Local earthquake monitoring with a low-cost seismic network: a case study in Nepal

Abstract

Seismic monitoring matters both for research and for populations living in areas of seismic hazard; however, it comes with a cost that is not fully affordable for developing countries. Compared to classical approaches with very quiet sites and high-quality instrumentation, it is therefore worth investigating low-cost seismic networks and how well they perform at detecting and characterizing seismicity. We analyze 1 year of seismic data from an educational seismology network in Nepal, create our own earthquake catalog, and compare it to the publicly available national observatory catalog. We find that despite the noisier seismic station sites, the overall results are comparable and all the main features relevant for seismicity are found. We present quantitative analyses of locations, magnitudes and their frequency distribution in our catalog, as well as differences with the observatory catalog. Differences between the two catalogs primarily stem from the respective network geometries and their coverage, as well as daytime noise level differences. We conclude that if properly planned and installed, low-cost seismic networks are a viable, feasible and significant complement to monitor seismic activity.

Graphical Abstract

1 Introduction

Seismic monitoring is of great significance as it yields valuable information regarding the dynamics of the Earth. For seismologists, acquiring comprehensive details about earthquakes including their location, frequency and magnitude distribution, and the maximum anticipated magnitude for a specific region, is crucial. Data obtained from seismic monitoring allows scientists to assess seismic hazard in a given area, and the authorities and policymakers to evaluate seismic risk, develop risk management strategies, and provide efficient information for education as well as public outreach.

In practice, professional seismic networks monitor seismic activity by installing various types of modern instruments, which are usually costly: on the order of 10 kUSD for the seismometer alone, and more for a complete station with high-quality digitizer and real-time data transmission, without considering costs for establishing a data center and employing observatory staff. Such costs for operating a professional permanent network make it hardly, or not at all affordable for less developed countries, such as Nepal, where no or only very limited internal funding is available for purchasing professional instruments. Besides funding seismic instrumentation, the cost of modern technology to install and operate a professional seismic network in a developing country, especially high-quality telemetry, is a major issue to solve. In addition, the long-term maintenance operation of a seismic network requires experienced seismologists and personnel capable of addressing any potential issues that may arise. These challenges in the past have been typically solved through collaboration and assistance with international partners.

In the context of limited funding, it is worth considering alternatives that may be available at a relatively lower cost. A low-cost seismic network is an appealing option not only for installing new, but also to complement (existing or newly built) professional networks: the densification of a seismic network, or an aftershock deployment could be more easily achieved. Naturally, the lower costs should also reflect easier installation and operation. Whether the effort is worth and the data quality sufficient, it is important to assess how well a low-cost seismic-network performs in monitoring local seismicity, despite not necessarily using more remote and therefore low seismic-noise sites as professional networks do. To examine these questions in a real context, we carry out a study in western Nepal, where a low-cost seismic network was established in 2018.

Seismicity in Nepal is primarily tectonic, and is the result of the deformation of the Himalayan arc, which started to form after the collision of the India plate with Eurasia (e.g., Molnar and Tapponnier 1975). The present-day convergence rate between the two plates is ca. 4 cm/year, half of which is accommodated by shortening on the Main Himalayan Thrust (MHT) at the foot of and below the orogen (e.g., Bilham et al. 1997; Lavé and Avouac 2001; Zheng et al. 2017). Several temporary seismic experiments were deployed in the past decades to investigate the structure and seismic activity of the Himalaya (HIMNT: Schulte-Pelkum et al. 2005; Hi-CLIMB: Hetényi et al. 2006, 2007 and Nábelek et al. 2009; Hi-KNet: Hoste-Colomer et al. 2018 and Subedi et al. 2018; NAMASTE: Karplus et al. 2018), but by design all operated for a few years only. The Himalayan region has a long history of frequent strong earthquakes and has been shaken by large (M7 +) and great (M8 +) earthquakes in the past centuries that have either partially or completely ruptured the north–south extent of the MHT. The most recent major earthquake in Nepal was the 2015 Gorkha earthquake, which was followed by tens of thousands of aftershocks (Adhikari et al. 2015, 2023; Avouac et al. 2015; Bai et al. 2019). Therefore, the seismic hazard in the Himalaya is very high, and is a primary concern for Nepal.

The long-term monitoring of earthquakes in Nepal is carried out by the National Earthquake Monitoring and Research Centre (NEMRC), at the Department of Mines and Geology: they run the Nepal Seismic Network since 1994, currently consisting of 40 broadband permanent seismic stations all over the country. Of these, 13 stations are installed in the area where our low-cost network, the Nepal School Seismology Network (NSSN) has been installed since 2018. In general, the NEMRC has been providing earthquake epicenter and magnitude information to the authorities and the public for every ML 4 + event. Earthquakes of ML 4 + occur frequently in the study region, nevertheless, on many occasions people felt shaking without finding a corresponding earthquake information from NEMRC. This is because earthquakes, especially shallow ones, are felt by the population even if their magnitude is below 4, but the corresponding detections by NEMRC are not made public. This is an additional reason, beyond testing the feasibility and performance of a low-cost approach to earthquake monitoring, for which we here undertake a comparison of observations by the NSSN with those of NEMRC, and in the process create our own “low-cost” catalog. We propose the potential use of low-cost seismometers for the purpose of seismic monitoring, especially in Nepal where obtaining government funding for purchasing further broadband sensors is extremely challenging.

2 Nepal school seismology network

The Nepal School Seismology Network was initiated in 2018 in the framework of the Seismology at School in Nepal program (Subedi et al. 2020a), which was motivated by educational seismology initiatives elsewhere in the world, and ultimately triggered by the devastating 2015 Gorkha earthquake. The NSSN deploys Raspberry Shake sensors, and the pilot station in 2018 was a vertical component (RS1D) instrument. In May 2019, the main deployment phase with 21 stations covered western Nepal (27.5–29.1°N, 82.5–85°E) over an area of ca. 280 km east–west, including Barpak, the epicenter village of the 2015 Gorkha earthquake (Fig. 1). The NSSN has been progressively deployed as a densification of the low-cost instrumentation in the region of interest as well as extensions throughout Nepal, to reach 33 sites in December 2022. The main goal in establishing the NSSN was to initiate earthquake education in schools and to build an educational seismic network that would provide useful data for both education and fundamental research. The network was installed to the west of Nepal's capital, Kathmandu, for the following reasons: it is a region where a major or great earthquake is expected (Dal Zilio et al. 2021; Bilham R., 2019), there were no earthquake education programs at the time (Subedi et al. 2020a), and the installation of sensors was feasible in terms of power supply and Ethernet connection. For further details of the program we refer to Subedi et al. (2020a,b).

Fig. 1

Map of Western Nepal, showing the Nepal School Seismicity Network (NSSN), and the locations of 328 earthquakes recorded during 2021. the black rectangle frames our network area. Earthquake locations inside the network are represented as red circles (n = 273) with size proportional to their respective magnitude. Earthquakes that occurred outside the network and with azimuthal gap > 250° are plotted in gray color. Yellow and magenta stars refer to the mainshock of the 2021 Lamjung earthquake sequence (Koirala et al. 2023) and the 2015 Gorkha earthquake (Adhikari et al. 2015), respectively. NSSN stations are depicted with green inverted triangles, and NEMRC (permanent network) stations in the region are with blue triangles. The main cities are represented by white squares: KTM, POK, JOM, and BTL stand for Kathmandu, Pokhara, Jomsom, and Butwal. Main tectonic features: MFT—Main Frontal Thrust, MBT—Main Boundary Thrust, MCT—Main Central Thrust, STDS—South Tibetan Detachment System; thin red lines represent normal faults. The cross-section line AA’ is for depth profile and is shown on Fig. 9

In practice, the context of the program means that NSSN stations are deployed in school buildings or on school campuses, typically on the ground floor, yet with inherent daytime noise. There is usually no or very little insulation of the sensors, except for a few stations at high altitudes. As an overview, most of the sites are reasonably calm for the context; however, a few are badly affected by daytime noise (e.g., road traffic, urban environment). Nevertheless, they still provide valuable and useful data at night (Subedi et al. 2020a).

The decision to use Raspberry Shake seismometers within the NSSN was determined by comparative testing and the ease of deployment (Subedi et al. 2020a). The Raspberry Shake Network is the largest citizen science network where community scientists, citizen science projects, universities, and individuals observe Earth’s movement and share data openly. The RS1D seismometer, which NSSN also uses for most of the sites, has been used to study different natural and environmental activities in many countries. Anthony et al. (2019) were among the first to use Raspberry Shakes for studying local and regional earthquakes. Since then, Raspeberry Shakes have been widely used for different purposes that include studying rockfalls (Manconi et al. 2018), monitoring debris flows (Chu et al. 2019), monitoring anthropogenic activities in cities (Grunberg and Schlupp 2019; Pulli 2018), studying seismic swarms (Hicks et al. 2019), observing seismic noise reduction due to COVID-19 lockdown across the globe (Lecocq et al. 2020), and educating citizens in general (Walter et al. 2020; Subedi et al. 2020a, 2020b). More recently, the 2021 Haiti earthquake and its aftershock sequence helped scientists test the applicability of RS1D to provide scientifically relevant data (Calais et al. 2022; Paul et al. 2023) following a Raspberry Shake network installed in Haiti (Calais et al. 2019).

During our own tests, we successfully detected an earthquake of ML 1.0 at a distance of 36 km, which confirms that RS1D is suitable for monitoring local events in tranquil environments. Easy installation, low price, time-stamping and real-time data transmission via internet connection could place Raspberry Shake sensors as the ideal candidate for the possible densification of operational seismic monitoring networks. The low cost of Raspberry Shake products comes with limitations compared to traditionally used, more expensive seismometers: the geophone (RS1D, RS3D, and the vertical component of the RS4D) will clip for strong ground motions, and the accelerometer (3 components of the RS4D) allows detecting large earthquakes only. Nevertheless, the RS1D can be expected to provide sufficient good-quality data for monitoring the bulk of background seismicity in Nepal.

3 Stations and data

In this paper, we study earthquakes that took place in central-western Nepal and were recorded by the NSSN. Particularly, starting with continuous waveform data and using modern earthquake location software (SeisComP version 5.0), we present a year-long “low-cost” earthquake catalog for the study area.

In 2021, the NSSN was composed of 31 stations where 2 are RS4D, 1 is RS3D and the rest of all are RS1D short-period seismometers consisting of a 4.5 Hz geophone, a 24-bit digitizer and a Raspberry Pi computer. The geophone includes an electronic extension to lower frequencies; of our sensors of different generations, nineteen sensors have flat velocity response in the 0.8–29 Hz range, nine sensors over 0.7–44 Hz and three sensors over 0.8–23 Hz. The RS sensor records data at 100 samples per second, stores it locally in miniSEED format, and transfers it to the central server in near real time via the Seedlink Protocol. Therefore the status of the stations is known by visiting the Raspberry Shake’s station view webpage (for example, for a station S8618: https://stationview.raspberryshake.org/#/?lat=28.19820&lon=84.38905:zoom=8.000:sta=S8618. In 2021, almost all stations were recording continuously except for four stations that stopped operating for about a period of 4 months (Fig. 2). Two instruments were replaced, leading to 31 operating sites.

Fig. 2

Data availability plot for 2021. The stations’ name is shown on the left y-axis and the data collection rate for each station is on the right y-axis. Stop or start of operation times for particular stations are marked by vertical dashed lines. The top plot shows the operational station count throughout the year. The aggregate data recovery rate for the whole network and considering the time of operation of each sensor is 79.75%. The dotted blue line on 2021.05.18 marks the start of the Lamjung earthquake sequence

Here, we use one full year of continuous data from the Nepal School Seismic Network, for the calendar year 2021. The average data recovery rate for the whole network considering the time of operation of each sensor is 79.75%. All the continuous waveform data used in this study are archived at the Raspberry Shake FDSN Web Service, where they are stored for at least 2 years and are openly available to the public. Three Power Spectral Density (PSD) probability density functions are shown in the Supplementary Material as examples, for a calm and a noisy site in comparison with a very calm site; the median PSD lines for the first NSSN 22 sites are reported in Subedi et al. (2020a).

To assess the affordability of this seismic network, we rely on our experience pertaining to the expenses associated with the instrumentation. A professional seismic station, with a broadband three-component seismometer, high-quality digitizer, real-time data transmission and vault costs ca. 20–25 kUSD, and it can be expected to operate trouble-free for 20 years. On the other hand, a low-cost, one-component seismometer with a minimalistic installation and readily available internet connection costs ca. 0.5 kUSD, and has an expected operational lifespan of 5 years. Therefore, while the richness of data use potential is lower, the low-cost implementation is ca. 10 + times cheaper.

4 Earthquake detection and location method

We employ available open source earthquake location tools, also in the spirit of low-cost approach and open data. Earthquakes are located as follows: first, we use a standard automatic detection and obtain initial earthquake locations using SeisComP (GFZ. 2008) and Locsat locator (Bratt and Nagy 1991). Second, we calculate the final absolute earthquake locations using the Hypo71 locator (Lee and Lahr. 1972) and the local 1D P- and S-wave velocity model developed by Pandey et al. (1985) for Nepal.

4.1 Automatic detection and initial locations

We pre-processed the seismic waveform data with an automatic location procedure using the SeisComP workflow, the same tool that is used by NEMRC. We first filtered the data between 0.7 and 8 Hz using a Butterworth filter of third order. This is almost the same frequency band that NEMRC uses for locating earthquakes (Adhikari et al. 2015), and, more importantly, it is within the bandwidth of the flat frequency response of Raspberry Shake instruments. To obtain an automatic catalog, we run SeisComP in the post-processing mode using three main modules: scautopick, scautoloc, and scevent. Scautopick is an automatic P- and S-phase picker based on an Short Time Average to Long Time Average ratio (STA/LTA) algorithm (Allen 1982), where the picker will be triggered once the average amplitude of the signal exceeds a certain threshold. STA/LTA with respective windows of 2 and 10 s is used in this case. Our automatic data processing for detecting earthquakes is based on a standard STA/LTA P-wave arrival time detection on vertical components only. The secondary picker we used is the Akaike Information Criterion (AIC) which works based on signal-to-noise ratio and refines the final phase picking (Bozdogan 1987). Scautoloc reads automatic picks and associated amplitudes, and then locates events. We use most of the default settings that enforce relaxed conditions on pick association and detection of the vast majority of earthquakes with signal-to-noise ratio sufficient for subsequent manual analysis. For automatic locations, we used LocSat locator (Bratt and Nagy 1991) and the global iasp91 velocity model (Kennett and Engdahl 1991) within the SeisComP workflow. After all, scevent associates an origin with an event or creates a new event if the number of contributing phases is over 5.

4.2 Manual inspection and final absolute earthquake locations

To remove any false detections and wrongly located events in the automatically generated catalog, we manually revised all the automatic phase picks. To improve the precision of the hypocenters we then located the events with the Hypo71 locator (Lee and Lahr 1972) using the local 1D-velocity model available for the region and used at NEMRC (Pandey 1985). For the sake of precision in our manual location, once we locate our catalogue manually, we include S arrivals on vertical components in our manual picks after a careful and lengthy analysis of the waveforms, spectrograms and predicted arrival times. Since we are using S phases from the vertical component, it might cause problems on picking, hence, we reduced the weight of S arrivals to 10% of the P arrivals to minimize the potential effects of S-picking error (see Fig. A1). As most of the located seismicity is relatively shallow (associated with the megathrust at ca. 12 km depth), at the recorded magnitudes there is a clearly discernible signal on the vertical component seismometers (see Fig. A2). This was made with thorough case-by-case analyses and only with data that allowed it. We picked the P- and S-phases assigning an uncertainty in time on the pick. A total of 3373 phase picks (1925 P-phases and 1145 S-phases) were manually determined leading to the locating of 328 earthquakes that occurred in a year. Almost all events have absolute travel time residuals lower than 1 s and a minimum of 5 phases (Fig. A3). The magnitude of the events is evaluated for each event at all available stations by measuring the peak amplitude on the vertical component (as required by MLv) in a time window (ca. 40 s for the closest station and more with increasing distance) around the P arrival, and considering the traditional attenuation function defined for California by Richter (1935). The final network magnitude MLv is the mean of all the available station magnitude contributions, with the outer 12.5 percentiles removed.

5 Results and discussion

5.1 Initial automatic earthquake locations

A total of 911 events are automatically detected by the network between January 1st and December 31st, 2021. However, as it is allowed by our automatic processing configuration, more than half of these events are found to be false events while looking closely at the seismic signals. False arrivals are generated mainly by high seismic noise related to human activities at or/and close to schools, regional events occurring in the west, east, and north of the study area, and teleseismic events. Mainly, earthquakes which occurred in the Tibetan Plateau in the north, in Eastern Nepal and Sikkim in the east, and in Western Nepal and north–west India are detected. On the other hand, at some stations, for example, R6EC4, R8C46, and R51F6, seismic noise is considered as a signal of an event and picks have been incorrectly identified on many occasions. Additionally, noisy seismic sites trigger many event notifications which correspond to human activity signals, and we ignored such detections during the manual re-location procedure.

A significant number of events are located outside of the NSSN network. These are not considered for the manual relocation procedure because we expect the location quality to be low. We define two criteria for analyzing the good quality events: (1) events that occurred within the network area (27.5–29.1°N, 82.5–85°E), and (2) events having an azimuthal gap ≤ 250°. We found a total of 273 events that passed the selection criteria.

5.2 Final absolute earthquake locations

We manually inspected the 911 automatic detections and a total of 273 earthquakes are located in close proximity to the network. A larger portion of detected earthquakes (about 80% of our catalog) is located near the eastern part of our network, which mainly reflects the Lamjung earthquake and its aftershocks starting on 18 May 2021 (Figs. 2, 3), as this sequence was well covered by at least four NSSN stations: in Lamjung (S8618), Gorkha (RA5AB and RD184), and Barpak (S8086). The magnitudes of the manually verified earthquakes vary from MLv 0.8 to 5.3, and 43% of these are smaller than MLv 1.8. The NSSN earthquake catalog in QuakeML format is available in Zenodo repository: https://doi.org/10.5281/zenodo.12626709.

Fig. 3

Earthquake magnitude versus time for the events of the NSSN catalog. The size of each circle refers to the respective magnitude. Solid black, dotted gray and dashed gray lines represent the cumulative number of events versus time for the full catalog, school open-time (9 h-17 h), and school off-time detection (17 h-9 h local time), respectively. The sudden increase of seismicity after 2021–05-18 is due to the Lamjung earthquake sequence. The Lamjung earthquake and the subsequent aftershock sequence is analyzed in detail in Koirala et al. (2023)

The average absolute travel-time residual is 0.36 s, and 82% of hypocenters are associated with less than 0.50 s residual (Fig. A3). These values possibly reflect the timing accuracy of the Raspberry Shake sensor, and the deviations of the true velocity structure compared to the 1-D local velocity model that is available and which we use.

Hypocentral depth calculations using mainly vertical component sensors would usually raise concerns about its quality and uncertainty; however, the observed average error on earthquake depth is 6 km. At the same time, the Hypo71 locator allows us to give a higher weight to the P phases compared to the S phases and consider the respective station elevation by accommodating additional time delay for each station. Although the mainly vertical component seismometers in our network are a limitation in estimating focal depths, our calculated hypocentral depth values using SeisComP workflow fall in a reasonable range: 90% of hypocenters are located at depths 0–15 km, with an average of around 11 km. This matches very well the average depth of the MHT in the region, at 10–15 km depth, where most earthquakes are expected to occur based on longer-term observations by permanent and temporary networks (Duputel et al. 2016).

The location quality and earthquake density are generally based on station distribution and seismic activity within the seismic network. Most events are located in the eastern belt of the network where a relatively large number of earthquakes have been recorded in general (Lamjung sequence) and rather quiet seismic stations are available (see PSD–PDF figure for S8618 in Fig. A4). The seismicity in the western part is less active in the analyzed time period; a similarly high activity could have been detected despite the somewhat sparser station coverage there.

Figure 3 shows the temporal variation of the 273 earthquakes detected by our network, with the Lamjung earthquake sequence in May 2021 being the most prominent feature. We also observe a significant difference in detection rates during the daytime and night-time (Fig. 4). The higher noise level during daytime compared to night-time is well observed at most of the NSSN stations which is demonstrated by the PPSD plots for 6 h (see example Fig. A5) and seismic signals and spectrograms for 2 h (see example Fig. A6) of daytime and night-time data. This diurnal variation of earthquake detections is not surprising, given that our stations are noisy during the school opening time compared to night. More than 80% of earthquakes in our catalog are detected during school-off time (17 h–9 h local time, Fig. 3), and only 18% of events are detected during the school open time (9 h–17 h local time). Schools are open at 10 h–16 h local time, Sunday to Friday; however, students start to arrive at school ca. 9 h and are supposed to stay at school until ca. 17 h local time. This is also the time for other major anthropogenic activities close to our sites, like factory operation time, businesses, and heavy traffic, which increases background noise on our seismic records. In calculating the number of detections (Fig. 3), we have not accounted for the school-free Saturdays and school holidays. In addition, Fig. 3 also shows that the level of the smallest magnitude detections is higher in the Summer due to the monsoon noise compared to the Winter; this phenomenon has been observed earlier in Nepal (Bollinger et al. 2007) and in Bhutan (Diehl et al. 2017).

Fig. 4

Manual P-picks at NSSN stations as a function of distance (hypocentral) and earthquake magnitude, cumulated over the magnitude range, separated into night-time (from 19 h—6 h local time) and daytime. P-wave arrivals from earthquakes of the same magnitude are picked at larger distances at night compared to daytime, demonstrating how data quality improves at night

5.3 Mc and b value estimation

The magnitude–frequency distribution of our earthquake catalog is shown in Fig. 5. We estimate a magnitude of completeness (Mc) at MLv 1.8, using the Maximum Curvature method within the ZMAP software (Wiemer 2001), which also provides a bootstrap-based uncertainty.

Fig. 5

Gutenberg–Richter distribution plot of the cumulative number of events located by NSSN in 2021. Frequency (in left axis) shows the number of earthquakes in each magnitude bin. The magnitude bin used here is 0.1 magnitude units. The number of events is represented by black dots and the cumulative number of events is shown in pink circles. Completeness magnitude is estimated at MLv 1.8 and the estimated b value is 0.71 ± 0.05. The earthquake magnitude and frequency equation is log₁₀(N) = 3.60–0.71 MLv

The slope of the Gutenberg–Richter distribution line, the b value represents the relationship between the relative frequency of earthquakes and their magnitudes, and is a parameter used in seismology for understanding seismicity levels and seismotectonics. The b value should remain a relatively stable feature of the background tectonic seismicity for a given region and it might change only as a consequence of a permanent or transient change in its deformation regime, e.g., an intrusion in the host material or a large event rupture (Amitrano 2012). Conversely, the magnitude of completeness is a result of the detection method. We estimate a b value for the catalog in this study of 0.71, which is close to 0.89, a recently observed b value for Nepal using 5 years of seismic record between 2015 and 2020 (Adhikari et al. 2023).

Previous studies in Nepal have reported various b values ranging between 0.6 and 1.0 (Pandey et al. 1999; Adhikari et al. 2015, 2021; Laporte et al. 2021; Hoste-Colomer et al. 2018; Michailos et al. 2021). The computed b value for Central Nepal is below 1.0, for example, 0.80 (Adhikari et al. 2015), 0.89 (Adhikari et al. 2021), 0.92 (Baillard et al. 2017), 0.93 (Thapa et al. 2018), and 0.75 (Chingtham et al. 2016). On the other hand, a high b value (> 1.0) is observed in central-western Nepal for a seismic swarm (Adhikari et al. 2021). For the 2021 Lamjung earthquake cluster, Koirala et al. (2023) suggested a b value of 0.64, this is the most recent estimate in our study area to which NSSN data also contributed. In western Nepal, relatively low b values (< 0.80) have been reported, in particular 0.77 (Laporte et al. 2021), and 0.75 (Hoste-Colomer et al. 2018). All these values mentioned above are derived using mostly shallow earthquakes. Nevertheless, for intermediate-depth earthquakes, a b value of 0.91 is reported in Central and Eastern Nepal (Michailos et al. 2021).

In summary, our b value finding falls within the range of b values reported for Nepal, even with using only 1 year of data. The relatively low b value in Western Nepal could be an indication of a relatively high regional stress. Similarly, a relatively high b value in Central Nepal could indicate a relatively lower stress level. In addition, looking at the broader Himalayas, the b values derived in other parts of the Himalayas are similar to the Nepal Himalayas. Namely, the estimated b value for Bhutan is 0.80 (Diehl et al. 2017), for Garhwal–Kumaun Himalaya is 0.83 (Singh et al. 2016), for North West Himalaya is 0.62 (Hajra et al. 2022), and further studies of the temporal and spatial variations including uncertainties of the b value are required. A high seismic activity rate (a value > 3.6 for 1 year) is observed in our study area.

5.4 Comparison with the observatory earthquake catalog

The comparison between the NSSN “low-cost” catalog and the NEMRC national observatory catalog is done by matching events by mainly comparing the event origin time and hypocenter location. To this end, NEMRC provided a catalog of earthquakes that happened in the region for the same period of 2021. The methodology used for event detection by the NEMRC is very similar to ours, including the velocity model and the software.

A total of 273 events are detected and located by NSSN during 2021. Out of 273 events, 250 are present in both NSSN and NEMRC catalogs. During the same period, the NEMRC catalog comprised of 505 events in the area (27.5–29.1°N, 82.5–85°E; Fig. 6). Possible reasons that our analysis has not detected all of NEMRC events include: (1) lower sensitivity of the low-cost sensors, (2) event time, we are missing mostly daytime (9 h–17 h local time) events because of high background noises, especially during high noise-level school assembly time (Fig. A5, Fig. A6), (3) offline station close to Lamjung earthquake sequence, especially some of the low-noise sites (e.g., S8086, R2109, and RA5AB).

Fig. 6

NSSN daily detections compared with NEMRC detections

It is encouraging to see that we can locate an additional 23 events in comparison with NEMRC, which is more than 8% in number. These events are located within the network of magnitude ranging from 0.8 to 3.4. This could verify that the lower detection threshold of NSSN compared to NEMRC has also been observed (Fig. 6). However, 232 earthquakes in the NEMRC catalog are missing from our event list (Fig. A8). A possible explanation for this could be that a low-cost seismometer in noisy environments or periods can pick P-wave arrivals only from closer distances and bigger events compared to quiet environments or periods (Fig. 4). These are not detected by our network because of the following possible reasons, (1) they are at the edge of the NSSN area therefore their azimuthal gap is > 250° so we consider them as poor quality events (Fig. A8), (2) they occur well inside the network but during daytime, (3) due to their small magnitude there are no sufficient stations to locate them, (4) the closest NSSN stations are offline. In general, the majority of the events missing one catalog but present in the other represent out-of-network solutions or an insufficient number of picks for one seismic network but are sufficiently covered by the other.

The horizontal distance difference between the NSSN and NEMRC locations is shown in Fig. 7. In comparison, 36%, 67% and 83% of locations are within respectively 5, 10 and 20 km in the two catalogs. Higher than 20 km horizontal differences only occur for MLv < 2.5 earthquakes (Fig. 7b), for events with insufficient SNR to allow using all available stations, and only for events beyond 150° azimuthal gap (Fig. 7c).

Fig. 7

NSSN and NEMRC catalog comparison in terms of horizontal distance between the respective locations of 250 commonly detected earthquakes. a Histogram with 1 km bin size. b Horizontal distance difference as a function of NSSN magnitude. c The same as a function of azimuthal gap

Figure 8 shows the hypocentral depth difference between the NSSN and NEMRC locations. This value is typically more poorly constrained. The NEMRC catalog reflects the pre-selected fixed depth choices of the observatory, at 2, 6, 10, 20, 25 and 30 km depth (Fig. 8a). The NSSN results are as described before, with an average of 11 km depth (Fig. 8a). Note that these values are not directly biased by the interfaces in the 1D velocity model, which are at 23 and 55 km depth, i.e. there is no hypocentral depth clustering in the catalogs around those depths. The hypocentral depth difference plot (Fig. 8b) is satisfactory, with a mainly symmetric shape around 0 km, and two additional local high-value bins related to the NEMRC fixed depth-value choices. The locations from the NEMRC and NSSN catalogs are projected along the N–S profiles (see profile AA’) to see more clearly the relationship between seismicity and the tectonics structures, mainly the Main Himalayan Thrust (MHT). As expected, we observe seismic clusters located along the MHT geometry (Fig. 9).

Fig. 8

NSSN and NEMRC catalog comparison in terms of hypocentral depth between the respective locations of 250 commonly detected earthquakes. a The hypocentral depth histograms of the two catalogs. Note that NEMRC uses fixed depth-values. b The histogram of hypocentral depth differences

Fig. 9

NEMRC (green, fixed depth) and NSSN (cyan) catalogs projected to the depth profiles. The red line is MHT geometry across central Nepal (from Duputel et al. 2016)

The magnitudes of the two earthquake catalogs are compared in Fig. 10. The NEMRC values are as reported in the official catalog. The NSSN values have been calculated using the calibration equation as described in Subedi et al. (2020a), based on data from 8 months at the first 22 stations of the NSSN and against NEMRC values in the year 2019. Therefore, we expect a good match, which is confirmed by a fit of N_NEMRC = 0.96 · N_NSSN + 0.07 (Fig. 10). Overall, for any given earthquake, the MLv magnitude value in the two catalogs may differ by at most 0.9 unit, and mostly at low magnitude values.

Fig. 10

Comparison of MLv magnitude values between the NEMRC and the NSSN catalogs. The regression line, N_NEMRC = 0.96 N_NSSN + 0.07 is plotted in black solid line; the black dashed line refers to equivalence, and the gray dashed lines are plotted at ± 1.0 unit

6 Conclusions

We located 273 earthquakes of magnitudes ranging from MLv 0.8 to 5.3 recorded during January–December 2021 by 31 low-cost, mainly one-component seismometers installed at schools in western Nepal. The findings reproduce well those of the national catalog from permanent stations. Our findings indicate that a low-cost network can detect about half of the earthquakes from the advanced national catalog, and events missing in the low-cost catalog are mostly because of high background noise during the school opening time. We also located 23 events within the limits of our network that were missing from the national catalog, suggesting that contribution from low-cost seismometers might enhance earthquake monitoring in Nepal. High event density on the eastern part of the network reflects the Lamjung earthquake on 18th May 2021 and its aftershocks. A large percentage (90%) of earthquakes are located at shallow depths, between 0 and 15 km; these match very well the average depth of the Main Himalayan Thrust in the region, where most earthquakes are expected based on longer-term observations by permanent and temporary networks as well as the geodynamic context. Irrespective of the high noise level at most stations during the daytime, the magnitude of completeness of the NSSN is found at MLv 1.8, which is almost the same as the estimated magnitude of completeness MLv 1.7 by Koirala et al. (2023) for the Lamjung earthquake sequence recorded with the permanent and temporary seismic networks. More than 80% of the identified events are detected during the school off-time, suggesting that low-cost seismic network stations are suitable for seismic monitoring if the site has low background noise. The estimated b value of our catalog is found 0.71 which is close to the b value recently observed by NEMRC for Nepal. When quantitatively comparing the location and magnitude of our earthquakes to that of the observatory catalog, we obtain very similar results in terms of detection completeness, location accuracy and magnitude determination. With this work we demonstrated the feasibility of employing low-cost seismic networks (like Raspberry Shake) for local seismicity monitoring in developing countries where institutional budgets are very tight. The cost–benefit of low-cost deployments compared to high-end professional solutions is about 10 times more advantageous, which highlights the importance and usefulness of (complementary) employment of low-cost instrumentation in seismic monitoring.