EQTransformer, AMAG, and U-Mamba; PhaseNet and EQTransformer use adapted regression heads.Source: §4.2 · p. 3, §4.1 · p. 3, Table 1 · p. 3

Regional hold-out

• 5,478 recordings are excluded from training in the Chile–Taiwan regional hold-out, creating testing on geographically unseen seismic regions.The held-out split contains a broad magnitude range, including a substantial proportion of medium-to-large events.Source: §4.3 · pp. 3–4, Figure 2 · p. 3
• R^2 = 0.8518 and MAE = 0.2808 are achieved on the harder Chile–Taiwan hold-out, with latency remaining nearly unchanged at 0.58 ms.The observed performance drop from the standard benchmark is attributed to regional variation in attenuation, site effects, source properties, and station conditions.Source: §4.3 · pp. 3–4

Ablation study

• Ablations show that deeper encoders increase latency, dense Mamba placement performs worse than sparse placement, and removing multi-scale fusion degrades accuracy.These results indicate that state-space modelling is most useful after local waveform compression and that shallow and deep features provide complementary magnitude evidence.Source: §4.4 · p. 4

• Selective state-space modelling can improve single-station magnitude accuracy while preserving low inference latency for spatially distributed earthquake early warning.This supports deploying models near individual stations.Source: §5 · p. 4
• Regional waveform shift remains unresolved, as attenuation, site effects, source properties, and station conditions affect transfer to unseen seismic regions.The Chile–Taiwan hold-out revealed a performance drop under regional distribution shift.Source: §5 · p. 4
• The auxiliary temporal head provides time-resolved magnitude supervision and shows how estimates evolve as waveform evidence accumulates, but it is not a complete uncertainty estimator.Operational alerting would still require calibrated uncertainty estimates and decision rules.Source: §5 · p. 4

• SeisMamba is a lightweight selective state-space architecture for single-station earthquake magnitude estimation.Source: §6 · p. 4
• The architecture combines local waveform encoding, sparse Mamba sequence modelling, multi-scale fusion, and auxiliary temporal supervision.Source: §6 · p. 4
• SeisMamba improves the accuracy–latency trade-off on a global seismic benchmark while retaining useful performance under a regional hold-out setting.Source: §6 · p. 4