Seismic data represents vibrations of the Earth's surface, commonly referred to as seismic noise. These low-amplitude movements are recorded across the Earth's surface and are traditionally used to study the Earth's internal structure and detect earthquakes. Recently, it has increasingly found applications in meteorology and hydrology, particularly for monitoring weather events (e.g. Diaz et al., 2023; Hua et al., 2023), destructive flood episodes (Burtin et al., 2016), ocean storms, and tropical cyclones (Gualtieri et al., 2018). Seismic noise can reveal the impact of atmospheric and oceanic conditions, providing valuable insights into weather events and climate changes (e.g. Bromirski and Duennebier, 2002; Aster et al., 2008, 2023). In particular, seismic data helps track variations in the Earth's surface caused by factors such as ocean waves, wind, and precipitation, offering a unique perspective on these phenomena (e.g. Grevemeyer et al., 2000; Borzì et al., 2022).

When the seismic noise is analyzed in the frequency domain, two clear peaks emerge in the spectrum (Fig. 3), reflecting distinct types of ocean wave interactions (Koper and Burlacu, 2015; Ardhuin et al., 2019; Tanimoto and Anderson, 2023). The primary peak, observed in the range of 10–20 s (0.05–1 Hz), is generated by the impact of “swell” waves traveling in the same direction, inducing pressure variations in the Earth's crust that match the period of the waves. The secondary peak, in the range of 5–10 s (0.1–0.5 Hz), is produced by wind-driven waves, which propagate in different directions and generate pressure oscillations on the ocean floor (Ebeling, 2012). These seismic signals directly link ocean conditions with seismic activity (Li et al., 2020), providing insights into large-scale weather phenomena like ocean storms.

Higher frequencies above 30 Hz are associated with the effects of precipitation and wind, as seen in studies like Rindraharisaona et al. (2022) or Diaz et al. (2023). These higher-frequency seismic signals help track more localized weather events, such as storms and heavy rainfall. Seismic data, when integrated with other meteorological tools, enhances the ability to monitor and predict weather events.

To analyse seismic data, the raw traces are first corrected for instrument response and converted to units of velocity. These are then filtered with bandpass butterworth filters adapted to capture the target signal: low pass filtering (<1 Hz) for wave-seafloor coupled interactions and high pass filtering (>30 Hz) to identify possible signatures of precipitation, essentially induced pressure fluctuations in the ground converted to weak seismic vibrations due to rain drops. Spectrograms of these filtered seismic traces were computed using short-time Fourier transforms implemented in the scipy.signal package, with the default 256-sample window length used for each segment, to visualise signatures of the hydro-meteorological phenomena in the frequency content of ground vibrations.

Potential environmental signals in the seismic data were also investigated using power spectral density (PSD) analysis. To account for variations over time, a Probabilistic Power Spectral Density (PPSD) method was applied. The continuous waveform was divided into 1 h time windows with 50 % overlap, and a PSD was computed for each window after instrument-response correction and basic preprocessing. These estimates were combined into a probability distribution, providing a statistical overview of typical and transient noise levels across frequencies. The PPSD was produced using ObsPy (Beyreuther et al., 2010), which handles data gaps and ensures reliable normalization.

Temporal variations in PSD amplitudes are also analyzed to track changes in seismic noise at specific frequencies. By extracting PSD values at selected frequencies that are expected to capture primary and secondary microseisms, time series of noise levels are generated. These temporal PSDs allow for the identification of trends and correlations with environmental factors, such as ocean wave activity or weather conditions.

Infrasound waves are low-frequency acoustic waves that are inaudible to the human ear, typically below 20 Hz. These waves are generated by a variety of natural and anthropogenic sources, including meteorological events, volcanic eruptions, earthquakes, and human activities such as explosions and industrial processes (Campus and Christie, 2009; Bondár et al., 2022). In particular, infrasound is often associated with phenomena like thunderstorms, ocean waves, and large-scale atmospheric events, which generate pressure fluctuations that propagate through the atmosphere (e.g. Stopa et al., 2012; Landès et al., 2012; Listowski et al., 2022). These waves provide valuable information about the dynamics of weather systems (e.g. Hupe et al., 2019), making them an essential tool for monitoring and understanding environmental processes (e.g. Brachet et al., 2009; Hupe et al., 2022). Infrasound associated with thunderstorms, primarily generated by acoustic waves from thunder, has been studied previously and shown to be detectable at distances ranging from tens to hundreds of kilometers (e.g., Assink et al., 2008; Šindelářová et al., 2015, 2021). Nevertheless, infrasound arrays detect signals from multiple storm-related sources, not just thunder (e.g., Waxler et al., 2024). In the present study, we build on this understanding by integrating these signals with seismic, satellite, meteorological, and water vapor observations to investigate what these complementary datasets reveal about storm evolution in a coastal environment.

For the monitoring of infrasound signals, we use data from an infrasound array system located at Eforie Nord-Agigea, Romania (AGIR, Fig. 2). This array consists of multiple sensors, including SIS-1 infrasonic sensors (Seismowave), equipped with global positioning systems (GPS) and noise reduction technology.

To analyze the seismo-acoustic characteristics of the 30–31 August Black Sea storm, we used a two-pronged approach: (1) single sensor signal analysis based on feature extraction and unsupervised machine learning, and (2) array-based analysis using all the sensors of AGIR and classic multi-channel correlation algorithms. Together, these methods provide complementary insights into the acoustic behavior of the storm, capturing both local signal characteristics and spatial coherence across sensors.

For the single-station analysis, infrasound data recorded at the AGIR sensor (Fig. 2) was segmented into 30 min windows, and a set of time-frequency features was extracted to characterize the signal dynamics (Supplement). These features describe how energy and frequency content evolve over time, providing insights into the structure of the infrasound signal. Parameters such as spectral centroid and spectral rolloff are standard descriptors in acoustic signal analysis and are suitable here because they effectively capture shifts in dominant frequency produced by lightning-generated acoustic waves or the passage of pressure disturbances, while spectral flux highlights changes in broadband acoustic energy (Pásztor et al., 2023). Spectral entropy reflects the complexity of the frequency distribution, which increases during turbulent atmospheric conditions, and the zero-crossing rate, mean, and variance of the power spectrum summarize overall activity and variability. This feature set provides a compact representation of the signal suitable for unsupervised machine-learning approaches such as clustering, techniques widely used in data mining to identify patterns in multidimensional time-frequency data (e.g., Coates and Ng, 2012), and allows us to distinguish physically interpretable stages of storm-induced changes in the infrasound wavefield.

The extracted features were used as input for K-Means clustering (MacQueen, 1967), an unsupervised machine learning algorithm that partitions data into a predefined number of groups. K-Means minimizes within-cluster variance by iteratively assigning feature vectors to the nearest cluster centroid and updating the centroids based on the grouped data. This clustering method enables the identification of distinct acoustic patterns in the signal (e.g. Pásztor et al., 2023), offering a data-driven way to segment the storm's infrasound profile without requiring prior labels or assumptions. Prior to clustering, the features were standardized using z-score normalization, to ensure comparable scaling across variables. The optimal number of clusters was determined using the elbow method, which evaluates within-cluster variance as a function of cluster number (Supplement). To select the most informative features, we applied covariance pruning, and the temporal evolution of the features was visualized to ensure meaningful representation. This procedure resulted in six clusters, providing a balanced representation of the infrasound dynamics while avoiding over-segmentation or overfitting. By combining multiple features in the clustering, this method captures the evolving acoustic states of the storm in a compact, interpretable form.

In parallel with the single-station analysis, we also applied the Progressive Multi-Channel Correlation (PMCC) method, as implemented in the DTK-PMCC software (Cansi and Le Pichon, 2008; Le Pichon et al., 2010) to detect and analyze coherent acoustic signals across an infrasound array. The PMCC method targets signals generated by atmospheric sources such as lightning (i.e., associated thunders) or other pressure disturbances, operating in the low-frequency range of 0.7 to 7 Hz. It is specifically suited for mini-array configurations, where signal coherence between closely spaced sensors can be exploited for precise signal detection and characterization.

The PMCC algorithm was implemented using a multi-resolution configuration following the standardization proposed by Garcés (2013), with window lengths and frequency bands arranged in third-octave bands. A total of 19 frequency bands were used, covering 0.1–7 Hz. Window lengths decrease logarithmically with frequency, ranging from 258 s in the lowest band to 4 s in the highest band. A 10 % time step was applied (corresponding to 90 % overlap between consecutive windows), and this scheme repeats every decade. Within each time-frequency segment, cross-correlations are computed between all sensor pairs to identify coherent wavefronts, signals that exhibit consistent arrival times across the array. From these detections, PMCC estimates several key propagation parameters, including backazimuth (the direction of arrival), horizontal trace velocity, amplitude, duration, and dominant frequency. This approach is particularly effective in noisy environments and enables the discrimination of storm-generated infrasound from background signals or unrelated acoustic sources. The algorithm's output consists of a time-frequency map of signal detections enriched with physical metadata, allowing for detailed interpretation of the storm's acoustic footprint and its temporal evolution.

We also incorporated data from the Meteosat Third Generation (MTG) satellite system (Holmlund et al., 2021), specifically from its Lightning Imager (LI) sensor (Viticchie et al., 2020). The MTG satellites operate in geostationary orbit at approximately 36 000 km altitude, providing continuous observations over Europe, Africa, and surrounding waters. The Lightning Imager detects cloud-to-cloud, cloud-to-ground, and intra-cloud lightning flashes using four cameras that collectively cover 86 % of the Earth's visible disc from the satellite's perspective.

For this study, we used Level 2 group data, which includes the geographical coordinates and timing of each detected flash. The MTG Lightning Imager detects total lightning (cloud-to-cloud and cloud-to-ground) optically at 777 nm, with 4.5 km pixel resolution at the sub-satellite point and 1 ms frame interval (Holmlund et al., 2021; Kokou, 2023). Level-2 achieves detection efficiencies of ∼80 %–90 %, capturing even weak flashes reliably, with false alarm rates <0.3 (Enno et al., 2025). Flash geolocation uncertainty reaches 5–10 km near the edge of the instrument's field of view, where off-nadir viewing geometry amplifies parallax effects (Bližňák and Sokol, 2026). By mapping these detections, we were able to analyze the spatial distribution and temporal evolution of the storm's lightning activity. The dataset also offered insights into the storm's intensity and structure, complementing other meteorological observations.

Associations between infrasound detections and lightning flashes detected by MTG within 50 km of the AGIR infrasound station were investigated by assuming direct-path acoustic propagation and a correspondence between infrasound time-of-arrival and the MTG lightning discharge time (after Assink et al., 2008): 1t=tMTG+d/c+Δt, where d is the distance between the lightning discharge and the infrasound station, c=340 m s⁻¹, and Δt=±10 s accounts for timing uncertainty associated with the simplified propagation assumption. In particular, infrasound travel time from thunder sources can vary due to atmospheric temperature and wind variations along the propagation path, which affect the effective sound speed and may introduce deviations from the assumed constant-velocity, straight-path propagation. Additionally, a maximum angular deviation of 10° between the observed infrasound backazimuth and the MTG-derived backazimuth is permitted for an association to be accepted.

The use of GNSS technology for atmospheric monitoring provides a powerful tool for analyzing extreme weather events. Beyond its well-known applications in navigation, timing, positioning and crustal dynamics (Nistor et al., 2021a, b), GNSS has become a reliable method for sensing tropospheric water vapour, an essential driver of weather systems and a key variable in forecasting models (Guerova et al., 2016; Vaquero-Martínez and Antón, 2021). Over the past two decades, ground-based GNSS networks in Europe have contributed significantly to operational meteorology by providing near real-time estimates of atmospheric water vapour, aiding in the detection and tracking of severe weather, including heavy rainfall and storms (Karabatić et al., 2011; Priego et al., 2017; Jones et al., 2020). These high-resolution observations have proven valuable for both nowcasting and validating numerical weather prediction models (Wilgan et al., 2015; Bosy et al., 2012; Awange, 2012).

In this study, GNSS data were collected from several networks (Fig. 2), including the International GNSS Service (IGS, Johnston et al., 2017), the EUREF Permanent Network (EPN, Bruyninx et al., 2012), the Romanian Position Determination System (ROMPOS, Iliescu et al., 2019), and GEOPONTICA (Dimitriu et al., 2017). A total of 37 permanent GNSS stations were analyzed over a 30 d period, with the rainiest interval selected at the midpoint of the study period. These stations provide high-quality, continuous observations critical for atmospheric monitoring.

The data were processed using a double-differenced, ionosphere-free combination of L1 and L2 carrier phases. This approach helps minimize errors such as ionospheric delays, satellite clock biases, and other common atmospheric effects. The resulting Zenith Tropospheric Delay (ZTD) values were then corrected using the Vienna Mapping Functions 3 (VMF3, Landskron and Böhm, 2018), which improves the accuracy of ZTD by accounting for variations in the troposphere's atmospheric conditions. Once the ZTD was refined, it was converted into integrated precipitable water vapor (PWV) using surface meteorological data (temperature and pressure) from co-located weather stations, following the method outlined by Bosy et al. (2012). This process allowed for the derivation of high-resolution atmospheric water vapor content, critical for analyzing the dynamics of the extreme storm event over the Black Sea. By combining GNSS-derived PWV with data from other observational sources, the study captured the temporal and spatial variations in atmospheric moisture, offering valuable insights into the storm's development and intensity.

To compare the infrasound signals captured during the Black Sea extreme storm event, we extracted meteorological data from the open-access ERA5 reanalysis dataset, produced by the European Centre for Medium-Range Weather Forecasts (ECMWF). This dataset provides a comprehensive record of global weather conditions from 1950 to the present (Hersbach et al., 2023). ERA5 combines observational data and advanced numerical models to generate relatively high-resolution atmospheric parameters compared with earlier global reanalyses, including precipitation (Fig. 1), wind speed, and wave height. ERA5 has been extensively validated (Jiao et al., 2021; Wu et al., 2022; Soci et al., 2024) and is widely used in studies of storm evolution and precipitation dynamics (e.g. Dullaart et al., 2020; Tiberia et al., 2021; Price et al., 2025), making it a suitable choice for the mesoscale processes examined here.

For our study, the ERA5 data was used to track the meteorological context of the storm, offering insights into the intensity of precipitation, the evolution of wind patterns, and the development of oceanic wave heights. With a temporal resolution of 1 h and spatial resolution of 0.25° × 0.25°, ERA5 allows for a mesoscale comparison of the storm's meteorological characteristics over time. While its spatial averaging cannot resolve localized convective-scale precipitation, it provides a vital benchmark for qualitative comparisons and for testing multi-sensor monitoring potential. These comparisons help us understand the storm's dynamics and assess its impact, further enhancing the interpretation of infrasound signals and aiding in future storm prediction and monitoring efforts. The open-access nature of ERA5 ensures broad accessibility, contributing to the transparency and reproducibility of our storm analysis (Copernicus Climate Change Service, Climate Data Store, 2023).

High frequency (>30 Hz) analysis of seismic noise reveals strong signals during periods of intense rainfall (Fig. 4). Specifically, the displacement envelope at station MANR and its spectrogram for 30 August, 12:00 UTC to 31 August, 06:00 UTC (Fig. 4b, c) reveal strong signal around midnight, when recorded precipitation exceeded 20 mm per 10 min. Similar temporal patterns in the seismic spectrogram were also visible when compared with hourly precipitation levels from ERA5, indicating that the high amplitude of energy observed above 30 Hz is most plausibly generated by raindrop impacts.

However, this correspondence is not uniform across all rainfall episodes. While the main precipitation maximum on 30–31 August produces a clear and sustained seismic response, several lower-intensity precipitation pulses show a much weaker or no recognizable signature in either the seismic envelope or spectrogram. This behaviour is consistent with previous work (e.g., Rindraharisaona et al., 2022), which demonstrates that only rainfall above a certain intensity, or involving sufficiently large drops, generates impact forces strong enough to be detected by broadband seismometers. Our observations therefore reflect both strong positive correlations during intense rainfall and the lack of seismic expression for weaker precipitation. This selective sensitivity supports the interpretation that high-frequency seismic noise can reflect strong rainfall peaks but is less responsive to light or moderate precipitation, an important nuance when interpreting multi-sensor relationships in this study.

Anthropogenic seismic noise is typically strongest at low to mid frequencies (<25 Hz), where day-night variations reflect traffic, human activity, and transient signals from machinery, while higher-frequency bands (25–45 Hz) may include periodic contributions from rotating equipment (e.g., Groos and Ritter, 2009; Díaz et al., 2017). The bandwidth targeting rainfall in this case is between 30–50 Hz, which is above the dominant frequency content of most anthropogenic sources and overlaps with raindrop-impact energy documented in recent rainfall-seismic studies.

To visualise the signature of the storm passing over the network of broadband seismic stations in the coastal area, we also plotted the hourly precipitation values with the hourly root-mean-square amplitudes of the high-frequency (>30 Hz) seismic velocity envelopes recorded at seismic stations. Figure 5 shows four snapshots of hourly plots of gridded precipitation data from ERA5, which have a lower amplitude than point measurements at the Mangalia station, due to the averaging over the grid block. This figure presents a temporal coincidence between changing precipitation patterns from ERA5 data and the amplitudes of high-frequency seismic noise. This observation further supports the likelihood of a causal relationship. These high-frequency seismic signals could potentially be explored as a near real time indicator of intense rainfall events, providing a conceptual basis for a simple streaming detection approach.

The analysis of the microseismic noise frequency band is closely linked to the interaction between ocean waves and the seafloor, which is influenced by storm conditions. To assess the storm's impact, we analyze the PPSD (Probabilistic Power Spectral Density) of noise recorded at several stations during both storm and quiet days, using the latter as a baseline. Figure 6 shows examples of PPSD at stations MANR and MFTR (Fig. 2), revealing differences in PSD amplitudes across the primary and secondary microseismic bands. These differences indicate the presence of high-intensity wind-driven waves and swell energy in the sea.

The secondary microseismic band, in particular, shows a significant rise in amplitude during storms, consistent with established mechanisms linking storm-driven wave activity and seafloor pressure fluctuations (Fig. 3) to enhanced secondary microseism generation, while local factors such as bathymetry or wave direction may modulate the response (Bromirski and Duennebier, 2002; Ebeling, 2012; Ardhuin et al., 2019). On quiet days, the PSD remains consistently lower, typically staying below the -120 dB threshold. This stark contrast emphasizes the role of atmospheric conditions in modulating seismic noise, with storms causing a notable increase in energy across both frequency bands. The temporal evolution of the PSD values (Fig. 7) further highlights the storm's impact, with fluctuations corresponding to changes in environmental factors, reinforcing the connection between storm activity and the observed seismic signals.

Anthropogenic seismic noise does not significantly affect the microseismic band (0.1–1 Hz). Human-generated vibrations predominantly occupy frequencies above 1 Hz, while long-period microseisms are produced by ocean wave interactions and are coherent over large distances. The temporal evolution of the microseismic energy observed in this study matches changes in wave state associated with the storm rather than any local activity. Similar to the findings of Groos and Ritter (2009), the sub-Hz frequency range is dominated by natural sources, with anthropogenic contributions being negligible.

The analysis of daily GNSS-derived precipitable water vapor (Fig. 11) reveals clear temporal variations, with the highest PWV values consistently recorded on stormy days (>40 mm on DOY 240–243, i.e. 27–30 August). Notably, the peak values occurred between DOY 241 and DOY 243 (Fig. 11b), when the heaviest rainfall was observed (Fig. 1). Coastal stations showed extremely high PWV values (>40 mm) compared to inland stations (<30 mm), with a slight decrease in PWV away from the coast (Fig. 11a). This spatial distribution highlights the geographical gradient of atmospheric moisture, with the highest PWV concentrations near coastal areas, also decreasing gradually toward the north away from the storm peak. Interestingly, some inland stations (BUCU, PGNL, RMSR) recorded their peak PWV on DOY 255, corresponding to the onset of the Boris storm, another significant extreme rainfall event that swept through Central and Eastern Europe (Athanase et al., 2024).

Elevated PWV was observed as early as 27 August (Fig. 12a), suggesting that the tropospheric moisture loading began to increase several days before the onset of the rainfall. This increase in PWV may act as an indicator of a developing weather system. Remarkably, although HAR1, located inland, did not directly experience the extreme rainfall, it exhibited similar PWV behavior to coastal stations, suggesting that GNSS stations, even outside the immediate storm zones, can capture atmospheric signals indicative of intense precipitation. This finding offers a valuable precedent, showing that PWV measurements at GNSS stations not directly in the storm path can still provide critical insights into moisture dynamics at the tropospheric level. The comparison with ERA5's total-column water vapour further supports this interpretation, as the broad temporal evolution of ERA5 humidity mirrors the GNSS-derived daily PWV patterns, despite the inherently coarser resolution of the reanalysis data.

Using the hourly PWV data, Fig. 12b illustrates the evolution of water vapour at the MNGA station, which recorded the heaviest rainfall in the study area. Notably, MNGA also showed a rapid buildup of PWV, reaching values greater than 44 mm just a few hours before the storm event. This rapid increase in PWV suggests that the accumulation of atmospheric moisture may precede extreme weather events, such as intense rainfall and storms. This observation aligns with known atmospheric dynamics, where a significant increase in water vapor content precedes heavy precipitation, highlighting the potential usefulness of GNSS-based PWV monitoring for studying pre-storm atmospheric moisture variability. The general rising trend toward the event is present in both GNSS-based and ERA5 reanalysis datasets, although some minor fluctuations are not matched. After the storm, the GNSS PWV drops sharply while ERA5 maintains elevated values for several hours. These differences show that GNSS can resolve rapid, real-time atmospheric changes that may be blurred in large-scale weather model products.