Application of the pair-input deep learning model for seismicity reassessment in Cameroon

Abstract

Applying advanced signal analysis and feature extraction methods to seismic datasets facilitates the extraction of key information from these valuable resources. In particular, advanced deep learning models can play a crucial role in this extraction process from seismic waveforms. This study aims to analyze the seismicity of the Cameroon area by deeply scanning an archived dataset from a temporary seismic network composed of 32 broadband sensors that operated in Cameroon from March 2005 to February 2007. We thoroughly scan the dataset to detect local earthquakes, calculate an updated crustal velocity model for the region by incorporating the acquired earthquake bulletin, and perform a joint inversion scheme. The DL-based workflow implemented in this study utilizes a pair-input deep learning model for scanning the seismic phases in the dataset, followed by a seismic phase picking process. Applying the Rapid Earthquake Association and Location method allows for the association of detected phases with local seismic events. By combining a set of 17,261 P-picked and 14,239 S-picked phases associated with 473 well-located earthquakes with magnitudes ranging from $$1.3 \le M_L \le 4.6$$ 1.3 ≤ M L ≤ 4.6 , we implement a joint inversion to estimate an updated 1D crustal velocity model and hypocentral parameters. The obtained model reveals layers with thicknesses of 82,034, and 54 km, extending to a depth of 80 km, with P-wave velocities ($$V_{{\text{p}}}$$ V p ) of 6.1, 6.4, 6.6, 7.6, 8.25, and 8.5 km/s, and S-wave velocities ($$V_{{\text{s}}}$$ V s ) of 3.57, 3.74, 3.86, 4.44, 4.82, and 5.0 km/s, respectively. The newly detected seismic events predominantly cluster in the Central Cameroon Shear Zone (CCSZ), the east flank of Mount Cameroon, the region between Mount Cameroon and Bioko Island, and the southern part of Bioko Island. The validity of this approach is further supported by a newly compiled catalog for this period, which is 1.35 times more extensive than the previously reported catalog for this dataset. Finally, the study presents a 3D time-lapse animation illustrating the sequence of detected earthquakes.