CRLBD: A deep-learning dataset for the Corinth Rift Laboratory Near-Fault Observatory

Abstract Deep-Learning (DL) phase pickers are becoming increasingly popular in seismology, permitting the processing of large data volumes and reducing manual biases. While general purpose models exist, trained on very large datasets, it is essential to benchmark them on diverse seismic scenarios to explore their potential as a future operational tool. To this end, we created the Corinth Rift Laboratory Benchmark Dataset (CRLBD), a collection of predominantly local earthquake recordings from the Western Gulf of Corinth, an area monitored by the Corinth Rift Laboratory Near-Fault Observatory (CRL-NFO). The dataset creation process involved the acquisition of phase picks from the Seismological Laboratory of the National and Kapodistrian University of Athens and the Geodynamic Institute of the National Observatory of Athens. Waveforms were recorded by 128 stations, belonging to 8 networks, at a maximum epicentral distance of 150 km (with over 99% being within 100 km). We performed several quality assurance steps and calculated extensive metrics. CRLBD consists of 224,573 recordings of 13,918 earthquakes, labelled by an equal number of P arrival times and 127,629 S picks. We then assessed the performance of two DL architectures (EQTransformer and PhaseNet) trained on three different datasets (ETHZ, INSTANCE and STEAD). Our results indicated that models trained on INSTANCE performed best in all evaluation tasks (event detection, phase classification, and picking accuracy). CRLBD is an important step towards DL monitoring in CRL-NFO, creating the opportunity for experimentation with training models on the dataset itself or transfer-learning. At the same time, CRLBD fills an important gap among published benchmark datasets, by providing a large number of near-source records. This work was conducted in the context of the TRANSFORM2 project, as it aims on improving the digital infrastructure in the NFO