Repopulating Earthquake Catalog in the Hindu Kush-Pamir Region using Attentive Deep Learning Model

Seismology data is overgrowing and is outpacing the development of processing algorithms. This tremendous increase in high-quality data can help better understand the earthquake processes related to the geology of active seismic regions such as the Hindu Kush. Most traditional detection algorithms are computationally inefficient compared to the amount of seismological data available and fail to detect low magnitude and noisy events. Deep learning algorithms are known for their applicability on large datasets with less runtime. Event detection and phase detection can be considered a supervised deep-learning problem quite similar to image recognition. In this study, we have implemented a hierarchical attention mechanism-based deep learning model for simultaneously phase picking and earthquake detection. This model is trained using Stanford Earthquake Dataset (STEAD), a globally disturbed labeled seismic dataset. We used this trained deep learning model to detect earthquake signals and pick P and S phases in the Hindu Kush - Pamir region for twelve months of continuous data spread across 45 different stations. A rigorous selection criterion based on detection, P and S phase probabilities, and other parameters has been used to associate the phases from different stations and to locate the earthquake. Our model detected almost six times more earthquakes than previously existed in the catalog. The algorithm picked P and S phases with a high level of precision, comparable to human analyst picks. Furthermore, pinpointing these events allowed us to define the S-shaped seismic zone in the Pamir-Hindu Kush region and better comprehend the deformation caused by Eurasian- Indian plate motion.