In this study, we developed a new real-time multi-station seismic monitoring system for volcanoes without any prior knowledge within a transfer learning framework. Although some classical ML techniques such as Markov models have been used in sequence modeling tasks, neural networks (NN), including both RNN and TCN architectures (LeCun et al., 2015; Lea et al., 2016), have optimal temporal modeling capabilities. By generating a spatio-temporal sequence of hierarchical features, both architectures have been applied in complex and emerging geosciences research fields, including seismo-volcanic monitoring (Titos et al., 2018; Bueno et al., 2021), climate change (Yan et al., 2020), remote sensing (Račič et al., 2020), and human activities recognition (Nair et al., 2018). Accordingly, in this work, an RNN based on long short-term memory cells (RNN–LSTM) and a TCN (Titos et al., 2018; Titos et al., 2022) trained with a master catalogue belonging to Deception Island volcano (Antarctica) were proposed as a baseline. These were then used as a blind-recognizer for the data from a different volcanic environment, Mount Bezymianny (Kamchatka). The master database belonging to Deception Island volcano is unbalanced; however, it has been thoroughly reviewed by experts on the volcano. According to (Titos et al., 2018), the Deception Island dataset is composed of five seismic categories: background noise (BGN), tremor (TR), hybrid (HYB), VTs, and LPs; Table 1 summarizes the performances of the two approaches (RNN–LSTM and TCN) using the master catalog, based on the percentage of events correctly recognized.

Then, assuming a scenario in which the monitoring agency does not have any previous information related to the seismic dynamics of a volcano, a new monitoring tool was obtained as follow (see Figure 1):

1. Data parameterization: Raw streaming data belonging to each seismic station within the new volcanic environment were parameterized following the parameterization scheme of (Titos et al., 2018) to obtain the baseline systems.

The pipeline used for this study is suitable for application to other baseline systems and parameterization schemes.

Bezymianny volcano (55.6°N, 160.3°E) is an explosive basaltic–andesitic stratovolcano belonging to the Klyuchevskaya (KVG) volcanic group on the Kamchatka Peninsula, Russia. It is located in the central depression of Kamchatka (CKD), which covers > 4,000 km² between the Sredinny and Eastern ridges. This region marks the northeastern corner of the Pacific subduction plate, which is formed by the Kuril–Kamchatka and Aleutian trenches (Figure 2A). According to its eruptive history, the volcano was considered inactive for more than 1,000 years (Braitseva and Kiryanov, 1982), until the lateral eruption in 1956. Bezymianny has experienced an active period since 2000, with more than 15 eruptive episodes (Van Manen et al., 2010). Among its recent eruptive episodes, that of 20 December 2017 (Girina et al., 2018) produced an eruption column that exceeded 15 km in height, representing a potential hazard to air traffic (Neal et al., 2009; McGimsey et al., 2014). The seismic database associated with this eruption is reliable and complete; therefore, it was selected for testing the approach developed in this study.

The seismic data used in this study were collected by a temporary network composed of 10 seismic stations, installed during the 2017–2018 period (Koulakov et al., 2021). However, only data corresponding to four stations (Figure 2B) were selected. Criteria for selecting the seismic stations were motivated by both the availability and quality (signal-to-noise ratio) of the data. In addition, to further determine the reliability of the monitoring system proposed, two additional eruptive phases (a pre-eruptive stage characterized by little activity and a syn-eruptive stage with tens of thousands of events) containing 6 months of seismic data from June to December 2017 were also selected.

Figure 3A summarizes the VTs detected by the pre-trained RNN–LSTM system before and after re-training using the new (Bezymianny volcano) dataset. Contrary to expectations, the number of VTs detected at some stations using the RNN–LSTM remained constant or decreased after being re-trained. Figures 4, 5 show comparisons of the monthly per-station cumulative distribution function (CDF) before and after the re-training process, representing the probabilities and normalized cumulative sums of events predicted as VTs. Before re-training, while a high number of events were detected, the confidence of such detections was low. More specifically, in August 2017 (Figure 4A), almost 70% of the events detected had probabilities of between 35% and 55%; in October and December 2017 (Figures 4B, C), except at station BZ06 (where 50% of the events detected had probabilities of < 55%), no recognized event exceeded 55%. After re-training, there was a clear change in the trend, with fewer recognized VT earthquakes depending on the station (Figure 3) but much higher confidences of the detections (Figure 5).

Figure 3B summarizes the VTs detected by the pre-trained TCN system before and after re-training using the new (Bezymianny volcano) dataset. Figures 6, 7 show comparisons of the monthly per-station CDF before and after the re-training process. In contrast to the RNN–LSTM, the TCN architecture saw an increase in the total number of earthquakes detected after being re-trained but a significant decrease in the confidence of the recognitions. Before re-training, 90% of events were detected with probabilities of > 80%; after re-training, depending on the station, only 40%–60% of recognized events had probabilities of > 80%.

To determine the robustness of our system, we compared our results before and after re-training to those of a classical approach, the short time average over long time-average (STA/LTA) trigger algorithm (Trnkoczy, 2009). We selected a single day on which several hundred earthquakes occurred and analyzed the results on an hourly timescale. Given that the TCN always detected a greater number of events than the RNN–LSTM, we assumed that the VTs detected by the RNN–LSTM were a subset of those detected by the TCN and selected only those events for analysis. On the chosen day, the RNN–LSTM did not recognize any VTs before it was re-trained; after re-training, all of the VTs recognized had previously been categorized as TR events.

Figure 8 presents an overview of the STA/LTA triggering thresholds. For proper operation of the STA/LTA algorithm, four parameters should be tuned: the short window length (STA), long window length (LTA), activation threshold level, and deactivation threshold level. The STA/LTA trigger parameter settings are always a tradeoff among sensitivity and specificity. While sensitivity may also include a tolerable number of false triggers, specificity correctly detects only particular instances, therefore decreasing the number of detections. Considering that the algorithm computes the average absolute amplitude of a seismic signal in two consecutive moving-time windows, only events exceeding pre-set values describing the triggering thresholds of both STA and LTA were identified. Figure 9 compares the number of VTs detected hourly during 14 August 2017, by the STA/LTA trigger algorithm and re-trained RNN-LSTM architecture; overall, the results show that the RNN–LSTM recognized a higher number of VTs than the STA/LTA algorithm (782 vs. 648).

The noticeable difference in the performance of the system after re-training can be explained from a geophysical perspective. First, owing to the noisy content registered in the new volcanic environment, it is possible that many of the events detected before the re-training were low probability VTs (or mis-recognized VT) corresponding to seismic traces characterized by high frequencies and variable length. Since the system fitting approach is a density estimation problem and such seismic traces partially match observed VT features in the master catalog (Deception Island), the estimated probability density function and its parameters cannot explain the underlying distribution of the new input data; as such, it assigns a low probability. Figure 10A provides a clear overview of this issue, in which a seismic trace partially matches with source earthquakes with a low probability (52%).

Second, many of the VTs detected after re-training were originally recognized as TR, which can be explained by the differences between the learned representation at source and target underlying distributions. Figure 11 shows the spectrograms and power spectral densities (PSD) of VTs from Bezymianny volcano and Deception Island, and a TR from Deception Island. The figure shows very similar spectral energy distributions. The beginning of the Deception Island TR (Figure 11B) has a short and overlapped package of high frequency waves (up to 20 Hz). These high frequency signals are associated with the explosive step of pressure in the source region when LP events are generated near the seismic station (no visible exponential decay in frequency is observed) and with small earthquakes. At Bezymianny volcano, many VTs have a higher energy component at low frequencies (Figure 11C); therefore, as our parameterization scheme performs energy analysis by frequency bands that are more sensitive in lower frequencies, the pre-trained system failed to recognizing these energetic low frequency VT events. After re-training, the global number of recognized VTs was similar, but confidence of the detections was much higher.

In summary, the system (i.e., the probability distribution and associated parameters) is fixed to maximize a likelihood function that best explains the joint probability distribution of the new volcanic dynamics (in this case, Bezymianny volcano). As a result, following re-training, only confident VTs were detected and labeled. Those previously mis-recognized as TR were now confidently detected, while those events with low probabilities ( < 65%) were labeled as undefined events (e.g., Figure 10B). Such events require careful review by experts. As conclusion, before re-training, only VTs with probabilities of > 80% could be included in the new catalog; after re-training, all VTs detected with probabilities of > 65% could be easily identified.

The noticeable difference in the performance of the system after re-training can be explained by the greater specialization ability of TCN compared with RNN–LSTM owing to the multi-resolution dilated skip connections between layers and deeper hierarchical features. The new Bezymianny volcanic dynamics often exhibit consecutive seismic events that bear partial resemblance to earthquakes occurring closely together in a short timeframe. Before retraining, the system could avoid recognizing such new volcanic dynamics based on high frequencies and short length as isolated events. Therefore, when such concatenated events were detected, focusing only on those volcanic dynamics that were similar to the master catalog, the system considered them all as a whole, not as isolated events (Figure 12A).

In summary, before re-training, the system labeled seismic traces partially matching with earthquakes as high probability VTs, while low probability events were labeled as ‘undefined’. After re-training, all seismic traces with high frequencies and variable length were detected as VTs, decreasing the number of undefined events. However, only clear VTs were detected with high probability. Figure 12B shows an example of a detected low probability (70%) VT. Before re-training, this seismic trace was labeled as ‘undefined’ with a high probability of assignment to VT ( > 90%). After re-training, the system decreased the probability of assignment to VT. In this way, before re-training, many VTs were mis-recognized; after re-training, all VTs detected with probabilities higher than of > 85% could be included in the new catalog.

However, in some time slots, STA/LTA detected a greater number of events. These results may be explained by the nature of the STA/LTA algorithm, its trigger parameter settings, and the grammar imposed on the proposed models, which was responsible for improving the interpretability of the models based on geophysical knowledge of the volcano (Titos et al., 2018).

Since there was no previous information related to the seismic catalog, the STA/LTA triggering thresholds were fixed so that the system was more sensitive than specific. The goal was to obtain as much information as possible, and all energy changes, even small ones, were detected. This scenario resulted in a tolerable number of false triggers. In contrast, the RNN–LSTM (and TCN) system imposed the use of grammar (a set of rules) based on geophysical knowledge of Deception Island volcano to improve interpretability. The average duration of seismic events belonging to the master dataset (Deception Island) in combination with the per-class probabilities output by the models in the new volcanic environment allowed us to check that the predictions were consistent with the expected lengths of events. Since no information was provided on the average duration of the seismic-volcanic events of the new volcanic environment, the grammar only recognized those events that, on average, had durations that were greater than or similar to those described in the master dataset. Events whose durations were less than the average duration of events in the master database, even if recognized with high per-class probabilities, were labeled as background noise or unknown events.

For the time slots in which STA/LTA detected a greater number of VT events compared with RNN–LSTM, many of the events recognized by the STA/LTA model corresponded to short duration energy changes (Figure 13). In contrast, the RNN–LSTM model discarded these events (i.e., labeled them as background noise when the output VT per-class probabilities were low and as unknown events when the output per-class probabilities were high) since the durations were shorter than the average duration of VT earthquakes at Deception Island.

For the time slots in which STA/LTA detected a lower number of VT events compared with the re-trained RNN–LSTM, a possible explanation is the behavior of the STA/LTA algorithm in a seismic swarm state. Seismic swarms, which are a common volcanic phenomenon, involve a sequence of seismic events that occur within a relatively short period of time within a very local area. Given that the STA/LTA algorithm computes an average absolute amplitude of the seismic signal in two consecutive moving-time windows, when a low energy event occurs immediately after a high energy event, the averaging process masks the occurrence of the least energetic one, decreasing the number of recognized events (Figure 14). In contrast, as the RNN–LSTM analyzes signals based on spectral features, it has the ability to analyze a concatenated occurrence of events, such as that observed during a seismic swarm.

Based on our results, once it has been re-trained and the average duration of the seismic-volcanic events has been fixed, our RNN–LSTM has a number of advantages over STA/LTA. In particular, the system will automatically detect VTs present in the seismic trace without searching for optimal parameter settings, which makes it a portable, scalable, and economical tool with relatively low computational cost. Another important advantage is that, besides obtaining a preliminary seismic catalog (composed of several types of events), it offers information on the confidence of the recognition. Importantly, for multi-station seismic networks, these probabilities will serve to obtain more reliable seismic catalogs. The recognition of an event characterized by high frequencies at one station provides an indisputable condition to obtain a reliable label at another station where the event has been attenuated or is virtually unrecognizable. An example of this scenario is shown in Figure 15, in which one earthquake is included in the new seismic catalog by both techniques (STA/LTA and RNN-LSTM), while an attenuated one, in addition to presenting difficulty during detection using classical techniques owing to threshold adjustment, could only be considered as an earthquake using our multi-station analysis. Finally, our approach provides a very useful tentative label for subsequent analysis carried out by a human operator.