The Korean infrasound catalogue (1999–2022)

SUMMARY

The Korean infrasound catalogue (KIC) covers 1999–2022 and characterizes a rich variety of source types as well as document the effects of the time-varying atmosphere on event detection and location across the Korean Peninsula. The KIC is produced using data from six South Korean infrasound arrays that are cooperatively operated by Southern Methodist University and Korea Institute of Geoscience and Mineral Resources. Signal detection relies on an Adaptive F-Detector that estimates arrival time and backazimuth, which draws a distinction between detection and parameter estimation. Detections and associated parameters are input into a Bayesian Infrasonic Source Location procedure. The resulting KIC contains 38 455 infrasound events and documents repeated events from several locations. The catalogue includes many anthropogenic sources such as an industrial chemical explosion, explosions at limestone open-pit mines and quarries, North Korean underground nuclear explosions and other atmospheric or underwater events of unknown origin. Most events in the KIC occur during working hours and days, suggesting a dominance of human-related signals. The expansion of infrasound arrays over the years in South Korea and the inclusion of data from the International Monitoring System infrasound stations in Russia and Japan increase the number of infrasound events and improve location accuracy because of the increase in azimuthal station coverage. A review of selected events and associated signals at multiple arrays provides a location quality assessment. We quantify infrasound events that have accompanying seismic arrivals (seismoacoustic events) to support the source type assessment. Ray tracing using the Ground-to-Space (G2S) atmospheric model generally predicts observed arrivals when strong stratospheric winds exist, although the predicted arrival times have significant discrepancies. In some cases, local atmospheric data better captures small-scale variations in the wind velocity of the shallow atmosphere and can improve arrival time predictions that are not well matched by the G2S model. The analysis of selected events also illustrates the importance of topographic effects on tropospheric infrasound propagation at local distances. The KIC is the first infrasound catalogue compiled in this region, and it can serve as a valuable data set in developing more robust infrasound source localization and characterization methods.

1 INTRODUCTION

Data from early seismic networks allowed seismologists to build the first catalogues of seismic events (Ammon et al. 2020). Today, earthquake catalogues are fundamental to seismological studies and are often the starting point for a variety of scientific and practical goals such as identifying faults, quantifying types and sizes of sources and assessing seismic hazards. In the same vein, networks of infrasound stations provide the opportunity to build catalogues of infrasound events for both scientific and practical purposes. On a global scale, infrasound data from the International Monitoring System (IMS) operated by the Comprehensive Nuclear Test Ban Treaty Organization (CTBTO) has supported the construction of catalogues that include a variety of natural and anthropogenic sources (Brachet et al. 2010). Hupe et al. (2022) published an infrasound detection list based on data from 53 IMS infrasound stations covering 18 yr (2003–2020) using the Progressive Multi-Channel Correlation (PMCC, Cansi 1995) method. At regional scales, networks in the Western United States (Walker et al. 2011; Park et al. 2014) and Europe (Pilger et al. 2018; Bondár et al. 2022) have enabled an assessment of the types of events that generate infrasound signals in these areas. A local or regional scale (<150 km) infrasound event catalogue is also important in studying repeated events such as experimental explosions, quarry and other engineering blasts and volcanic activities. Rodd et al. (2023) document a multimodal event catalogue over a 2-week period including infrasound events in the southwestern United States. Such catalogues can be used in machine learning-based infrasound studies, particularly in supervised learning studies that need large, labelled data sets (Albert & Linville 2020; Witsil et al. 2022; Ronac Giannone et al. 2024). This study focuses on the development of such a catalogue of infrasound events using a dense and long-running network of infrasound arrays in South Korea (SK), supplemented by two nearby IMS infrasound stations in Russia and Japan. The Korean Infrasound Catalogue (KIC) is unique in terms of its extended operational time span and the rich variety of source types, including underground nuclear tests, mining explosions, accidental industrial explosions and various other engineering activities.

Infrasound studies experienced a renaissance beginning in 1996 because of the initial investment in a global network of IMS stations and the support of data processing by CTBTO, which now provides more than 20 yr of global infrasound data. Infrasound event catalogues that span these long durations are used to identify repeating sources from common locations, quantify the distribution of sources in time and space and document the effects of a time-varying atmosphere. These catalogues provide the basis to quantify seasonal variations in infrasound wave propagation as well as identify sources for more detailed analysis. Walker et al. (2011) produced an infrasonic catalogue (2007–2008) based on acoustic-to-seismic coupled signals observed across the dense set of USArray seismic stations deployed in the western United States. In a complementary study, Park et al. (2014) used infrasound recordings from 12 regional arrays in the western United States to produce an infrasound catalogue covering the time period from 2010 to 2012. There was significant commonality between the two catalogues although they used different types of data (seismic versus infrasound), time periods and methods suggesting common source areas. Pilger et al. (2018) located infrasound events using 24 European infrasound arrays (including 8 IMS infrasound arrays) covering 16 yr (from 2000 to 2015). Bondár et al. (2022) documented infrasound-only and seismoacoustic events from 2017 to 2020 based on data from multiple infrasound arrays and seismic stations in Europe.

The primary purpose of this study is to produce and assess a regional infrasound catalogue for the Korean Peninsula for the time period of 1999–2022. The KIC includes a variety of source types with some repeated sources that illustrate spatiotemporal variations. Since 1999, Southern Methodist University (SMU) and Korea Institute of Geoscience and Mineral Resources (KIGAM) have cooperatively designed, built, operated and upgraded six seismoacoustic arrays in SK. This collaboration has developed instrumentation efficiencies based on an evolutionary design, testing, installation and operation of collocated seismic and infrasound sensors as well as supported monitoring tools designed to assess both natural and anthropogenic sources across the Korean Peninsula (Stump et al. 2004; 2022). These SK seismoacoustic arrays provide an integrated data set that is used to study a variety of infrasonic sources as well as to discriminate human-induced events from natural processes, including earthquakes. The data have been previously used for signal characterization of mining explosions (Che et al. 2002, 2009a, 2019, 2022), underground mine collapses (Che et al. 2018), North Korea (NK)’s underground nuclear explosions (UNEs) (Che et al. 2007; 2009b; 2014, 2022, 2022; Kim et al. 2009; Park et al. 2018a), industrial accidents (Arrowsmith et al. 2021), large earthquakes (Kim et al. 2004; Che et al. 2013) and early warning of the discharge of large volumes of water from dams (Che et al. 2023). Other studies have assessed infrasound sensor calibration (Kim et al. 2010), the performance of infrasound detectors (Park et al. 2017), the effects of infrasound propagation (McKenna et al. 2008; Che et al. 2011), local site effects on signals (Park et al. 2016) and atmosphere model inversion using infrasound observations (Park et al. 2023). Based on a detection capability study by Che et al. (2017), the Korean infrasound network with IMS infrasound data provides a detection threshold that is three times lower than that with only IMS data, while network performance is significantly enhanced during the summer when the atmosphere supports more efficient propagation to the arrays. Recently, Stump et al. (2022) documented the history and details of the six seismoacoustic arrays in SK accompanied by the release of the data from the six UNEs conducted by NK. The seismoacoustic data quality is routinely monitored to identify and respond to technical issues to ensure high-quality data as well as quantify seismic and infrasound noise characteristics (Park et al. 2023).

This paper focuses on documenting the regional KIC for 1999 to 2022 and includes assessments of events and atmospheric (wind and temperature) variations that affect infrasound propagation across the Korean Peninsula. Section 2 summarizes infrasound array locations and configurations including sensor details. The procedures used to produce the KIC are outlined in Section 3, including automated signal detection at each array, the association of detections between arrays and finally, the formal location procedure with error estimates. Results based on the assessment of the KIC are summarized in Section 4: first, the automated infrasound detections and event locations increase as the number of arrays and instruments in each array increases; second, most infrasound events within the Korean Peninsula are related to human activities based on both location and time of occurrence; third, seasonal, weekly, and diurnal variations of infrasound sources are documented, providing insight into the temporal and spatial variations of the atmosphere; and finally, the KIC is used as a trigger to review events of interest and comparison of observations with predictions based on a variety of atmospheric models. Section 5 contains a closing discussion, and Section 6 summarizes our findings and provides recommendations.

2 DATA

Infrasound events are formed using the data from a relatively dense (average interarray spacing of ∼180 km) and long-running (∼24 yr) network of six infrasound arrays in SK, supplemented by data from two nearby IMS infrasound stations in Russia and Japan (Fig. 1). The five seismoacoustic arrays, BRDAR, CHNAR, KMPAR, KSGAR and YPDAR, and one infrasound array, TJIAR, are distributed across the southern Korean Peninsula (Fig. 1). The average distance to the closest three arrays across the study area is estimated and contoured in Fig. 1 as it is critical to the formation of infrasound events. The densest array coverage is along the border between NK and SK where the average distance to the three closest arrays falls below 200 km. The array coverage decreases beyond the Korean Peninsula with locations to the northeast and southeast having improved azimuthal coverage from the IMS arrays (IS45 and IS30).

At CHNAR and KSGAR, four-element 1-km aperture seismo-acoustic arrays were installed, with each element composed of one seismometer and a small (60–70 m) aperture infrasound subarray. BRDAR is similar with an additional element. CHNAR and KSGAR have a total of 11 infrasound sensors and 4 seismometers while BRDAR has a total of 13 infrasound sensors and 5 seismometers. KMPAR, TJIAR and YPDAR are composed of 6, 5 and 4 infrasound sensors configured with apertures ranging from 0.2 to 1.0 km. KMPAR and YPDAR have one and two, three-component seismometers, respectively, while TJIAR has no seismometer within the array. Infrasound sensors, originally installed at CHNAR and KSGAR in 1999 and 2003, respectively, were based on an SMU design using Validyne electronics (Stump et al. 2004) and were replaced later in January 2004 with Chaparral Physics Model 2 sensors (flat response from 0.1 to 200 Hz) recorded by 24-bit digitizers. KMPAR is sampled at 100 samples s⁻¹, while all other arrays are sampled at 40 samples s⁻¹. Each infrasound sensor is attached to ten, 8-m porous hoses connected at the centre in a radial configuration to reduce the background noise generated by turbulent winds at the boundary layer. McComas et al. (2021) quantified the performance of noise reduction systems (4-, 10- porous hoses and fabric dome), suggesting that the 10-hose filter system provides better signal detectability. Damaged hoses due to moisture, rain, snow, animal incursion and human activities can degrade the noise reduction system. Hoses are replaced immediately upon identification of damage and at regular intervals of approximately 3 yr. The large-aperture arrays have weather sensors (4 samples s⁻¹) measuring wind speed, wind direction and temperature, installed at a height of 2 m above the surface. Details of the seismoacoustic arrays, including a description of the seismometers, can be found in Stump et al. (2022).

IS30 and IS45 are equipped with 6 and 4 infrasound MB2000 sensors each with a flat frequency response from 0.01 to 8 Hz and sampled at 20 Hz (Assink et al. 2018). The sensors are attached to pipe arrays deployed in a rosette to reduce wind-generated noise and enhance low-frequency infrasound signals (Christie & Campus 2010). At these sites, an anemometer is installed at a height of 2.0 m above the surface, and the temperature sensor and an absolute barometer are installed at a height of 1.0 m.

3 METHODS AND DATA PROCESSING

Construction of the KIC follows a workflow like that used in seismology: (1) signal detection, where signals are detected at individual arrays; (2) signal association, where signals at different sites are grouped to build events and (3) source location, where an optimal event location is produced with an associated uncertainty estimate. In this study, subsequent event characterization for selected events or event clusters (e.g. event identification or estimate of atmospheric propagation path effects) relies on an analyst review process as discussed in Section 4.3.

The automated detection algorithm used in this study has been published and subsequently tested and tuned for the Korean Peninsula (Park et al. 2016). It uses the Adaptive F-Detector (AFD, Arrowsmith et al. 2009) which is based on array cross-correlation estimates and the F-statistic accounting for temporal changes in noise using an adaptive moving window analysis. The arrival times and backazimuths from AFD are input into a grid search to associate signals between multiple arrays (Arrowsmith et al. 2008b). The association procedure searches for groups of arrivals that have backazimuth and delay times corresponding to each gridpoint covering the monitored area. The associated signals are then used in the Bayesian Infrasound Source Location (BISL, Modrak et al. 2010). The BISL method is based on a statistical approach that evaluates a likelihood function, which is a product of backazimuth and arrival time likelihood functions over all arrays and accounts for model and observation uncertainties. The association and location procedures rely on a priori estimates of variations in azimuth and group velocities across a range of travel paths and atmospheric conditions.

We processed nearly 24 yr of infrasound array data using these automatic detection, association and location procedures (codes modified from the InfraMonitor software, see Section 'Data and Resources'). Park et al. (2016) quantified the physical characteristics of the regional infrasound signals across the Korean Peninsula and documented tuning parameters for detection processing. The main cause of the temporal adaptation in AFD is documented as variations in wind speed with time where an adaptive window length of 1 hour captures time-varying high noise levels. Background noise is found to be highest for arrays at or near the ocean with less variation for the inland arrays. BRDAR (array near the ocean) has the highest background noise levels with lower noise for CHNAR (inland) and KSGAR (near the ocean but surrounded by mountains) as quantified by AFD in the detection processing. Park et al. (2017) document the performance of the AFD by comparing the time-dependent F-distribution for different array configurations (small, large and small + large aperture arrays), and background noise levels affected by weather conditions at the individual arrays (low and high wind speed) compared to analysts’ reviews using different frequency bands and p-value thresholds. Additional AFD parameters used in this study include a 1–5 Hz frequency band designed to avoid contamination from microbaroms and ocean-related energy (0.1–0.5 Hz) and a p-value of 0.01.

Multiple short-period signals within a processing time window can be labelled as a single event even though multiple sources may contribute to the waveforms. All array elements in each array were used in the processing. Note that individual arrays have different apertures with varying numbers of array elements (BRDAR, CHNAR and KSGAR—small and large aperture array, KMPAR and YPDAR—small aperture array and TJIAR—large aperture array). In Fig. S1, the array response function for 1–5 Hz at all eight infrasound arrays used in this study is displayed, documenting the effects of array geometries on slowness estimates. Park et al. (2017) show that the optimum performance of the AFD for local infrasound signal detection occurs when all array elements are used under low background noise conditions, while the best performance occurs when only small arrays are used under high background noise conditions. It was found that the use of all array elements in the AFD processing provides the largest number of detections from different array element configurations. Since our study focuses on both local and regional infrasound signals, we use all available elements for each array in this study, realizing that a focused study of a particular event could be improved in some cases.

Infrasound data quality is routinely monitored for infrasound polarity reversal and noise level to identify technical issues related to broken sensors or hoses and quickly trigger routine maintenance activities to improve data quality (Park et al. 2023). Periods of problematic array data were excluded from data processing. For example, data from KSG20 and CHN00 were not used due to an inverted polarity from May 2013 to Mar 2017 and for 2019, respectively.

The association algorithm uses the first detected arrival from the source to reduce computational time. Detections associated with three or more arrays are required to define an event. The association and location procedures focus on tropospheric and stratospheric returns, with group velocities from 240 to 350 m s⁻¹ and standard deviations of azimuth and arrival time set to 5° and 50 s based on experience including the detailed infrasound location study of the North Korean underground nuclear explosions in 2016 (Park et al. 2018a). A uniform prior Probability Density Function (PDF) is used for BISL processing to accommodate unknown group velocities between the sources and arrays. The grid corners of the study area are 27°N, 111°E and 46°N, 145°E. Further analysis of ground truth data could refine the assumed input parameters, including the use of priors that change with seasons (Che et al. 2011) in order to take advantage of enhancements to BISL that incorporate seasonal variations in the ground-to-space (G2S) atmospheric models (Marcillo et al. 2014) into the priors. This initial study does not take these additional seasonal effects into account, and thus the estimated location errors are expected to be large and might be reduced in subsequent studies (but the 95 per cent confidence contour of the event includes the possible location as shown in Fig. S3a).

4 RESULTS

This section summarizes the KIC illustrating infrasound detections (Section 4.1) and event locations (Section 4.2) from the automatic procedures. Most infrasound events across the Korean Peninsula appear to be related to human activities based on the time of occurrence and association of infrasound hotspots with areas of known human activity (Section 4.3). The seasonal, weekly and diurnal variations of the infrasound sources in some of these areas are illustrated. Finally, the KIC provides an opportunity to use events of particular interest to trigger further event analysis using more complex atmospheric models to improve event interpretation.

4.1 Infrasound detection

Detection density plots (Fig. 2) document the total number of automated infrasound detections for a particular array as a function of time and backazimuth for the time period of July 1999 to December 2022 with an expanded scale for CHNAR to illustrate details. Infrasound detections (detection time, backazimuth, phase velocity, correlation, F-statistics and RMS amplitude) are included in the supporting material (Appendix A). CHNAR, KSGAR, BRDAR, KMPAR, TJIAR and YPDAR were installed on 22 July 1999, 21 November 2003, 16 August 2004, 19 May 2005, 14 February 2011 and 24 May 2011 with IS30 and IS45 added to the SMU database on 17 and 22 March 2012 (IS30 and IS45 are in the CTBTO's database from 2005 and 2008 but are not included in this study). At CHNAR there are consistent backazimuth from the northwest during spring, fall and winter, while backazimuths change to the east during the summer. The arrays are sorted by year of first data availability in Fig. 2(b). The largest number of detections occur during the winter and are typically from the northwest, while those during the summer are generally from the southeast (i.e. CHNAR, KSGAR, BRDAR, KMPAR and YPDAR). Park et al. (2016) document a strong correlation between seasons, the number of detections and backazimuth estimates. These results correlate with seasonal changes in tropospheric, stratospheric and, to a lesser extent, thermospheric variations in the northern mid-latitudes, predicted by atmospheric models such as G2S. Hupe et al. (2022) similarly note that there are strong correlations between the middle atmosphere wind conditions and detected signals at 53 IMS infrasound stations as quantified in their study of data from 2004 to 2020. Che et al. (2017) illustrate that network performance is strongly affected by seasonal variations in stratospheric winds, with enhanced array detection performance in the summer. Hedlin et al. (2002) note that detection performance improves when there is little variation in wind direction and speed (i.e. consistent wind direction). Local site environments at and around the arrays also contribute to detection performance.

Detections at BRDAR and KSGAR are strongly associated with temporal changes in the surrounding ocean environment based on high local wind speeds and associated ocean noise (Park et al. 2016). At TJIAR which is located in an urban area in the centre of the Korean Peninsula, detections are often associated with local human activities rather than wind directions. IS45 has the most detections from the south and west. IS30 has many detections from the surrounding island, mostly during the winter. Detection statistics at CHNAR, KSGAR, BRDAR and KMPAR are typically associated with high correlation estimates as a result of using both small (∼0.1 km) and large (∼1 km) aperture array data, while TJIAR, YPDAR, IS30 and IS45 produce lower correlation values as these arrays rely on large, ∼1 km, or medium aperture, ∼0.5 km, array data (Fig. S2). Initially, CHNAR and KSGAR had only large aperture array elements that were subsequently upgraded with the addition of small aperture sites in 2005 to improve performance. The initial configuration produced lower correlations until the 2005 upgrade. The arrays that now have both large and small apertures, CHNAR, KSGAR, BRDAR and KMPAR, produce a relatively larger number of detections compared to the other arrays (Fig. 2b), illustrating the advantage of multiple array scales. KMPAR experienced data gaps on some elements, as shown in the irregular pattern of detections. Note that data gaps at IS30 and IS45 are due to data loss in our database system and may not reflect the total IMS data delivered.

Fig. 3 (top panel) plots the cumulative infrasound detections by array as a function of time. CHNAR, KSGAR, BRDAR and KMPAR produce more total detections because of their early installation dates of 1999, 2003, 2004 and 2005, respectively. Many detections at CHNAR are interpreted as local sources during time periods of relatively low wind speed that coincide with stable propagation conditions (McKenna et al. 2008; Park et al. 2016). YPDAR and TJIAR were installed in 2011 and data from the IMS arrays was added in 2012 (when we started acquiring data). The slopes of the cumulative number of detections versus time for individual Korean infrasound arrays are steeper than those for the two IMS arrays, possibly reflecting the added contribution of the small aperture arrays (100–500 m) to the detection of local and regional signals. TJIAR is a single large aperture array (∼1 km) like the geometry of the IMS arrays but also produces a steeper slope plot, suggestive of either its proximity to an urban environment or the high-frequency impact of using porous hoses for noise reduction (McCommas et al. 2021).

There is strong evidence that a significant proportion of the infrasound detections and events identified in this study are related to human activities. This conclusion may also be impacted by focusing the detector on high-frequency data, 1–5 Hz, observed on arrays located in proximity to densely populated urban areas. Histograms summarizing the number of detections as a function of the day of the week (Fig. 4a) and the time of day (Fig. 4b) support the human impact interpretation. Event histograms as a function of local time-of-day rise from 8 to 11 a.m. with a noticeable dip between noon and 2 p.m. and then decay as night approaches. The day-of-week histograms increase from Tuesday through Thursday, with the fewest events on Sunday. The only exceptions to these patterns are observed at TJIAR and IS30 with increased numbers of detections at night and on weekends, suggesting a contribution at these arrays from continuous noise sources, in some cases associated with an urban environment, may be detected independent of normal human activity time and day.

4.2 Infrasound event location

All 38 455 KIC events from December 2004 to December 2022 are documented in this study, and origin times, event locations and uncertainty estimates are included in the supporting material (Appendix B). The cumulative number of events as a function of time is compared to the number of infrasound arrays in the bottom panel of Fig. 3. Event locations begin in December 2004 following the installation of BRDAR, when data from a minimum of three arrays became available. There has been a dramatic increase in the number of events beginning in 2010, which coincides with additional array installations and access to additional data in 2011 and 2012, which improves azimuthal station coverage and improved locations as the number of associated arrays (NAA) evolves over time. The number of events decreases slightly from 2016 to 2018, possibly associated with data gaps at KMPAR and IS30 (white area in Fig. 2b). The number of events decreases again in 2020, possibly related to COVID-19 and the associated reduction in human activities or data outages.

Event locations document areas of repeated infrasound sources, defined as ‘infrasound hot spots’ (Walker et al. 2011). Fig. 5a is a 2-D map in and around the Korean Peninsula depicting the density of infrasound source locations over 18 yr. There is a good correlation between these hot spots and infrasonically active areas associated with industrial surface explosions, limestone mines and quarries in the Korean Peninsula (Che et al. 2002, 2009a, 2011, 2019) as well as areas with other known engineering activities. Events along the west coast of SK do not correlate with known sources and may warrant further investigation. In the western United States (Walker et al. 2011; Park et al. 2014) and Europe (Pilger et al. 2018; Bondár et al. 2022), some infrasound hot spots were spatially associated with military exercise areas. While infrasound events in the western United States are not strongly correlated with shallow mining and quarry blasts (Walker et al. 2011; Park et al. 2014), there is a significant correlation between these source types and infrasound events across the Korean Peninsula (Che et al. 2019).

All infrasound events are plotted in Fig. S3, with events that have 95 per cent uncertainty contours (<25 000 km²) noted. The 95 per cent uncertainty contours depend on the number of observing arrays as well as the azimuthal distribution of the arrays around the event with ‘good’ events defined as relatively small uncertainty areas (<25 000 km²), as shown in the example events in Fig. S3a. This level of uncertainty area was selected based on an empirical study that documented that most real events, excluding falsely associated events, have uncertainty areas smaller than 25 000 km², as illustrated by the NK UNEs (Park et al. 2018a). Locations with smaller uncertainties occur in areas where several arrays provide good azimuthal coverage (coloured events in Fig. S3b), primarily across the middle of the Korean Peninsula, where five infrasound arrays have an average distance between the three closest stations of <200 km (Fig. 1). The edges of the designated study area are outlined by events with large uncertainties (grey events in Fig. S3b) where azimuthal array distribution and the spacing between arrays degrade event locations. This effect was investigated using a simulation technique that perturbed arrival times in the location procedure (Arrowsmith 2018), with results documented in Text S1. The simulation locates many events based on the perturbed detection times in China and across middle and southern Japan, all areas along the edges of the study area. Most events located across the Korean Peninsula are far from these areas, which is consistent with real sources. A seasonal effect is also associated with these false associations to the northwest during the winter, spring and fall and to the southeast during the summer (Fig. S4).

Variations in the number of observed infrasound events as a function of season, weekday and time reflect both the anthropogenic origin of the sources as well as seasonal variations in propagation path efficiencies to individual arrays (Fig. 5b). Seasonal variations are primarily impacted by the direction of stratospheric winds across the Korean Peninsula, with the most detections in the winter from the northwest and in the summer from the southeast (Fig. 2). Summer sources from the southeast include mining blasts and quarry operations in SK. During the winter, event locations are dominant along the west coast of NK because of the southeastward stratospheric winds. There is a strong infrasound hot spot in China to the northwest of the arrays with relatively large uncertainties (>25 000 km²), as shown in Fig. S3(b), since the location is outside the footprint of the arrays. In both summer and winter, there are locations with significant infrasound source densities near the border between NK and SK as well as along the west coast of SK. Strong infrasound event densities are documented in multiple regions, including mining and quarry operations, engineering and industrial activity areas on Monday through Friday (Figs 4c and d), consistent with other studies (Walker et al. 2011; Park et al. 2014). These activities are dominant during the daytime (6 a.m. to 6 p.m.) but not during the nighttime. We observed a few sources that occurred in different locations from 6 p.m. to 9 p.m., but no events from 9 p.m. to 6 a.m. A few sporadic infrasound sources are observed in NK and SK during weekends.

The addition of data from IS45 and IS30 in 2012 enhances azimuthal coverage, especially for IS45, which covers the area around the NK test site at a unique azimuth. Processing with data from the two IMS infrasound arrays (April 2012–March 2014) produced a similar number of total events (1674 without IMS and 1680 with IMS), including new events near the two IMS infrasound arrays. It shows a reduced number of events with large uncertainties (Fig. 6a) and more infrasound events near NK, Russia and Japan (Figs 6b and S5). The inclusion of data from two IMS arrays provides a slightly reduced average uncertainty by ∼5000 km² than those without two IMS arrays for the events summarized in Fig. 6(a). These improvements are further documented by comparing results for the events that have a 95 per cent uncertainty of less than 25 000 km² with and without data from the two IMS arrays for 2 yr (Fig. 6b).

We further investigated the quality of the infrasound events over these 2 yr to evaluate the performance of the detection, association and location framework used for this study. An analyst reviewed all 529 events with a 95 per cent uncertainty of less than 25 000 km² for the events with data from the two IMS arrays (blue circles in Fig. 6b). 80.5 per cent of these events (yellow bar in Fig. 6a) are good quality with coherent signals at each array where detected and associated with the given infrasound group velocity range, mostly located in and near the Peninsula (yellow circles in Fig. 6b). It was also found by analyst review that low SNR signals were detected and associated with producing some false events. The addition of data from the two IMS arrays also produces more edge events with much larger error estimates, which may represent false associations as a result of continuous sources, as shown in Fig. S4.

4.3 Signal review of selected events

The KIC provides a basis to identify special events for additional investigation using signal characteristics and atmospheric modelling in order to quantify the quality of the catalogue, further assess signal association and add to source type identification. An important part of this assessment is the documentation of the specific effects of seasonal winds on the detecting station distribution. We review selected events that use many associated arrays for their location estimates and are accompanied by ground truth information. Events from all six areas highlighted in Fig. 5(a) are included. We also subsequently investigate the number of seismic signals that are associated with infrasound events to better constrain possible source types in the different areas.

4.3.1 Ground truth events

4.3.1.1 North Korean underground nuclear explosions

NK conducted six UNEs in 2006, 2009, 2013, 2016 (January and September) and 2017 at the Punggye-ri Nuclear Test Site (black star in Fig. 5a). Hereafter, these UNEs are abbreviated as NK2006, NK2009, NK2013, NK2016J, NK2016S and NK2017. Infrasound signals are generated from the UNEs in three ways; local infrasound (due to direct seismic ground motion at the infrasound array), epicentral infrasound (by atmospheric pressure changes at the source epicentre) and diffracted infrasound (secondary; by interaction of surface waves propagating away from the source with topographic features). Previous studies (Che et al. 2009b, 2014, 2022; Park et al. 2018a, 2022) assessed the detection of these three types of infrasound from the UNEs. Infrasound detections from the six UNEs depend on explosion size, atmospheric conditions and local environment (Che et al. 2019). Comparisons of ray tracing simulations using the G2S atmosphere specifications (Drob et al. 2003) for the six UNEs that occurred at different times of the year, document strong seasonal atmospheric effects on the predicted infrasound propagation (Park et al. 2022). The data processing used in this study assumes detections are from epicentral infrasound and estimates the event location as described in section 3. We reviewed the detection and location results for the six NK UNEs from the automatic processing and summarized the results in Table S1. The key points are:

• No distinct infrasound signals were associated with NK2006 due to the small size of the explosion, as Che et al. (2009b) and Park et al. (2022) document. Automatic processing identifies infrasound detections at one or more infrasound arrays from all the remaining UNEs.

• In the case of the NK2013, there were detections at KSGAR and IS30 from the automatic processing, while Che et al. (2014) document an additional detection at IS45 based on PMCC processing. The detection missed by AFD was due to a relatively low amplitude signal combined with overlapping high amplitude signals from other sources observed at this array.

• Since the location procedure relies on detections from at least 3 arrays, only NK2016J and NK2017 were automatically located (Fig. 7a). The uncertainty contours for these two events include the test site, while the mean location is biased to the north from the true source location due to wind effects on propagation combined with poor azimuthal station coverage. Park et al. (2018a) illustrate that the location uncertainty can be reduced by correcting the propagation model for stratospheric winds and adding a detection at ULDAR, another infrasound array in the East Sea operated by KIGAM, with a complementary propagation path to the southeast.

• In the case of NK2017, there are multiple infrasound detections associated with both diffracted and epicentral infrasound (Che et al. 2022). To reduce computational time, the automated association uses only first arrivals within a given celerity range, and therefore, in this case, the initial location estimate is based on first arrivals which are from diffracted infrasound that can arrive early since a portion of the travel path is at seismic velocities. The automated processing then ignores the later arrivals associated with the direct epicentral infrasound (Table S1). The location estimates and origin times from the original automatic processing and a new location estimate using only the epicentral infrasound arrivals are compared in Fig. 7(a) and Table S1. This comparison illustrates that the corrected location and origin time are closer to the seismic origin time and location, suggesting that phase identification that separates multiple signals from the same source can improve event location estimates.

4.3.1.2 An accidental chemical explosion

A large, accidental chemical explosion occurred at a Lotte Chemical Plant in the city of Seosan in southwestern SK on 4 March 2020 at approximately 18:00 UTC (green star in Fig. 5a). Seven infrasound arrays at distances up to 1260 km (IS30) detected infrasound signals from the explosion. Detection results are summarized in Table 1. The estimated origin time and location using these infrasound detections (arrival times and backazimuths) is 17:55:19 UTC and 37.0008°N/126.4762°E (Fig. 7b). The ground truth source location based on a newspaper report is 36.9946°N/126.3697°E (see Section 'Data and Resources'). The difference between the plant location and the estimated infrasound location was 9.4 km documenting that with a robust set of automated detections with good azimuthal coverage that high-quality automated locations can be estimated relying on backazimuth estimates.

Negative backazimuth deviations (difference between the observed and true backazimuth) are observed at CHNAR, KMPAR and YPDAR to the northeast of the explosion, while the arrays to the southeast (TJIAR and IS30) have positive backazimuth deviations (Table 1). These differences are consistent with ray tracing simulations using the European Center for Medium-Range Forecasts (ECMWF, see Section 'Data and Resources') by Arrowsmith et al. (2021) due to crosswind deviations of ray paths. These comparisons suggest the possibility of improving event locations while reducing uncertainty by correcting backazimuth biases using an atmospheric model. The assessment of this event illustrates how an automated KIC can be a first step in forensic analysis as well as a tool to motivate extended studies of infrasound propagation using atmospheric models (Blom et al.  2023).

4.3.2 Selected events from infrasound hot spots

Fig. 8 summarizes the NAAs from 3 to 7 that contribute to the association and location of events in each infrasound hot spot area outlined in areas A-D in Fig. 5(a). The NAAs for a given event are a proxy for the confidence that the automatically generated event is real. Most events with more than four NAAs are clustered in specific areas (Fig. 8) where repeated events are suspected to be from open-pit mines, rock quarries or other types of engineering activities. As documented by Park et al. (2014), event type, location, array azimuthal distribution and seasonal wind directions all impact the detection, association and location process. For each focus area, we review a selected event, describe signal characteristics and model propagation to enhance the interpretation. The G2S atmospheric model (zonal and meridional wind velocity profiles, temperature, density and pressure; Drob et al. 2003) at the closest time to each event was used in ray tracing (Blom & Waxler 2012). The waveform data reviewed in this section are available in the supplementary material.

Infrasound signals from 298 events with locations along the west coast and western inland of NK were recorded by the infrasound arrays (Fig. 8a). These events occur primarily in the winter when stratospheric winds are complementary to the array distribution to the southeast (Fig. 5b). Che et al. (2019) document infrasound signals from NK that are associated with industrial mining activities, water channel construction, or other industrial human activities such as public construction in support of reclaimed land development. The sources may have variable frequency content as impulsive explosions can produce higher frequency signals than some delay-fired mining explosions with complex timing intervals between individual explosions resulting in increased signal duration (Heuze & Stump 1999).

We reviewed an event from this area (star in Fig. 8a) with 7 NAAs that occurred on 22 December 2021 as shown in Fig. 9(a). The origin time and location estimate using infrasound arrivals are 06:00:04 UTC and 39.5519°N/126.1390°E. This event is believed to be an explosive event producing an impulsive infrasound signal of short duration, and a clear P-wave arrival observed on the co-located seismometers associated with the infrasound arrays (blue waveforms in Fig. 9a). An improved origin time estimate (06:01:19.5 UTC) uses the seismic arrivals and an assumed constant seismic velocity of 6.5 km s⁻¹ for the P wave (Lee et al. 2020). Table 2 summarizes the seismic arrival times, estimated seismic origin time, infrasound arrival time, azimuth, range and celerity at the seven infrasound arrays. The infrasound celerities (range divided by infrasound travel time) are based on the estimated infrasound location and seismic origin time. The closest three arrays, CHNAR, KSGAR and KMPAR, have relatively faster celerities from 342 to 361 m s⁻¹, corresponding to tropospheric (Iw) phases. While BRDAR, YPDAR and TJIAR have slower celerities from 286 to 309 m/s, corresponding to stratospheric (Is) phases. The relatively high celerity estimate at KMPAR could be impacted by: (1) An inaccurate distance from the estimated infrasound location (not seismic location or known location information) or (2) Propagation conditions different from other arrays (i.e. CHNAR and KSGAR). Note that KMPAR has two infrasound arrivals (Iw and Is) based on the source location and celerity estimates (360 and 304 m s⁻¹, respectively) with only the Iw phase used for the association and location. IS45 recorded multiple signals, possibly a mixture of Iw and Is phases, with a later, fast celerity (319 m s⁻¹). Ray tracing suggests triple-directional stratospheric ducting conditions at the time of the event due to a sudden stratospheric warming (SSW) event (Park et al. 2016; 2018a). Ray tracing predictions well match both Iw and Is phases, except for those from YPDAR and TJIAR, possibly due to the missing fine-scale structure of the atmosphere in the model used (Text S2 and Fig. S6a).

314 infrasound events with large NAAs are located along the west coast of the Korean Peninsula (Fig. 8b). We reviewed an event on 15 September 2017 during the Fall Equinox with 7 NAAs (star in Fig. 8b) from this area and found that there were no associated seismic arrivals recorded at the seismometers colocated at the arrays for this event (Fig. 9b). The detailed information for the associated signals is summarized in Table S2. We observed two short-duration infrasound signals at the three closest arrays, while signals gradually overlap and decay in amplitude due to multipathing at the more distant arrays, out to IS45 array (Fig. 9b). These infrasound observations are not explained by the G2S ray tracing simulation (Text S2 and Fig. S6b), possibly indicating an undocumented change in the atmosphere during the equinox period, similar to observations accompanying NK2017 (Park et al.  2022).

An infrasound hot spot that contains 206 events with NAAs of 3–5 is located between CHNAR and KMPAR (Fig. 8c) at local distances. An event in this area on 24 May 2015 (03:12:14UTC) (Table S2) generated impulsive infrasound signals at all Korean arrays except for KSGAR (Fig. 10a), while no associated seismic arrivals were observed at the colocated seismometers. Infrasound observations are interpreted as direct arrival or tropospheric phases, although ray tracing using the G2S model fails to predict those returns (Fig. 10b). Additional ray tracing was conducted using local weather data from a nearby a wind profiler (WP) operated by the Korea Meteorological Administration and radiosonde (R) data distributed by the University of Wyoming (see Section 'Data and Resources'). The R data are taken at Osan ∼103 km from the source and the WP data from a site close to CHNAR (∼15 km from the source, Fig. 10b). The detailed descriptions of these data are summarized in Text S3 and Fig. S7. The simulations using the R profile and WP data better capture the effects of the troposphere that are applicable to local propagation distances (Fig. 10b).

Surface topography adds additional complexity that needs inclusion in the ray tracing (Blom 2023). The topography used in this additional simulation is from ETOPO1, a global relief model of Earth's surface that integrates land topography and ocean bathymetry (see Section 'Data and Resources'). The spatial resolution is 15 arcsec longitude and latitude. The simulations with topography using WP data and R profile continue to predict arrivals at CHNAR to the north, the three arrays to the west (KMPAR, YPDAR and BRDAR) and TJIAR to the south. TJIAR has relatively low SNR signals, suggesting that this array may be affected by topographic effects. No arrivals are predicted using either the WR or R data when topography is included along the path to KSGAR, consistent with observations. Note that both KSGAR and TJIAR have noise levels similar to the other arrays (Fig. S8). These results suggest that tropospheric propagation can be blocked by mountainous areas consistent with topography between the source and KSGAR (Fig. 10c) and illustrate the importance of including both the fine structure of the troposphere as well as the topography in assessing infrasound propagation at local distances.

220 events are located in Area D (Fig. 5a). One of these events occurred on 22 July 2015 (star in Fig. 8d) associated with 6 NAAs and is subsequently reviewed (Fig. 11a). The four closest arrays, TJIAR, KSGAR, CHNAR and KMPAR, all have simple and clear arrivals, while the farther arrays, BRDAR and YPDAR, have multiple arrivals consistent with multipathing. Ray tracing using the G2S model predicts stratospheric arrivals at the infrasound arrays to the west, while tropospheric arrivals are predicted using local weather data (R and WR) to the northeast (Fig. 11b). YPDAR and BRDAR record stratospheric arrivals with high celerity consistent with predictions, while KSGAR and TJIAR have relatively slower stratospheric arrivals (Fig. 11c). Ray tracing using the local weather data suggests a broad ensonified area for Iw to the east but not to the west where Korean arrays are located, consistent with the observations (detailed description in Text S4 and Fig. S9). Strong infrasound arrivals at TJIAR and KSGAR are not predicted by ray theory but could reflect contributions from gravity waves, effective jet speed or other fine-scale inhomogeneities in the atmosphere (Hedlin & Drob 2014; Chunchuzov & Kulichkov 2019). This example suggests the need for further analysis using other modelling techniques, including full waveform simulations, in order to assess these additional atmospheric effects.

Events with large NAAs near Russia and Japan (Areas E and F in Fig. 5a) have scattered locations due to either inconsistent source locations, unaccounted long-distance propagation effects, or lack of good azimuthal control. We selected the events with a large number of the associated arrays for review in Areas E and F and found that these events in these areas were confirmed as real sources and can be interpreted with propagation model predictions. The results show that the multiple stratospheric arrivals detected at the arrays from our processing are matched with the G2S model predictions during summer and winter (including a SSW period). The detailed information of these events is summarized in Text S5 and Fig. S10.

4.3.3 Seismoacoustic events

Many of the infrasound events in western NK (Area A) and in the inland region of SK (Area D) are suspected to be generated by near-surface mining explosions (Che et al. 2019), as some of these events are reported to be associated with both seismic and infrasound signals. The KIC provides a valuable data set to analyse complementary seismic arrivals that might subsequently be used to document seismoacoustic events. In this section, infrasound events that have accompanying seismic arrivals (seismoacoustic events) are reviewed to illustrate improved source type assessment under the assumption that events generating seismic and infrasound signals must couple energy into both the atmosphere and the solid earth. Events that generate only infrasound might be associated with atmospheric events and/or events near the surface that poorly couple seismic energy, while events that generate both seismic and infrasound can be from shallow-depth earthquakes or more likely buried explosions.

Seismic waveforms from the colocated seismic arrays for the 1637 events in the infrasound hot spots (Areas A–F in Fig. 5) were reviewed by an analyst to associate complementary seismic arrivals generated based on the estimated infrasound origin time. A visual search for seismic arrivals in the 1–5 Hz pass band was conducted over the time range of ±10 min from the estimated infrasound origin time for each event, as shown in Fig. 9(a). Locations of seismoacoustic events in Areas A–F are documented in Fig. S11 with summaries of the total number of seismoacoustic events in each area (percentages of seismoacoustic events relative to total infrasound events and total number of events as a function of the number of associated arrays in the location) in Fig. S12. There are both seismic and infrasound signals from many events in NK (Area A) (118 events, 39.6 per cent of the events) consistent with the assessment that many events in this region are surface explosions. The majority of these seismoacoustic events are located along the northwestern coast and in two clusters in NK, known for either water channel construction or near surface mining (Che et al. 2019). In Area C, near CHNAR and KMPAR, a cluster of seismoacoustic events in SK may also be surface explosions, as documented by Che et al. (2019). There are few seismoacoustic events in Areas B and D (Fig. S11b and d), 7.5 and 5.0 per cent, respectively (Fig. S12). Che et al. (2019) have shown that there are a few near-surface explosions along the western coast of SK (Area B). As a result, these infrasound events may be associated with underwater or atmospheric disturbances that poorly generate seismic signals. These events occur primarily during the daytime and on weekdays, supporting the interpretation of a human cause. Infrasound hot spots in Area C in SK correspond to known locations of open-pit mines, and quarries confirmed by field surveys (Che et al. 2019). Few seismic arrivals were observed from many of these events as well, possibly as the explosions are in open-pit mines with surface/atmospheric source locations that poorly couple energy into seismic arrivals but effectively generate acoustic energy. This lack of seismic observations may also be due to the spatial energy decay of the seismic wavefield below the background noise level at regional distances from the source to the arrays. The event cluster in Fig. S11(d) (around 37°N/129°E) is in a mountainous area where there are no known infrasound sources, such as mines, based on public information. These infrasound events may be due to other types of engineering activities. There are a few sporadic seismoacoustic events located in Russia and Japan (Figs S11e and f), with percentages of seismoacoustic events at 9.0 per cent for both areas (Fig. S12). The source types in these areas are unknown but considered anthropogenic since most occur during weekdays (Fig. 5b). The scattered locations for these seismoacoustic events result from the uncertainty in the infrasound event location, which could be improved with any observed seismic arrivals.

The development of seismoacoustic events from the KIC is based on first producing infrasound events. Che et al. (2019) triggered their assessment with seismic events from a dense seismic network and then associated with infrasound arrivals from at least three arrays, taking a seismic-based approach to event construction. Event origin times in our study have larger errors than this earlier work as ours are based on infrasound observations which can make association of seismic and infrasound observations more difficult. This study, because of its design, may also document more near-surface and/or atmosphere sources as indicated by events in ocean areas or areas without known mining in SK (Figs S11b and d). Che et al. (2019) identify more events in inland areas, capturing mining or quarry operations based first on seismic arrivals.

5 DISCUSSION

The KIC is used to assess events in and around the Korean Peninsula. Signals from selected events are compared with predictions from ray-tracing models to improve the interpretation of infrasound propagation and identification of arrivals under similar but changing atmospheric conditions. The automated KIC is also used to assess the detection capabilities of regional infrasound arrays in Korea. Assessment of the catalogue benefits from extended analysis of special events such as UNEs conducted by NK. Based on ground truth information for these UNEs, the KIC documents detections of infrasound signals (from at least one infrasound array) for all the NK UNEs except the first and smallest explosion in 2006 and includes estimated event locations. These automated results motivated further analyses characterizing these nuclear explosion sources (Che et al. 2009b, 2014, 2022; Park et al. 2018a) including the use of seismic data. Detailed analysis of infrasound arrivals motivated the implementation of an atmospheric model inversion to refine the atmospheric specification for consistency with observed infrasound arrivals from one of the explosions conducted near the Fall Equinox when the atmosphere undergoes a rapid change (Park et al. 2022). These studies illustrate the importance of azimuthal station coverage in assessing locations and their error estimates with the inclusion of two additional infrasound arrays (YAGAR and ULDAR) operated by KIGAM.

This study focused on signals in the 1–5 Hz band to capture events such as mining explosions based on analysis of regional infrasound signals in previous studies (Park et al. 2014, 2018a, 2023). Che et al. (2019) illustrated the benefit of removing repetitive low-frequency signals for detection, including multifrequency signals, based on the results of PMCC analysis. Multifrequency detections using the entire data set, including low (<1 Hz), mid (1–5 Hz) and high (>5 Hz) frequency bands, may provide additional information concerning source characterization. Additional parameters such as array geometry and total number of elements are also found to affect detection, association and location.

The current automatic association algorithm uses only the first arrival from the source to reduce computational time, but the KIC contains multiple infrasound detections for individual events including arrivals associated with both diffracted and epicentral infrasound as documented for NK2017. This result illustrates that when multiple signals are observed from a single source, individual phase identification is critical to producing appropriate source location estimates based on model predictions.

Based on a simulation using the perturbed detection times from the entire collection of KIC events (Arrowsmith et al. 2008a), most infrasound events formed on the Korean Peninsula are believed to be real sources. However, events along the edges of the study area bounding all the arrays (excluded from this study, Fig. S3) may be incorrectly formed from quasi-continuous signals such as dams, gas flares, or other unknown sources. Data from infrasound arrays outside the Korean Peninsula would increase azimuthal coverage, help reduce these false associations and provide better source estimates along the edges of the study area.

The KIC is also used to quantify temporal variations in the atmosphere that can provide a basis for improved event location. Ray tracing using the G2S atmospheric model generally correctly predicts infrasound arrivals when strong stratospheric winds exist during the winter and summer. Some events (as in Section 4.3.2) have detections from multiple arrays that are consistent with model predictions of stratospheric paths during times of SSW events during the winter (Park et al. 2016, 2022) and can be used to explore how the atmosphere changes with time during these events. However, atmospheric propagation predictions often fail during the equinox period (September month) when the atmosphere is transitioning from summer to winter conditions and the main stratospheric wind direction changes. Infrasound arrivals are not well predicted by existing atmospheric models during these time periods of change as illustrated by the detailed analysis of infrasound arrivals from NK2017 (Park et al. 2022). The additional infrasound events identified by this study could be used to quantify better and understand these atmospheric variations.

Local weather data (radiosonde and wind profiler) provide complementary information for interpreting infrasound propagation in the troposphere observed at local distances. Simulations predict tropospheric propagation at short distances using fine-scale atmospheric structures from local wind profiles, while the global models do not have enough detail for adequate prediction. Simulations using local weather data also demonstrate that topographic effects can impact local infrasound propagation. However, local weather data is only available at limited locations, covers a limited range of altitudes (up to 5 and 30 km for wind profiler and radiosonde, respectively), and is made at specific times (every 12 hours for radiosonde) as documented by Park et al. (2022). The application of these models to local infrasound events in the KIC can provide insight into the practical use of local weather data when it exists and illustrates the shortcomings of ray-trace predictions using global models for observations at local distances.

The KIC provides data to support the conclusion that many active infrasound hot spots in SK are associated with open-pit mines, quarries, or other sources (accidental explosion or possibly construction activities) with poor seismic coupling, including events underwater, in the atmosphere or poorly coupled at the Earth's surface. Further work is needed to improve the characterization of the unknown event types with multifrequency analysis, using the KIC, with either additional data or refined wave propagation models. It is also documented that there are few associated seismic arrivals with the infrasound events in some areas, possibly indicative of poor seismic coupling and rapid spatial decay of the seismic waveforms due to the small sources. Therefore, a local dense seismic network (seismometer closer to the sources) may help identify the seismic arrivals from some of these infrasound events and improve the source characterization. A more detailed study of infrasound and seismic signal association is also motivated, possibly complemented by additional data such as overhead imagery.

The KIC includes source locations from several areas with repeating infrasound sources. We found that the infrasound hot spots in NK often generate strong seismic signals, while there were not a significant number of associated seismic signals for infrasound events in other areas identified in this study. The KIC provides a basis to explore seismoacoustic events further.

6 CONCLUSIONS AND RECOMMENDATIONS

The regional KIC is based on the automatic, unreviewed, processing of 24 yr (1999–2022) of data from up to eight regional infrasound arrays that produced a total of 38 455 infrasound events covering December 2004 to December 2022. This KIC is the first published infrasound catalogue compiled in this region that can serve as a valuable data set in developing more robust infrasound source localization and characterization methods. Analysis of the catalogue illustrates that these locations are reliable enough to trigger additional studies focusing on atmospheric wave propagation, source type determination and constraint of time-varying atmospheric conditions covering time spans of tens of years.

Conclusions from this research include:

• Adequate station distribution, including good azimuthal coverage, is necessary to study regional signals and minimize false events.

• Complementary analysis of seismic and infrasound signals can contribute to improved event location and source identification.

• Infrasound-only locations can be improved using multiple arrivals, local/regional signals, backazimuth corrections based on atmospheric models and application of more complex priors in the location procedure.

• The KIC provides an approach to event detection and location during times of atmospheric variations (Equinox versus SSW events) which can be used to study details of atmospheric change.

• Infrasound interpretation can benefit from including local weather observations and topographic effects to constrain propagation over local distances.

This work also highlights potential improvements to the creation of infrasound catalogues.

• Data processing that identifies and uses optimal array geometries or uses an algorithm to select the best element pairs, such as the study by Bishop et al. (2020), might be explored. The optimized array geometry should consider the different noise environments at the arrays in South Korea. As a result of these differences, appropriate noise reduction systems might be further explored.

• A more detailed wave propagation model based on a time-dependent atmosphere could be incorporated into BISL to further decrease the estimated error of the location contours (Marcillo et al. 2014; Blom et al. 2015).

• The current association algorithm might be compared to the implementation of the Global Association algorithm (Brachet et al. 2010) or Network Processing Vertically Integrated Seismic Analysis (NETVISA; Arora et al. 2013) with an infrasound model (Mialle et al. 2019), where additional geophysical parameters are considered to validate the association.

• The interpretation of infrasound propagation at local and regional distances might be further improved by using high-resolution atmospheric models such as the fifth generation ECMWFs reanalysis (ECMWF ERA5), available every hour up to 80 km altitude with a horizontal resolution of 0.25° × 0.25° (see Section 'Data and Resources'). The use of local meteorological data is also motivated where available.

• A hybrid technique that utilizes a combination of infrasound arrays, single seismic stations, or collocated seismometers at infrasound arrays could be explored and used to enhance the infrasound catalogue. Several studies (Walker et al. 2011; Park et al. 2018b) used acoustic signals coupled on seismometers to assess the infrasound signals and investigate infrasound detectability on a regional scale. Since the Korean infrasound arrays are colocated with seismic arrays, this assessment could be extended to complementary seismic arrivals that might subsequently be used for seismoacoustic association using sparse observations and the Probabilistic Global Search location method (Arrowsmith et al. 2020). The combined data set can also be explored using artificial intelligence or machine learning techniques.

• There is a need to develop a simple methodology to make infrasound source size estimates which is a logical next step in the development of a regional catalogue.

The data analysis reported in this paper motivates additional research that has the potential to improve the interpretation of regional catalogues and provide physical insight into problems of interest in wave propagation and source characterization.

• Analyst review can refine the observations and subsequent event analysis, which although time-consuming, can be used to assess selected events triggered by the automated catalogue. Analyst review of events occurring over 2 yr (2012–2014) found that ∼80.5 per cent of the events are confirmed to be real events. All events identified in infrasound hot spots were reviewed for complementary seismic arrivals. Further data review of infrasound signal characteristics can identify and separate multiple arrivals from the same source which can improve source location and interpretation.

• Integration of infrasound data with seismic signals can improve the understanding of large and complex sources such as UNEs. For some source types, the seismic energy is not large enough to be detected at colocated seismic arrays depending on the size of the source, its coupling, propagation distance and noise conditions. Expanded studies of these physical processes are motivated by this initial catalogue. Dense seismic network data that might incorporate single-station infrasound sensor could provide a basis for quantifying the distance gap between regional infrasound arrays and local events.

• The current catalogue illustrates that it can be used to explore events occurring during equinox periods to better understand atmospheric variations in stratospheric winds during this time. Ray tracing using the G2S atmospheric model generally predicts infrasound arrivals when strong stratospheric winds exist, while the predictions fail during the equinox period due to insufficient information about the changing nature of the stratosphere. Additional modelling tools, such as full-wave modelling, might be used to enhance wave propagation interpretations.

• Events with local atmospheric propagation paths provide a basis for incorporating local weather data into wave propagation simulations. In some cases, local infrasound observations are not consistent with the G2S model, but the inclusion of local weather data offers the opportunity to quantify the importance of small-scale variations in wind velocity in the troposphere. We also note topographic effects can be important in assessing infrasound propagation at local distances.

• Infrasound events from the KIC can be used in forensic analysis of damaging accidents such as chemical explosions and can be improved using an atmospheric model-based correction to backazimuth bias (Arrowsmith et al. 2021). In a similar way, even though events have smaller uncertainty estimates, stratospheric winds often lead to deviations in the observed azimuth from the true source azimuth, and thus uncertainty estimates might be improved by correcting azimuth errors associated with predicted stratospheric winds, that can be utilized in the location processing.

DATA AND RESOURCES

Infrasound data, which are acquired by the joint operation of seismoacoustic arrays between the Southern Methodist University (SMU) and Korea Institute of Geoscience and Mineral Resources (KIGAM), are not publicly open currently. Data from seismo-acoustic arrays in South Korea used in Section 4.3.2 is provided in the supplementary materials (“Selected_Events_Data.tar”). Data from the infrasound stations, IS45 and IS30, were obtained from the International Data Center of the Comprehensive Nuclear Test Ban Treaty Organization in Vienna. KIGAM (an affiliation of one of the coauthors of this article) is part of National Data Centres. We used the data wind profilers operated by the Korea Meteorological Administration, which release data on their website (https://data.kma.go.kr/cmmn/main.do) and radiosonde data provided by the University of Wyoming (http://weather.uwyo.edu/upperair/sounding.html). The GeoAc package is from https://github.com/LANL-Seismoacoustics/GeoAc. Current G2S data is being maintained with an experimental prototype system by the National Center for Physical Acoustics at the University of Mississippi and is under active development (https://g2s.ncpa.olemiss.edu). The European Center for Medium Range Forecasts (ECMWF) ERA5 atmospheric model is available at https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-pressure-levels?tab=overview. ETOPO1 data (NOAA - National Centers for Environmental Information 2022) is available at https://www.ngdc.noaa.gov/mgg/global/. A news link describing the accidental explosion in Seosan can be available at https://en.yna.co.kr/view/PYH20200304018700315.

ACKNOWLEDGEMENTS

This research was in part supported by the Basic Research Project of the Korea Institute of Geoscience and Mineral Resources (KIGAM) GP2020-017 funded by the Ministry of Science and Information Communications Technology of Korea. We are grateful for the helpful comments from two anonymous reviewers. Their thorough and insightful reviews led to significant improvements to the paper.

CONFLICTS OF INTEREST

The authors declare no competing interests.

REFERENCES