Deep learning application to earthquake detection and characterisation

While large earthquakes are intensively studied, it is also important to improve the number of small magnitude event detections since their spatial and temporal evolution could help us to constrain and understand the seasonal stress changes. Considering existing CNN models for local earthquake detection and characterisation developed for the Central Apennines, Italy (developed by J. Majstorović), the models will be first applied to a new study area, the Pollino region in the Southern Apennines, Italy. For this we will use a station that is located close to the seismic fault and that has been recording for 11 years without interruption. The new detections will be compared to the catalog of existing events maintained by the Istituto Nazionale di Geofisica e Vulcanologia. This catalog might miss some small magnitude events and the CNN pipeline will help to extend it. Once the near-fault catalog will be developed, we will search for seasonal correlation patterns between the Global Positioning System (GPS) time series and the local seismicity. Previous studies have already shown that the dominant signal in GPS data is a yearly modulation that correlates with the rain drop season. Yet, a question remains : does this variation in the amount of underground water also affect the local seismicity ? The last part of the internship will be to perform a transfer learning by fine-tuning the existing detector CNN model with the local events extracted for the Pollino region. The performance of the new CNN model will be compared with the previous one to study how the developed CNN models can be best generalised in order to be used in different regions.