The three datasets presented here are high-resolution catalogs containing origin time of seismic events for the same region and time range that have derived using AI-based techniques and a matched filter search. The corresponding standard catalogs from the agencies AFAD and KOERI are available under https://tdvms.afad.gov.tr/ (last accessed 28/07/2022) and http://www.koeri.boun.edu.tr/sismo/2/earthquake-catalog/ (last accessed 28/07/2022), respectively, when searching in the bulletin for longitude 28.80-29.10, latitude 40.4-40.625, and from November 1st 2018 to January 31th, 2019.

Specifications for the three catalogs are.

(i) Catalog derived utilizing AI-based techniques. We applied the PhaseNet deep learning method (Zhu & Beroza, 2019) to detect and pick the P-and S- waves of seismic events embedded in continuous seismic recordings from 16 stations surrounding the region of interest resampled at 100 Hz. The method was trained on a dataset from Northern California, but has been shown to generalize well to other tectonic settings. The picks were associated into seismic events using the GaMMA association method (Zhu et al., 2022). Manual check of the waveforms from all detections led to 516 seismic events with clear waveforms retained for further processing.

(ii) Template matching catalog A. We applied the matched filter algorithm EQcorrscan (Chamberlain et al., 2017) to the two nearby seismic stations with the largest data recovery during the period of interest, ARMT and MDNY. We utilized 14 manually picked template events with M > 2 that occurred in the region of interest during the analyzed time period, which were recorded in both stations. As a first criteria to remove false detections, we retained only detections exhibiting a Median Absolute Deviation (MAD) larger than eight. We required detections from different templates to be at least 1.5 seconds apart. To remove duplicate detections (e.g., detections of the same event by different templates), we retained the detections with the highest average correlation if multiple detections occurred within 2.5 seconds. As a second criteria, we calculated cross-correlation derived phase-picks. A pick was declared if the maximum normalized correlation between the signal of the template event and of the detection exceeds 0.7. We correlated the signals in a short window of ±0.3 seconds around the assumed pick time based on a time-shifted version of the template phase-pick. We retained the S-pick exhibiting the higher cross-correlation value with respect to the template. Following this step, we considered only detections with ≥ 2 picks. In case of events with only two picks we ensured that that were from the same station to have control on the ts-tp and therefore the distance of the event from the detecting station. This catalog contains 2,462 seismic events (all manually reviewed) with magnitudes MW in the range [-2.4, 4.5]. Since we were not able to locate the events from this catalog, we considered as “origin time” the time of the first arrival.

(iii) Template matching catalog B. We derived a second template matching catalog utilizing twelve of the closest seismic stations displaying high seismic data recovery during the analyzed time period. An initial list of detections was generated following the same steps as for the Template Matching Catalog A, with the additional requirement that all detections must contain at least one picks from one of the two closest stations, ARMT and MDNY. All detections from this catalog were also manually reviewed. The full description of the data processing and creation of the catalog is provided in the article “Stress changes can trigger earthquake sequences in a hydrothermal region south of Istanbul” by Martínez-Garzón et al., currently under review in Geophysical Research Letters.