Automatic detection of atypical seismic events through machine learning models trained on modulation spectrum representations of OBS datasets

The Ecuadorian subduction zone is one of the few subduction zones where aseismic slip occurs in the shallow segment of the megathrust fault. This aseismic slip appears to be characterized by seismic swarms. So far, no non-volcanic tremor has been detected using classical methods. This may be partially attributed to the fact that the previous deployments were mainly on land and only sparsely offshore, away from the expected locus of potential tremors.

During the HIPER marine campaign (2022, 15/03-12/04), we deployed around 40 OBS on a 3D grid to  image the structure of the Ecuadorian subduction zone in the region of the 2016 Mw 7.8 Pedernales earthquake. 

An automatic machine-learning CNN model was developed, relying on modulation spectrum representations of the seismic signals acquired from the OBS network. This approach is rooted in the detection of typical/atypical patterns in animal vocalizations or human speech, as it has been demonstrated to be highly effective in profiling and detecting the "natural" variations from noise - how the modulation patterns (the “timbre” and “prosody”) evolve around the carrier frequency (the “pitch"). 

Thus, the representation dataset in our approach consists of streams of time-varying images 2D+t (carrier vs modulation frequencies) computed for each unidimensional directional seismic time series. This approach was tested and proved to be both discriminatory and efficient in validating the detection of tremors obtained on OBS seismic signals extracted from the SEIS-PNSN tremors dataset from the Cascadia subduction zone.

For the first time, we have recorded seismic activity on a dense offshore network over a one-month-long period, which will reveal whether tremors occurred in the region of the Pedernales earthquake, a region which is prone to aseismic and seismic slip.