SCALODEEP: A Highly Generalized Deep Learning Framework for Real‐Time Earthquake Detection

The detection of earthquake signals is a fundamental yet challenging task in observational seismology. A robust automatic earthquake detection algorithm is strongly demanded in view of the ever‐growing global seismic dataset. Here, we develop an automatic earthquake detection framework based on a deep learning approach (SCALODEEP). It extracts high‐order features embedded in three‐component seismograms by encoding a time‐frequency representation of the data (scalogram) into a deep network with skip connections. The SCALODEEP is trained and validated on an open‐source dataset from North California, and then employed to seismicity detection in four areas, including Arkansas, Japan, Texas, and Egypt. Despite vastly varying characteristics of regional earthquakes (e.g., focal mechanism, duration, and noise level), SCALODEEP successfully detects seismic signals over a broad range of local magnitudes (as low as −1.3ML) and outperforms conventional algorithms such as STA/LTA, FAST, and template matching. Compared to recently proposed deep learning based frameworks (e.g., CRED and Earthquake transformer), SCALODEEP achieves a superior generalization ability via a sophisticated network architecture. In summary, our study offers a promising new tool to improve existing earthquake detection systems and, as importantly, sheds light on designing an effective deep learning network for generalized earthquake detection. Plain Language Summary An efficient and reliable earthquake detection based on deep learning is of great interest to a broad scope of geoscience community. One question that has been constantly raised by seismic practitioners is whether/how we can generalize the trained network to be widely applicable. We introduce a new deep learning method for generalized earthquake detection. Our network includes a very deep architecture with 24,629,053 parameters, and its generalization ability is further augmented by implementations of time‐frequency representation of seismic data and skip connections. We compare the performance of the new framework with the state‐of‐the‐art FAST, template matching, and CRED methods, and demonstrate its advantages. To demonstrate the potential in practical usage, we train the network using a community dataset from North California, and test its generalization ability on four independent regional datasets from Arkansas, Japan, Texas, and Egypt. Key Points We introduce a new deep learning method, SCALODEEP, for generalized earthquake de