RECOVAR: Representation Covariances on Deep Latent Spaces for Seismic Event Detection
Organizations: Department of Physics, Bogazici University, Istanbul
Abstract
Modern deep learning methods have improved earthquake detection, but the most widely used approaches typically rely on supervised learning from large labeled waveform datasets. Constructing such datasets is time-consuming and introduces biases from existing catalogs, detection pipelines, and human review, which can limit cross-dataset generalization and sensitivity to underrepresented events. We present RECOVAR, an unsupervised method for seismic event detection that can be trained directly from raw waveforms without using event/noise labels. The method uses an ensemble of convolutional autoencoders and detects events by measuring temporal alignment in their compressed latent representations through a cross-covariance-based metric. The results suggest that RECOVAR can reach the performance of supervised detectors on public benchmark datasets, remain robust under cross-dataset scenarios, and stay effective on imbalanced continuous waveform data where event-containing windows make up less than of the training stream. The method can also complement supervised pipelines by acting as a filter improving precision in a multi-station setting. Because the detection principle makes no domain-specific assumptions, the approach may extend to signal-detection problems beyond seismology.
Figures & tables
| Training dataset | Model | Testing dataset | |
| INSTANCE | STEAD | ||
| INSTANCE | Ours | ||
| PhaseNet | |||
| EQTransformer | |||
| STEAD | Ours | ||
| PhaseNet | |||
| Testing dataset | Training dataset | ||||||
| INSTANCE | Silivri | Balıkesir–Sındırgı | |||||
| Standard | No dilation | Dilation | No dilation | Dilation | |||
| INSTANCE | 0.977 | 0.896 | 0.969 | +0.073 | 0.973 | 0.970 | -0.003 |
| Silivri | 0.977 | 0.698 | 0.950 | +0.252 | 0.965 | 0.961 | -0.004 |
| Balıkesir–Sındırgı | 0.996 | 0.960 | 0.994 | +0.034 | 0.995 | 0.996 | +0.001 |
| Best recall | Best | |||||
| RECOVAR model | FP filtered | TP lost | FP filtered | |||
| INSTANCE-trained | 10.1% | 9.0% | 44.9% | |||
| Balıkesir–Sındırgı, no dilation | 12.3% | 5.1% | 42.8% | |||
| Balıkesir–Sındırgı, with dilation | 10.9% | 4.5% | 38.4% | |||
Seismology
Seismology studies earthquakes and seismic waves to understand Earth’s structure, monitor faults, and assess geological and infrastructure hazards.
See all 48 papers