The Effectiveness of Automatic Seismic Phase Picking and Detection Capabilities of Deep Learning Methods for Local On- and Offshore Seismic Data: The Case of the Terceira Rift, Azores, Portugal

High-quality earthquake catalogues for seismic hazard and tectonic assessment of a region, primarily require high-accuracy and high-precision hypocentral locations, along with a low completeness magnitude. The Terceira Rift is an active structure which accommodates a slow transtensional deformation of about 5 mm/yr, induced by eastward differential displacement between the Nubian and Eurasian plates. Due to its active and intense seismicity and volcanism, the Terceira Rift constitutes a natural laboratory to investigate active rifting processes. In our research, novel data from the UPFLOW project, encompassing 49 Ocean Bottom Seismometers, covering the Azores-Madeira-Canary Islands region, will be combined with existing land stations, to analyse the seismicity of the Terceira Rift. The detection capabilities of the existing traditional land seismic network have demonstrated weaknesses in detecting smaller events, highlighting its limitations in precise event location when classical analysis methods are applied. It is expected that well distributed network, and the use of Machine Learning methods, will provide us with the possibility to detect events of smaller magnitude with high-accuracy and high-precision. In this study, we tested the detection and phase picking capabilities of the deep learning phase picker EQTransformer for the land network alone and compared its performance with a manually analysed catalogue from the same network. Data used consist of ten days of continuous waveform from IPMA stations network, between October 10 and 20, 2021. Waveform pre-processing included the removal of instrument response, detrending, applying maximum taper of 1%, and high-pass filter at 2 Hz. For pick classification we used the cut-off threshold of 0.20 and 0.15 for P and S phases, respectively. Although there are some outliers for both P and S pick probabilities, we found that the median probability is approximately ~90% for P phase, and ~70% for S phase. Time differences between the catalogue pick-time and EQTransformer pick-time range approximately between -0.5 and +0.5 seconds for P phase and -1.0 and +1.0 seconds for S phase, denoting a high picking precision of EQTransformer. Within the time window analysed, deep learning methods detected more events than those in manually analysed catalogue. We also present initial results of our analysis using both deep learning networks EQTransformer and PickBlue applied to ocean bottom seismology recordings from the UPFLOW passive array deployed in the Azores-Madeira-Canaries region between June 2021 and August 2022.