Introduction

The Comprehensive Nuclear-Test-Ban Treaty Organization (CTBTO) is an international organization that is fully committed to achieving a world free of nuclear testing. It was established in 1996, and its main responsibility is to oversee the proper implementation of the Comprehensive Nuclear-Test-Ban Treaty (CTBT). This treaty has a global objective of banning all nuclear explosions worldwide for both civilian and military purposes. The CTBT was officially adopted by the United Nations General Assembly in 1996 and currently has 196 Member States. Out of these, 187 countries have signed the treaty and 178 countries have ratified it as of May 2024. The participation of Member States from all regions of the world is of utmost importance for the effective implementation and enforcement of the treaty. These Member States are actively working toward the entry into force and complete implementation of the treaty. The CTBTO operates the International Monitoring System (IMS), a network of monitoring stations around the world that detect and monitor seismic, hydroacoustic, infrasound, and radionuclide signals associated with natural and man-made activities. Upon completion, the seismic component of the IMS will consist of 50 Primary Seismic and 120 Auxiliary Seismic stations. The International Data Centre (IDC) is a key component of the CTBTO, which serves as the central hub for collecting, processing, analyzing the IMS data, then distributing the IDC products to the Member States (https://www.ctbto.org).

Network processing is pivotal in seismic analysis, aiming to detect and associate seismic events across stations. However, current methods like Global Association (GA) (Le Bras et al. 1994) have limitations in global event association due to complex wave propagation, signal variations, and distinguishing events from noise. The inaccurate association leads to false alarms (incorrectly associating unrelated seismic events), missed events (failing to link signals from multiple stations), analysis uncertainty, and challenges in monitoring that could result in missed early warnings, delayed emergency responses, and inaccurate hazard assessments. These factors can undermine the reliability of global seismic analysis.

The data, retrieved from the IMS stations, undergo a multistep processing and analysis procedure. Initially, the waveform data are automatically processed through detectors and a set of detections for each station is obtained. The next step is a network processing step, currently the Global Association (GA) system, where all detections are jointly processed, and a series of automatic bulletins is generated. The Standard Event List 3 (SEL3) bulletin is generated as an automatic outcome of the Global Association (GA) process. Following this automated event creation, the dataset is exhaustively examined through an interactive analysis, leading to the generation of the comprehensive International Data Centre (IDC) bulletins (Ali and Shanker 2016; Shanker et al. 2017; Ali et al. 2022), including the Reviewed Event Bulletin (REB). The complete processes, from the reception of raw IMS station data to the final IDC bulletins, are illustrated in Fig. 1. During the process of interactive analysis, analysts undertake a thorough review of the SEL3. This initial review serves as a crucial starting point, allowing analysts to record the real events within the dataset. Following the initial review, analysts further examine additional events flagged by an automated scanner tool to avoid overlooking any potential event. Through automated scanning of extensive data sets, the system detects signals meeting predefined criteria and extracts relevant information about identified events, including their location, magnitude, and arrival times. This automated procedure accelerates data analysis, enabling analysts to promptly identify and explore events without manually reviewing each data point.

Fig. 1
figure 1

Workflow of waveform data processing and analysis pipeline at the International Data Centre (IDC)

Full size image

An important improvement in the analysis process is the integration of the Network Processing Vertically Integrated Seismic Analysis (NET-VISA), a cutting-edge technology for automatic event detection in seismic, hydroacoustic, and infrasound data. NET-VISA uses advanced signal processing and data analysis algorithms to process seismic data, remove noise, enhance signal-to-noise ratio, detect seismic events, and estimate their locations and magnitudes. It aims to improve the efficiency and accuracy of event detection and location by incorporating Bayesian inference principles, which allow for the incorporation of prior knowledge and uncertainties into the estimation process. This approach proves especially valuable in regions with complex seismic activity or limited data availability (Arora et al. 2013; Calò et al. 2018; Calò and Fichtner 2020a, b). Since January 2018, NET-VISA has been actively used alongside the automatic scanner in an operational capacity. It operates concurrently with GA to generate a parallel bulletin called Variational Standard Event List 3 (VSEL3). The VSEL3 bulletin typically provides details such as event detection time, location (latitude, longitude, depth), magnitude, and associated uncertainties.

The primary objective of this research is to evaluate the impact of NET-VISA on the contents of IDC bulletins. This examination entails a comparison of distinct time periods within the bulletins, specifically examining two periods devoid of NET-VISA alongside a single period with NET-VISA integration. Our investigation will primarily focus on internally evaluating the contents of IDC bulletins and how they change under the influence of NET-VISA, rather than conducting comparative analyses with external catalogs. Such an approach is deliberate and strategic, as it allows for a focused and nuanced exploration of the specific impact of NET-VISA within the context of IDC operations.

Waveform data processing and analysis

Waveform data are obtained by monitoring seismic and acoustic waves propagating through the Earth, oceans, and atmosphere. These are recorded by the seismic, hydroacoustic, and infrasound stations of the International Monitoring System (IMS). The International Data Centre (IDC) in Vienna constantly receives waveform data from the International Monitoring System (IMS) stations, which is then subjected to automatic processing (Fig. 1) (Cansi 1995; Bondár and North 1999).

The current approach to network processing employed at the IDC, known as a global association (GA) (Le Bras et al. 1994), utilizes a grid search and various algorithms based on accumulated seismological knowledge to group incoming detection data and then determine the event’s location. The algorithm for determining the event’s location is based on the original iterative linear least squares method of Geiger 1910, and Geiger 1912. Throughout the years, this algorithm has seen significant enhancements, including the application of singular value decomposition to solve matrix equations (Menke 1989; Lay and Wallace 1995) and the integration of azimuth and slowness to refine the location solution (Magotra et al. 1987; Roberts et al. 1989; Bratt and Bache 1988). However, according to Myers et al. (2007), seismic event location fundamentally revolves around minimizing the difference between observed and predicted arrival times. Moreover, all these conventional approaches rely solely on the arrivals associated with them to establish the event’s location. Even multi-event location algorithms, such as those developed by Waldhauser and Ellsworth (2000) and Myers et al. (2007), exclude data from stations that do not detect an event (Arora et al. 2012). The NET-VISA exclusively focuses on network processing and using the IDC’s existing signal detection algorithm, which transforms the raw waveforms into a series of detections.

Station processing

Upon being stored at the International Data Centre (IDC), data from each individual monitoring station undergoes individual analysis to identify significant signals resulting from seismic or acoustic events. This process, referred to as station processing, is fully automated (Fig. 1). Upon detection of a signal, its characteristics are measured and documented in a database. These characteristics encompass time, magnitude, and azimuth, which denote the direction from which the signal was received at a station. Distinct methods of automated processing are employed for seismic, infrasound, and hydroacoustic data.

Network processing

In general, an event, provided it is sufficiently energetic and of interest to the CTBTO, is recorded by multiple IMS monitoring stations. Consequently, the subsequent stage involves deciding which signals from distinct stations stem from the same event. This intricate step, referred to as network processing, is particularly challenging due to the likelihood of numerous events taking place within a day and the presence of numerous detections that cannot be attributed to an event. By systematically scrutinizing all available data and linking them to a specific event, a clearer picture of the actual event gradually emerges, including an approximate location. An accurate location estimate is crucial in determining the magnitude of an event, as the signal intensity typically diminishes with distance at local and regional distances.

Standard event lists

The automatic processing of data yields three consecutive lists of events. The first of these lists, Standard Event List 1 (SEL1), is generated within an hour of the data being recorded at primary seismic and hydroacoustic stations. At this stage, infrasound data are not yet made available for processing as the signals travel slowly through the atmosphere. Based on the events listed in SEL1, waveform segment data from auxiliary seismic stations are requested to provide additional data to refine the location of events already detected and listed. These requests are formulated automatically and dispatched to selected auxiliary seismic stations. With the additional information and incoming infrasound data, a more comprehensive and superior quality Standard Event List 2 (SEL2) is created within four hours of the initial data recording. The final and most refined of these event lists is Standard Event List 3 (SEL3), which incorporates any additional, late-arriving waveform data and is completed within six hours of the first data recording (Fig. 1). The entire process, from data collection to event detection and refinement, is fully automated and carried out by specially designed computer programs.

Interactive analysis

It is crucial for the results of automated processing to be reviewed by analysts to provide Member States with accurate and thorough information. Occasionally, signals may be incorrectly linked to the wrong event, or real events may be missed or mislocated. Therefore, analysts at the IDC cautiously examine every event listed in SEL3. They play a crucial role in the process, reviewing all automatically generated event lists carefully within strict time limits. Their role demands experience, judgment, and sharp detection skills. They discard events that are not real, include signals that have not been linked to an event, and enhance the location estimates of real events. Analysts are also tasked with resolving cases where the automatic processing has combined signals from two events, known as a “mixed event,” or where signals from one event have been automatically interpreted as two events, known as a “split event.” Additionally, they scan the data for any events that may have been completely missed. The scanner tool takes advantage of the specific characteristics of the IMS stations to detect smaller-magnitude events.

Following interactive analysis, the confirmed and corrected events are listed in a daily Late Event Bulletin (LEB) (Fig. 1), prepared for the subsequent stage of processing called automatic waveform event screening. The automatic waveform event screening process involves screening the detected events to determine their characteristics and assess whether they meet predefined criteria for further analysis. This screening helps filter out background noise and non-seismic events, focusing on events of interest, including those potentially related to nuclear explosions.

Methodology

The Global Association (GA) approach applies an event association technique, using spatial and temporal information to associate seismic phases across the network. It initially identifies signals from individual stations that have a non-negligible probability of being the earliest arrival for an event in a particular grid point on the Earth (e.g., Pn and P). Subsequently, detections across the entire network are grouped together with this arrival into events based on various criteria such as consistency in arrival time, azimuth, slowness, amplitude, and phase label. During this stage, phase labels may be adjusted within defined transformation rules. Event locations are then computed by integrating time, azimuth, slowness of associated detections, and the assigned phase labels. Further event attributes, including magnitude, are computed based on amplitude measurements. Finally, detection attributes such as signal-to-noise ratio (SNR) are assessed using straightforward procedures to verify event quality, and afterward, they are determined whether to retain or discard the event. The analysts review and save a subset of events in the REB. This subset includes events that meet the event definition criterion (EDC). The EDC guarantees that the events possess at least three primary stations and a cumulative weight of 4.6 or higher. (The weight of the events is determined by various defining factors such as detection time, slowness, and azimuth at all detecting stations.) All events saved by the analysts, regardless of whether they meet the EDC or not, are incorporated in the Late Event Bulletin (LEB). While effective in many cases, GA struggles with accuracy in associating events, especially under varying geological and seismic conditions. Addressing this, the Network Processing Vertically Integrated Seismic Analysis (NET-VISA) approach, a Bayesian seismic monitoring system designed to process the IMS data (Sereno and Patnaik 1993; Coyne et al. 2009; Russell et al. 2010) to reduce the number of missed and false events in the automatic processing, as introduced by Arora et al. 2013.

NET-VISA, as introduced by Arora et al. 2012 and further developed by Arora and Russell 2012, has been designed to initially tackle the network processing challenge within the field of seismology and then extended it to hydroacoustic and infrasound acoustic detections. Its primary objective is to infer a set of seismic events based on the detections obtained from a network of stations. The overarching approach involves constructing a generative probabilistic model of seismology and acoustics at a global scale. This model is then utilized to infer the set of seismo-acoustic events with the highest posterior probability, considering the observed detections as well as any misdetections.

NET-VISA integrates the entire seismic event detection process into a unified generative model, leveraging comprehensive information about detections, including their timing, directionality, and clarity. The labels for different detection phases can be adapted to associate them with the right event (Arora et al. 2012; Le Bras et al. 2020). NET-VISA uses a generative model, which is made up of several subcomponents that are linked together by the model’s conditional independence assumptions. These subcomponents represent various elements, such as a distribution of potential event locations gathered from the historical seismic activity records, supplemented by a uniform distribution to allow for explosions occurring worldwide. In addition to various parameters, it may include seismic event characteristics such as event magnitude, location, depth, waveforms, arrival times, and other relevant attributes. This probability model is calibrated using historical data and the previous two months of data for the acoustic networks. These data are listed in the daily Late Event Bulletin (LEB). This model is updated on a weekly basis, as described in detail in Arora et al. (2013). NET-VISA calculates the location and confidence ellipse of seismic, hydroacoustic, and infrasound (SHI) events by using its probability density function derived from the Bayesian inference methodology. This approach offers a higher level of accuracy in determining the location and confidence ellipse compared to traditional methods. In practice, it has been observed that the newly developed algorithm for estimating the location and confidence ellipse yields theoretically sound results, surpassing those generated by GA.

Data and analysis

In order to assess the performance of the NET-VISA scanner and its impact on the released bulletins, we divided our study into two specific time periods, each spanning 1200 days. The first period precedes the integration of the NET-VISA scanner, providing a comprehensive view of the seismic data landscape before its integration. The second period starts right after the implementation of the NET-VISA scanner, allowing us to observe and analyze the changes and improvements brought about by the inclusion of the NET-VISA scanner into operations. To exclude any potential influences from various external elements, such as global seismicity and network performance, we also considered a validation period before the first period as period 0, which also includes 1200 days. The evaluated time intervals are indicated in the following manner: period 0, which extends from June 5, 2011, to September 17, 2014; period 1, which includes the period from September 18, 2014, to December 31, 2017; and period 2, which encompasses the timeframe from January 15, 2018, to April 29, 2021. We excluded the initial 15 days of January 2018 because of the extended series of aftershocks that lasted for several days. We intended to ensure that these aftershocks did not impact our findings. Within the specified time intervals, period 2 takes center stage in our study as it marks the initiation of NET-VISA event implementations, as evidenced by the occurrence of VSEL3 events during this period. A summary of period 2 reveals that SEL3 includes 195,941 events ranging in magnitudes from 0 to 6.7, while VSEL3 encompasses a larger dataset of 222,918 events with magnitudes ranging from 0 to 7.1. As previously mentioned, NET-VISA utilizes advanced data analysis algorithms to estimate event locations and magnitudes, differing from GA, which accounts for the magnitude range difference. Additionally, human-produced data (LEB) encompasses 160,470 events within the magnitude range of 0–6.3. It is noteworthy that these events are distributed across varying depths, ranging from 0 to 700 km. This spatial distribution and depth analysis are shown in Figs. 2–7, offering a comprehensive visualization of the event magnitude and depth distribution across the specified range.

Fig. 2
figure 2

Comparison of the number of events for the three evaluated periods. The left panel presents a comparison of the SEL3, rejected, scanned, LEB, and REB events across the three periods. On the other hand, the right panel illustrates the variation of scanned events in relation to the number of SEL3 events (highlighted in yellow), as well as in relation to the number of valid SEL3 events (denoted by rejected events) displayed in gray

Full size image

Results and discussion

When examining the number of the SEL3 events across different time periods (specifically, period 0 spanning 2011–2014, period one covering 2014–2017, and period two from 2018 to 2021), we aim to evaluate the influence of NET-VISA. Additionally, we seek to demonstrate that other potential factors, such as global seismicity and network performance, can be ruled out as significant contributors due to the absence of remarkable increases in seismic activity or significant changes in network performance. For a visual representation of the SEL3 event distribution in these periods, please refer to Supplementary Fig. 1.

In our evaluation of NET-VISA’s impact, we also examined the quantity of LEB events within these three periods. Our findings reveal a 4.6% rise in the count of LEB events subsequent to the implementation of the NET-VISA scanner. This translates to an average increase of 7 events per day, along with a significant upsurge of 17.90% in the number of scanned events. In a more detailed analysis (as shown in the left panel of Fig. 2), a comprehensive comparison of event numbers is presented for the three evaluated periods, encompassing SEL3 (representing automatically generated events), the events rejected during the interactive analysis, and the output bulletins; Late Event Bulletin (LEB) and the Reviewed Event Bulletin (REB).

During the period 2, the count of LEB events exceeds that of the other two periods. Specifically, compared to period 0, the count of SEL3 events remains relatively consistent in periods 0 and 2. It is noted that during period 2, a substantial increase in scanned events is observed. This is related to the operational integration of the NET-VISA scanner, thereby contributing to the significant difference in LEB event count during period 2. The gray bars in Fig. 2 (left panel) illustrate the number of scanned events for each of the three periods, displaying a remarkably larger count for period 2. The ratio of scanned events to the SEL3 events for each period is also presented in Fig. 2 (right panel). Post NET-VISA scanner implementation (indicated by the yellow bars), we observed a 6 and 4% increase in the ratio compared to periods 0 and 1, respectively. For a more in-depth insight into the NET-VISA scanner’s impact, we also focused on the valid SEL3 events, excluding the rejected ones. This closer examination revealed a more significant effect, with a 12% increase for period 0 and a 9% increase for period 1.

In Fig. 3, we present an annual comparison from 2012 to the end of 2021. The decline in the numbers of SEL3 events can be clearly seen by observing the black line above the light blue bars. On the other hand, the black line above the gray bars, which represent scanned events, shows an increase in their numbers starting from 2018. It is worth noting that the blue and green bars, representing LEB and REB events, respectively, do not show the same decreasing trend observed in SEL3 events, which is consistent with the rise in scanned events. This indicates that the counts for both LEB and REB have shown a consistent trend, deviating from the downward path observed in SEL3 counts, which emphasizes the remarkable impact of NET-VISA scanned events on the LEB and REB bulletins.

Fig. 3
figure 3

The annual analysis involves comparing the quantities of SEL3, rejected, scanned, LEB, and REB events across three distinct periods spanning a period from 2011 to 2021. The graphical representation showcases two black lines starting from 2017 (a year preceding the implementation of the NET-VISA scanner) to 2021, one indicating the variation in the number of SEL3 events (upper line), and the other representing the change in scanned events (lower line)

Full size image

In addition to evaluating the impact of the NET-VISA scanner, we also examined the performance of the NET-VISA on the network processing. This analysis involved a comparison of the distributions of the three bulletins, SEL3, VSEL3, and LEB events during period 2 (Fig. 4). The distribution of events in SEL3 (Fig. 4a) appears relatively random, initially creating the impression of a higher event count compared to VSEL3 (Fig. 4b). However, this is somehow misleading. VSEL3 indeed encompasses more events, but their distribution is not random, unlike SEL3. This observation suggests that VSEL3 may offer a higher number of valid events, potentially resulting in an increase in the number of LEB events. The LEB bulletin, which contains refined events after interactive analysis, shows a more clustered spatial distribution (Fig. 4c).

Fig. 4
figure 4

Maps showing the distributions of SEL3 A, VSEL3 B, and LEB C events in the period 2

Full size image

Figure 5a, b illustrates the distribution of events with varying magnitudes across the three bulletins during period 2. It is evident that LEB either contains an equal or a higher number of events with magnitudes below 3.5 mb when compared to SEL3 and VSEL3. Both SEL3 and VSEL3 exhibit a higher frequency of events with magnitudes exceeding 3.5 mb. A significant portion of these events, sometimes more than half, either undergo rejection by analysts or experience magnitude alterations due to changes in location, depth, or the number of recording stations. The number of VSEL3 events slightly exceeds those of SEL3, particularly in moderate-magnitude events falling within the range of 3.5–5 mb (Fig. 5a). In July 2018, there was a noticeable rise in the cumulative moment of SEL3 and VSEL3, which could be connected to specific aftershock sequences. This surge has a clear impact on the LEB. However, the significant increase in the cumulative moment in VSEL3 in August 2018 does not correlate with the shape of LEB, as shown in Fig. 5b.

Fig. 5
figure 5

A Histogram representing the magnitude (IDC mb) distribution of the SEL3, VSEL3, and LEB events in the period 2. SEL3 and VSEL3 events show almost the same distribution and mean value; the LEB has lower mean value. B Graph of cumulative magnitude for the SEL3, VSEL3, and LEB events between 2018 to 2022

Full size image

The distributions of SEL3 and LEB, both pre and post implementation of NET-VISA as a scanning tool in interactive analysis operations, are illustrated in Fig. 6. Figure 6d reveals an increased number of events in specific regions compared to Fig. 6c, suggesting a potential correlation with the additional scanned events in Fig. 2 attributed to NET-VISA. This leads to the inference that NET-VISA generates a larger number of real events, particularly within aseismic regions.

Fig. 6
figure 6

The spatial distributions of the SEL3 and LEB events. A SEL3 events in the period 1, B SEL3 events in the period 2, C LEB events in the period 1, and D LEB events in the period 2

Full size image

NET-VISA operates under the a priori assumption that seismic events are evenly distributed across all depths, presenting a slight drawback in regions with documented deep seismic activity. GA, on the other hand, uses deep grid cells in areas of deep seismicity. In these areas, having prior insights into probable event depths could probably have refined the NET-VISA event localization. Figure 7 outlines a comparative analysis of SEL3, VSEL3, and LEB bulletins, categorized by event’s depth. The outcomes indicate that NET-VISA (presented by VSEL3) exhibits slightly inferior performance for events occurring at depths exceeding 300 km. However, it is important to note that the dataset contains a limited number of deep events, making it challenging to arrive at a definitive conclusion. In addition, the deep events that may be interpreted as shallower by VSEL3 will be reviewed by the analysts, who will assign them a more accurate depth if they are indeed deep. Considering the relative sparsity of the IMS, it can be challenging to accurately determine the depth of events unless the depth phases are identified and utilized. This is critical for the CTBTO’s mission, as incorrectly attributing the source of an event as deep when it is shallow could have more severe consequences for the IDC. Conversely, the reverse scenario is not a significant issue. At depths below 40 km, there is a substantial surge in the count of VSEL3 events in all magnitudes. For magnitudes exceeding 4.0 mb, VSEL3 displays a significant rise in event count within the 100–300 km depth range. Interestingly, this surge does not result in a proportional increase in the count of LEB events at these depths. Conversely, there is a decline in the count of deep LEB events in comparison to SEL3 counts.

Fig. 7
figure 7

Depth Histogram of the SEL3, VSEL3, and LEB events in the period 1 (right corners) and period 2. a all magnitudes in the period 2; b depth histogram for magnitudes above 4 IDC mb in the period 2

Full size image

Conclusion

In this study, we assessed the impact of the Network Processing Vertically Integrated Seismic Analysis (NET-VISA) performance on the International Data Centre (IDC) bulletin production. The results revealed that the integration of the NET-VISA scanner resulted in a noticeable 4.6% increase in the number of events recorded in the Late Event Bulletin (LEB). On average, this equates to approximately seven extra events per day. Additionally, there was a significant 17.90% increase in the count of scanned events, highlighting the positive influence of NET-VISA on the overall process of monitoring seismic events. When comparing the spatial distributions of SEL3 and VSEL3 in relation to LEB, it becomes evident that VSEL3 displays a more clustered distribution, which bears similarities to the distribution of LEB. This indicates that NET-VISA performs better than GA in reproducing the Earth’s seismicity. However, in terms of the depth distribution of events, NET-VISA exhibits relatively weaker performance for events occurring beyond 300 km in depth. Nevertheless, below the 40 km depth threshold, there is a significant increase in the number of VSEL3 events across all magnitudes. Particularly for magnitudes exceeding 4.0 mb, VSEL3 demonstrates a substantial surge in event count within the depth range of 100–300 km. However, this increase does not result in a proportional rise in the count of LEB events at these depths. In conclusion, the scanned events generated by NET-VISA have played a crucial role in the rise of LEB events since January 2018. This indicates that a substantial portion of these events were real, resulting in a higher workload for analysts.