Identifying attacks in the Russia-Ukraine conflict using seismic array data

The study addresses the challenge of obtaining comprehensive, objective, and real-time information about ongoing military attacks in active conflict zones, where traditional and social media can be incomplete or biased. The authors investigate whether seismic and associated acoustic signals from explosions can be leveraged to detect, locate, and characterize individual attacks in near real time. They motivate the work by noting that while the International Monitoring System is optimized for global nuclear test monitoring, detecting lower-yield, regional military explosions requires nearby dense sensor networks. The purpose is to demonstrate an automated seismic monitoring approach applied to the Russia–Ukraine conflict, providing accurate timings, locations, and magnitudes of explosions and enabling verification of reported incidents and identification of unreported attacks.

The paper situates its contribution within prior work on conflict monitoring using remote sensing and seismoacoustic observations. Satellite imagery provides high spatial resolution but requires prior knowledge of time and place and lacks real-time regional completeness. Seismoacoustic methods have historic roots in sound ranging for artillery location during World War I and have evolved to locate artillery and infer properties of large explosions using acoustic sensor networks. The International Monitoring System today includes over 200 seismic and infrasound stations, and abundant open seismic data exist for earthquake and Earth structure studies. Prior literature has focused on experimental datasets for artillery localization or characterization of individual large explosions, with a gap in real-time analysis of seismic and acoustic signals from active conflicts due to limited data availability. The authors build on migration/stacking approaches and characteristic function methods (e.g., STA/LTA) developed for automatic detection and location in seismology and adapt them for conflict-related explosions near a seismic array.

Study area and sensors: The work uses the IMS Malyn AKASG seismic array (PS45), about 100 km northwest of Kyiv, Ukraine. The array comprises 23 vertical-component broadband seismometers plus one three-component broadband station, with an aperture of ~27 km and ~2 km inter-sensor spacing. Originally optimized for teleseismic plane-wave detection, the array is repurposed for local/regional events by abandoning the plane-wave assumption and exploiting individual station observations. Data flow and processing: Continuous waveform data are transmitted from Ukraine to the International Data Centre in Vienna and then to Norway for automatic processing, achieving close to real-time results. The automatic detection and location pipeline is based on migration/stacking methods applied to characteristic functions derived from short-term average to long-term average amplitude ratios (STA/LTA), tuned to detect P- and S-wave onsets at each sensor. Coherent signals are stacked and time delays used to estimate source locations near the array. Air-to-ground acoustic (infrasound) arrivals, although sometimes observed on seismometers, are omitted from the automatic location stage due to their infrequent, less-impulsive nature and the risk of false detections and reduced sensitivity. Event characterization: For each detected event, origin time, location, and local magnitude are automatically estimated. Acoustic arrivals are sought in post-processing by stacking seismic envelopes in time windows corresponding to acoustic propagation velocities, which can refine both spatial and temporal constraints where present. Spatial coverage and detection completeness: The monitoring region spans roughly 300 km by 222 km covering parts of the Zhytomyr, Kyiv, and Chernihiv provinces. Detection capability diminishes with distance from the array, reducing magnitude completeness at larger ranges (e.g., around Chernihiv ~170 km away). Yield estimation: Seismic local magnitudes are computed automatically. Empirical relationships between magnitude and explosive yield are used to estimate upper and lower bounds, acknowledging that underground nuclear explosion scaling is a poor analogue for surface blasts due to coupling differences. Additional yield estimates are explored using acoustic-phase amplitudes. Validation and quality control: The catalogue is compared against pre-invasion data (from 1 January 2022) to establish a baseline of mining/quarry explosions. Publicly reported attack data from Liveuamap are used for qualitative temporal comparison. The automatic detection threshold is set relatively low to favour true detections at the expense of more false positives; manual screening is performed to reduce false positives, which can arise from distant earthquakes/explosions causing aliased locations or from closely spaced multiple explosions causing misassociation of phases.

- Total detections: 1,335 automatic detections between 1 January and 3 November 2022, comprising 53 pre-invasion daytime mining/quarry explosions and 1,282 explosions from 24 February to 3 November 2022 within the ~300 × 222 km region around the Malyn array. - Spatiotemporal patterns: Clusters of military-related explosions were observed around Zhytomyr, Korosten, northwest of Kyiv, Chernihiv, and Malyn. The most prominent activity occurred northeast of Malyn in late February–March 2022, aligned with reported front lines and occupied regions. - Rates: From 24 February to 31 March, an average of 29 detected explosions per day was observed; the peak day was 7 March with 64 explosions. The last heavy bombardment was on 31 March; only two explosions were detected on 1 April. After the reported Russian withdrawal from the Kyiv region (2 April), activity largely returned to pre-invasion levels with sporadic targeted strikes. - Acoustic detections: Clear acoustic signatures were identified for about 29% of seismic events, improving spatial and temporal constraints when present. Some events were only acoustically observable (likely higher altitude, larger distance, or lower-yield explosions), underscoring the complementary role of acoustic monitoring. - Magnitudes and yields: Automated local magnitudes ranged from −1.25 to 2.24. Manual checks agreed within ~0.3 magnitude units; lower-magnitude events appeared closer to −0.6 in manual analysis, corresponding to an estimated 0.03–9.00 kg TNT (consistent with a 152-mm OF45 projectile yield of ~7.65 kg TNT). The largest clearly military-related explosion had M ≈ 1.7 (Chernihiv, 10 March 2022) with yield estimated between ~352 and 3,083 kg TNT; the lower bound is plausible for an air strike (e.g., Iskander missile ~700 kg), whereas the upper bound appears too high. The largest-magnitude events (M > 1.7) were associated with mining/quarry activity near Korosten. - Accuracy: The automatic seismic-phase method achieved spatial accuracy better than ~5 km and timing accuracy better than ~1 s for regional events within ~100 km of the Malyn array. Incorporating acoustic phases in post-processing further improved spatial accuracy. - External validation: The temporal trend in the seismic catalogue broadly matched aggregated public reports (Liveuamap), with seismic detections exceeding reported attacks on most days during the most active period (except 25–27 February). A case study on 20 May 2022 in Malyn showed three explosions automatically identified ~4 h before public reports; manual analysis located them within ~100 m of the ground-truth crater (automatic location ~1.4 km offset).

The findings demonstrate that a local seismic array can objectively monitor an active conflict in near real time, providing accurate origin times, locations, and magnitudes for hundreds to thousands of explosions. The spatial distribution and temporal evolution of detections correspond well to known zones and phases of military activity, validating the method’s relevance. The approach complements satellite imagery and media-based reporting by offering continuous, unbiased, and high-temporal-resolution coverage, enabling verification of reported incidents and discovery of unreported attacks. The detection of acoustic signatures for nearly a third of seismic events allows additional refinement of locations and timing, and highlights the importance of integrating seismoacoustic observations. Although yield estimation from seismic and acoustic data remains uncertain for surface explosions, the magnitude and approximate yield bounds provide rapid assessments of relative explosive strength. The methodology is transferable to other arrays and dense sensor networks near conflict zones and could support automatic characterization of weapon types, thereby enhancing scrutiny of potential breaches of international law.

This study provides the first known near-real-time seismic monitoring of an active military conflict, producing a comprehensive catalogue of explosions in northern Ukraine that surpasses publicly reported counts and offers precise timing and location information. By adapting migration/stacking detection to a local array and leveraging post-processed acoustic arrivals, the authors show that seismology can yield an objective, continuous view of conflict dynamics and validate specific incidents. Future research should focus on integrating acoustic phases into the automatic workflow, improving association of closely spaced multiple explosions, conducting yield calibration experiments for surface/airburst scenarios, combining seismic/acoustic observations with satellite and other open-source data, and developing automated classifiers for artillery and munition types to better quantify and interpret conflict activities.

- Detection completeness decreases with distance from the array, limiting sensitivity to smaller explosions in distant areas (e.g., ~170 km away at Chernihiv). - Automatic locations omit acoustic phases to reduce false positives, foregoing potential improvements in accuracy; acoustic arrivals are also often absent due to atmospheric propagation conditions and high-frequency attenuation. - False positives arise from distant earthquakes/explosions producing aliased locations and from closely spaced multiple blasts causing misassociation of phases; a low detection threshold increased true detections but also false alarms, necessitating manual screening. - Yield estimates based on seismic magnitude and acoustic amplitudes are poorly constrained for surface/airburst explosions due to coupling variability and lack of calibration; acoustic-based models tended to overestimate yields. - Most array sensors are vertical-component only, limiting full waveform characterization and source-type discrimination. - Some explosions are only acoustically detectable (e.g., higher-altitude or more distant, lower-yield events), and thus are underrepresented in the seismic-based catalogue.