Bangladesh, one of the most densely populated countries in the world, lies on the eastern side of the India-Asia collision zone, where the Indian plate subducts obliquely beneath the Burma microplate (Steckler et al., 2016; Bürgi et al., 2021). The subducting plate interface may be capable of producing great earthquakes, posing a considerable hazard to the high population (Akhter, 2010; Steckler et al., 2016). The Bengal Basin sits on the downgoing plate and is characterised by extremely thick Cenozoic sediments (up to 20 km thick, Johnson and Nur Alam, 1991; Singh et al., 2016). Bangladesh also hosts the world’s largest delta, fed by great rivers that transport sediment from the Himalaya to the Bay of Bengal. The annual monsoon, together with frequent cyclones, carries massive rainfall to Bangladesh, causing damaging flooding (Hossain et al., 2020). This dynamic setting, with tectonic changes below and climatic changes above, results in high exposure to Earth hazards in Bangladesh.

In 2016, the TREMBLE (Temporary REceivers for Monitoring BangLadesh Earthquakes) network, consisting of 28 seismometers, was deployed in north-eastern and south-eastern Bangladesh. Although low-noise sites were chosen for sensor deployment, background noise is always present. Background noise affects the ability of the network to detect earthquakes and to be used for imaging studies, for example, constructing unbiased noise correlation functions (Tsai, 2009; Ayala-Garcia et al., 2021). Therefore, evaluating the background noise recorded by the network is a crucial step in assessing its ability to detect solid Earth signals. Understanding the sources of seismic noise also improves our knowledge of physical processes, for example, ocean wave dynamics and river transport. TREMBLE presents an opportunity to better characterise seismic noise in a little-studied tropical environment, characterised by extreme monsoonal weather patterns, large tropical storms, extremely thick sediments, and located next to a marginal sea.

Our aim is to characterise the origins and spatiotemporal variations of seismic noise in Bangladesh. We first present an overview of our seismic network and use ambient noise to characterise the sites in terms of resonance frequency and absolute noise levels. We then analyse the frequency, as well as temporal and spatial variations, of seismic noise from low (∼0.02 Hz) to high (15 Hz) frequencies, using a combination of probabilistic power spectral density calculations and spectrograms. We find strong variations in noise between monsoon and dry seasons across a range of frequencies. Distinctive primary, secondary and local microseisms are present in the wavefield, with differing seasonal changes in power and horizontal-vertical ratios, that we correlate with changing environmental conditions. We further investigate the impact of rivers and human movement on higher frequencies (∼>3 Hz). Our results have implications for seismic deployments in tropical environments, microseism generation, seismic hazards, and earthquake detection.

We characterise the site response for each station by using ambient noise to calculate the ratio of horizontal to vertical spectra (Nakamura, 1989; Nakamura, 2019). A peak in the horizontal to vertical spectral ratio (HVSR) corresponds to the resonance frequency and is related to the depth of strong impedance contrasts (such as the sediment–bedrock interface). HVSR peaks, often used as input for ground motion models (e.g., Zhao et al., 2006), result from a combination of S-wave resonance within the sedimentary section (Nakamura, 1989) and surface wave horizontal polarisation (Bonnefoy-Claudet et al., 2008).

HVSR is computed for each station using 2 weeks of data (in the dry season). Amplitude spectra are computed for each component in 50 s time segments using Welch’s averaged, modified periodogram method filtered with a Hamming window (Welch, 1967). We use a 50 s window to capture signals down to ∼0.2 Hz (10 cycles), to capture our lowest expected resonance frequency. For each time window, horizontal components are averaged using a geometric mean, and divided by the vertical component spectra. The average HVSR is calculated as the mean for all segments. Using over 24,000 segments, transient noise, such as earthquakes, have minimal influence. Results are validated using Geopsy software (Wathelet et al., 2020).

Background noise is analysed using probabilistic power spectral densities (PPSDs). The power spectral density (PSD) for each station and component is first computed from the discrete Fourier transform. PSDs represent the proportion of the total signal power contributed by each frequency component of a signal and are computed in units of decibels (dB) with respect to acceleration per frequency (m/s²)²/Hz. PSDs are then analysed statistically to produce probability density functions, following the method of McNamara and Buland (2004) implemented in Obspy (Beyreuther et al., 2010). We remove the instrument response and calculate PPSD’s in 30-min windows with 50% overlap. PPSDs are smoothed in period bins of half an octave. To investigate seasonal changes in noise, we extract the mean from PPSDs computed per day. To analyse specific frequencies, the required centre frequency is extracted from each daily mean PPSD function. Note that the amplitude at the centre frequency is an average over the period bin (Supplementary Table S2.) Variations in the horizontal to vertical ratio at specific frequencies are calculated from the mean daily power at each component. Horizontal components are averaged using a geometric mean. For long-term HVSR curves at specific frequencies, we remove outliers from the time series, defined as more than three MAD (median absolute deviations) away from the local median within a sliding window (usually 10 days).

We compute spectrograms to look at noise variations over short time windows. Spectrograms display power at different frequencies over time and are useful to identify signals from different sources [e.g., Lythgoe et al. (2021)]. Spectrograms are computed by calculating the short-term Fourier transform for overlapping time windows. We choose a time window of 600 s, balancing the need to account for the longest period signals of interest (∼100 s), while maintaining adequate time resolution.

Data recovery at TREMBLE stations was highly variable, even with regular station maintenance. We had the least success obtaining data from outdoor stations, with station performance principally depending on environmental factors. We have documented recommended best practices for station installation and maintenance in similar tropical areas based on our experience in Supplementary Table S3. All stations have ambient noise levels below the Peterson New Noise Model (Peterson, 1993), although there are considerable variations with frequency, location, and season. Table 1 provides an overview of the frequencies of the microseisms and other signals of interest in the region.

Analysis of horizontal to vertical spectral ratios indicate that TREMBLE sites in the north are class IV (soft soil), whereas in the south they are generally class II (hard soil), which matches the known geology of Bangladesh. Site resonance frequencies vary between approximately 0.9 and 4 Hz, although many stations did not have a clear HVSR peak, indicating that they are located above a thick sedimentary column with little impedance contrast. We find high HVSR amplitudes (>5) at all stations with a measurable peak, with correspondingly high fragility indices. The identification of high fragility indices serves as a clear indicator of an elevated risk of liquefaction and severe shaking in an earthquake. Sites in the north on soft soil coincide with the highest seismic hazard in the region in terms of expected ground motion (BNBC, 2020). Future studies should be directed towards determining site classification and expected ground motion in the very densely populated areas of northern Bangladesh.

In the microseism frequency band, stations from the coast to far inland displayed two peaks, at around 0.08 and 0.25 Hz, corresponding to the primary and secondary microseism, respectively. These frequencies are higher than the global average of around 0.07 and 0.14 Hz, respectively (e.g., Stutzmann et al., 2000), likely due to shorter ocean wave wavelengths caused by the restricted fetch in the marginal sea of the Bay of Bengal.

A local microseism between 0.4–0.8 Hz is generated by wave interactions near the coast in the Bay of Bengal, with its amplitude strongest at stations near the coast. There is a positive correlation between local microseism power and coastal wind when comparing daily signals over a multi-year time scale (Figure 8). This is true for coastal and inland stations (Figure 8). However, no strong short-term relationship is evident when directly comparing raw seismic data and coastal wind speed data (Figure 9). This differs from Zhang et al. (2009), who found a strong correlation between local microseism amplitude and near-coast wind speeds, recorded by a buoy located near shore in California. The lack of strong short-term correlations in Bangladesh is likely because we do not have good observations of offshore oceanic windspeed: the weather station is located on land, and the ocean model does not have sufficient resolution.

The horizontal-vertical spectral ratio of the local microseism is approximately one, indicating a dominance of Rayleigh waves in the local microseism wavefield. Examining the local microseism in Australia (0.325–0.725 Hz), Gal et al. (2015) find that Rayleigh waves also dominate, although Love waves increasingly contribute at higher frequencies, possibly due to the conversion of Rayleigh waves at irregular coastlines, as also argued by Koper et al. (2010). Further analysis is needed to show if Love waves make up a small, possibly frequency dependent, component of the local microseism in Bangladesh.

Additionally, we observe a modulation of the locally generated seismic wave energy with ocean tides, with the microseism peak preceding high tide by ∼1 h. Our results are consistent with Becker et al. (2020), who find a local microseism (at 0.2–1 Hz) in the southern North Sea, with a peak in amplitude that precedes the water level maximum recorded at a nearby tide gauge. Wide shallow waters in both the Bay of Bengal and the North Sea are expected to produce larger amplitude and shorter wavelength tides than in the deep ocean, which may explain why there is a clear correlation between a local microseism and ocean tides in these locations. Becker et al. (2020) interpret the observed local microseism as generated in shallow water at near coastal areas, due to the interaction of incoming wave activity and tidal currents. This is consistent with our observation that high coastal wind speeds are needed to generate the local microseism and that the microseism power is highest on a rising tide, indicating that large wave heights in shallow, tidal areas are necessary to produce the local microseism. The observed time lags between microseism and water level peaks could be due to the microseism being generated in a different location to where the tide gauge is located and/or the dominant influence being when the tide is at the fastest velocity, rather than peak water level. Further work to precisely locate the microseism using array beamforming, as well as ideally using a tidal model rather than point measurement from a tide gauge, is needed to better understand its generation mechanism.

The power of the secondary microseism in Bangladesh is strongly correlated with mean significant wave height in the Bay of Bengal (Figure 10). The power changes with seasons, with up to 20 dB difference between monsoon and dry seasons recorded by coastal stations, and ∼5 dB seasonal difference recorded by stations >200 km inland. As wave height increases in a tropical depression, the secondary microseism power increases, and the frequency decreases. This is because it takes bigger storms (with greater wind speed, length, and duration of fetch) to generate ocean waves with longer periods (Bromirski et al., 1999), and may partially be due to the movement of the source areas of the secondary microseism to shallow water depths (Kedar et al., 2008). There is a lag of 1 day between the secondary microseism power and significant wave height (Figure 10C). A delay in seismic microseism amplitude with respect to significant wave height has been attributed to complex interactions between the incoming ocean swell with coastal reflections (Bromirski et al., 1999; Ardhuin et al., 2012).

There is a seasonal dependence of the horizontal-vertical ratio (between ∼1–4) at station BB02, whereas the horizontal-vertical ratio at station BB03 is nearly constant over time at 0.6–1. Generally, the secondary microseism is dominantly composed of Rayleigh waves (Stutzmann et al., 2009) due to its generation mechanism (Longuet-Higgins, 1950). However, in some cases the secondary microseism has strong horizontally polarised energy due to Love waves (e.g., Juretzek and Hadziioannou, (2016); Xiao et al., 2018), which are thought to be generated by steep bathymetry and subsurface geological boundaries (such as sedimentary basins), or through conversion from Rayleigh waves at heterogeneities along the microseism path (Gualtieri et al., 2020; Le Pape et al., 2021). BB02 shows increased horizontal energy in the monsoon period when wave height is highest, indicating a greater component of Love wave energy at this time. This may indicate that the location of the secondary microseism generating source changes by season, possibly due to tropical storms, which are frequent in the Bay of Bengal during the monsoon. Different Rayleigh-to-Love wave ratios between stations has been documented in several studies (Nishida et al., 2008; Juretzek and Hadziioannou, 2016), and a weak seasonal variation for a coastal region of California is reported by Tanimoto et al. (2016). BB02 is close to the coast, similar to the station analysed by Tanimoto et al. (2016), who suggested that the propagation distance at a near-coast station may be too short to create a diffuse wavefield at these long wavelengths.

The primary microseism shows positive horizontal to vertical spectral ratios, suggesting a significant non-Rayleigh wave contribution to the primary microseism in Bangladesh. The primary microseism has been found to contain dominant horizontal motion in other regions, due to horizontally polarised Love waves [e.g., Nishida et al. (2008); Juretzek and Hadziioannou, (2016)]. Unlike the secondary microseism, Love waves are generated directly in the primary microseism, likely from shear traction due to the shoaling of waves on inclined seafloor topography (Saito, 2010; Ardhuin, 2018). While the primary microseism power displays no strong seasonal changes, the horizontal to vertical ratio does, with higher horizontal energy in the dry season. The ratio of Love to Rayleigh waves in the primary microseism depends on the local source topography as well as the azimuth of the station with respect to the source, since Rayleigh waves and Love waves radiate differently (Ardhuin et al., 2015; Juretzek and Hadziioannou, 2016). Therefore, a changing horizontal to vertical spectral ratio could indicate a changing source location of the primary microseism with different seasons. Temporal trends in the horizontal-vertical ratio for the primary microseism are different to that of the secondary microseism–for example, BB03 displays no temporal variation for horizontal-vertical ratio in the secondary microseism, but a clear temporal variation for the primary. Further work is necessary to ascertain the source locations of the primary and secondary microseisms and how they change over time.

Large amplitude horizontal noise with a 24-hour periodicity is prevalent on all of our broadband stations at frequencies less than 0.05 Hz. A diurnal pattern in horizontal noise power at frequencies less than 0.05 Hz has previously been documented (Beauduin et al., 1996; Wilson et al., 2002; Koper and Hawley, 2010; De Angelis and Bodin, 2012; Wolin et al., 2015). Wilson et al. (2002) find no spatial correlation in long-period noise between regional stations, and conclude that the noise source must be local, from thermal or atmospheric driven ground tilt. Beauduin et al. (1996) and Wolin et al. (2015) utilise a co-located barometer to correlate the variation in long-period noise with atmospheric pressure. They conclude that this noise is likely produced by the elastic response of the Earth to atmospheric pressure fluctuations and wind, as proposed by Sorrells (1971). Atmospheric tides modulate surface pressure and wind speed. In particular, a large-scale atmospheric convective cell produces a 24-h periodic cycle (Dai and Deser, 1999). In continental areas, wind speed tends to peak in the afternoon, influenced by heating from the ground surface, and drops at night (Dai and Deser, 1999). This is similar to the pattern recorded in Bangladesh: for example, weather stations Kutubdia (near BB03) and Srimangal (near BB04) show that temperature and wind speed have a 24-hour cyclicity with peaks in the afternoon and low or no wind recorded at night (Figure 12). Chittagong weather station is on the coast and is likely influenced by coastal winds, so does not show a clear 24-hour cyclicity in wind speed. Peaks in horizontal seismic power for BB02 and BB03 occur during high winds (Figure 12), supporting the generation of this noise by instrument tilt due to the elastic response of the Earth to atmospheric pressure fluctuations. However, high seismic power for BB04 occurs during the night, which is not correlated with high wind at the nearest weather station (Figure 12C). BB04 also has extremely large power on the east-west component, ten times larger than the north-south component; this is likely an instrumental issue, since it first appears in June 2017 and then persists (Supplementary Figure S12). The station is located 14 km from the weather station, and on top of an exposed hill, so may experience local weather patterns–without a co-located barometer it is not clear why BB04 shows a different pattern than the other stations.

Wolin et al. (2015) note that the amplitude of the power change varies depending on the time of year and the soil conditions at the station: stations installed in fine silt material show the largest diurnal noise change, strongest in the summer months when the ground is dry (May-August in their location). Our stations in Bangladesh may therefore be particularly susceptible to atmospheric pressure variations, since they are often installed in shallow ground made of loose sediment. Stations installed indoors also record this diurnal noise change, although these indoor installations are not sealed from the outside and are often placed directly on the ground (e.g., Supplementary Figure S1). Installing seismometers at depth or in a vault can reduce long-period noise from atmospheric perturbations, but even in these deployments the noise cannot be completely removed [e.g., Wolin et al. (2015)], while budget and infrastructure constraints make boreholes unsuitable for many regional networks like TREMBLE. We do not see any seasonal variation in mean daily noise at < 0.05 Hz at the broadband sites in Bangladesh (Supplementary Figure S12).

Station SP02, which is located within 40 m of the major Karnaphuli River, records significant seasonal changes in seismic noise at ∼ 5 Hz. High seismic power occurs ∼8 days after high rainfall. The delay between local rainfall and seismic power is explained by the wide drainage area of the major Karnaphuli River, since river discharge is controlled by run-off from a wide area including the Indo-Burmese Mountains. Future stations located near large rivers in Bangladesh will also record high noise levels during the rainy season at frequencies up to ∼5 Hz, and this should be a consideration for future deployments.

Seismic noise due to anthropogenic activities is site specific. The seismic noise changes recorded by TREMBLE during the COVID pandemic demonstrate that lockdown measures were largely followed in most parts of Bangladesh, although a few stations were unaffected by lockdown measures. Roy et al. (2021) investigated pandemic lockdown noise changes at sites in remote areas of Gujarat, India, also finding negligible changes at some sites. Proximity to police stations appears to be one factor controlling the amount of noise reduction, with stations close to police stations showing large reductions (e.g., BB03), whereas those in tribal areas with little police influence show smaller noise reduction (e.g., SP07). We find that stations with the least reduction in seismic noise are located in highly populated urban locations (e.g., SP19), with more significant noise reduction in less populated areas (e.g., BB05); this may be due to the high population density in Bangladesh’s cities, where physical lockdowns may not be practically possible. Seismic noise is also affected by national holidays and religious events, particularly Eid-al-Fitr and Buddha Purnima. Many stations record a noise decrease during these holidays, including stations SP01, BB03, and BB04, located at a university, boarding school, and tea factory, respectively, while other stations report a noise increase, such as BB05, located at a rural resort. These patterns were largely reversed during the 2020 lockdown, reflecting that people could not leave their residence in the 2020 holiday period and festivals were instead celebrated locally.

Our study yields valuable insights into the multifaceted realm of seismic and environmental conditions in Bangladesh, with implications for research across diverse fields, including seismic deployments, seismic hazard, and imaging studies. Seasonal changes in background noise may manifest as seasonal changes in the amount and lower magnitude bound of detected seismicity. It has been observed, notably in Nepal and India, that earthquakes are more frequent in the dry season than the monsoon season, possibly due to loading from the summer monsoon’s rain supressing tremors (Bollinger et al., 2007). However, careful analysis of seismic noise and seismicity changes is required to fully account for changing noise on seismicity detection. Spatiotemporal noise variations also impact imaging studies. Seismic interferometry methods commonly assume a diffuse wavefield, which we have shown is not the case in the microseism frequency band in Bangladesh. Interferometry studies should particularly note the presence of high amplitude noise from the coast in the local microseism band which may bias resulting Green’s functions.

Further investigations of microseisms recorded in Bangladesh has the potential to improve our understanding of the generation of microseisms in a marginal sea setting like the Bay of Bengal. Additionally, the high number of tropical storms that occur in the Bay of Bengal and have been recorded by TREMBLE (∼35 from 2017–2021 including the strongest super cyclonic storms), present the opportunity to better understand storm generated microseism, and how microseism strength and frequency is affected by cyclone speed, water depth, coastal geometry, bathymetry, and crustal structures [e.g., Lin et al. (2017); Retailleau and Gualtieri, (2019); Park and Hong, (2020)]. Our analysis of one tropical storm raises several topics for further investigation, including why the frequency of the secondary microseism decreases as the tropical depression moves landwards, and why the primary microseism does not change in power or frequency as the depression strengthens. To improve our understanding of the microseisms identified here, future work to constrain the exact source locations is needed Further investigations into the composition of the wavefield would also give greater insight into the generation of microseisms. In particular, the changing horizontal-vertical ratio of the primary and secondary microseism might indicate a seasonal change in the dominant source generation location.