The monitoring period for the QJ spring spanned from 1 June 2019 to 21 May 2024 (approximately 5 years), while the WN spring was monitored from 3 October 2021 to 21 May 2024 (approximately 2.5 years). A synchronous monitoring of hydrogen and oxygen isotopes was conducted at both springs between 1 January 2023 and 21 February 2024. Water temperature, pH, electrical conductivity (EC) and hydrochemical components for the thermal springs were measured every three days. Rainfall data were collected through continuous in situ monitoring using an RTP-II tri-element meteorological instrument with a resolution of 0.1 mm. Before collecting the thermal water samples, high-density polyethene (HDPE) bottles were thoroughly rinsed thrice with deionised water and twice with the thermal water. Water samples were then filtered through 0.45 µm micropore membranes and stored in HDPE bottles. Samples intended for cation analysis were acidified with high-purity nitric acid. During collection, care was taken to prevent the introduction of air bubbles, and the samples were immediately sealed hermetically to preserve them.

The concentrations of major ions (Na⁺, K⁺, Ca²⁺, Mg²⁺, Li⁺, F⁻, Cl⁻, SO42-, Br⁻, NO3-) were analysed using a Thermo Scientific Dionex Aquion IC system equipped with an AS40 autosampler, which had a detection limit of 0.01 mg L⁻¹. HCO3- and CO32- concentrations were determined via standard titration procedures using a ZDJ-3D potentiometric titrator with 0.05 mol L⁻¹ HCl. δD and δ18O values were determined using a Picarro L2140-i water isotope analyser, with precisions of 0.015 ‰ and 0.05 ‰, respectively. All analyses were conducted at the Key Laboratory of the Institute of Earthquake Forecasting, China Earthquake Administration. The monitoring data are detailed in the database of Shao (2025). To ensure data accuracy, cation–anion electrical balance (EB) tests were performed for each sample as a quality control measure, with all EB kept within ±5 %. Data that fulfilled this criterion were included in the subsequent analysis. The EB (Appelo and Postma, 2005) is calculated as below: 1EB=∑mc-∑ma∑mc+∑ma×100% where ∑mc represents the sum of cation concentrations (in milliequivalents per liter, meq L⁻¹), and ∑ma represents the sum of anion concentrations.

The anomaly detection model developed in this study aimed to forecast earthquakes with M≥4 by determining pre-earthquake anomalous signals. To identify earthquakes that might affect hydrochemical component variations while excluding those unrelated to precursors, and to establish a precise correlation between variations in hydrochemical components and seismic activity, an earthquake screening method based on the earthquake preparation zone radius (R, km) formula (Dobrovolsky et al., 1979) was employed: 2R=100.43M where M represents the earthquake magnitude. Based on long-term observations of 27 radon-involved earthquake cases in China between 1997 and 2020, 9 widely applied prediction methods were evaluated. The findings demonstrated that Dobrovolsky's formula achieved the highest applicability rate, reaching 96.30 % (Li et al., 2023). Its practical utility is further confirmed by recent studies in complex tectonic settings (Yakupoğlu et al., 2025), establishing it as an empirically robust and widely adopted reference scale for selecting potentially correlated earthquake events (Li et al., 2023; Seminsky and Seminsky, 2024; Zhu et al., 2024).

Earthquakes were selected as study events based on the criterion that the epicentral distance (Δ) from the thermal spring monitoring sites did not exceed R (Fig. 1a). The QJ site was within the preparation zones of 22 earthquakes with M≥4 during its monitoring period (1 June 2019–21 May 2024), whereas the WN site was within the preparation zones of 12 earthquakes with M≥4 during its observation period (3 October 2021–21 May 2024) (Table S1). All earthquakes had focal depths ranging from 8 to 16 km, and were classified as shallow-focus events. The earthquake catalogue was obtained from the National Earthquake Data Center of China (http://data.earthquake.cn, last access: 7 August 2024).

Seismic moment (M0), which directly reflects fault geometry parameters and the rigidity of the surrounding medium, accurately quantifies earthquake rupture processes and the mechanical energy released during the event. Compared with magnitude scales, the seismic moment is more suitable for analysing the seismic impact on hydrochemical component changes in thermal springs. The commonly used empirical formula for estimating seismic moment based on magnitude (Hanks and Kanamori, 1979) is expressed as follows: 3lgM0=1.5M+16.1 Stress attenuates with increasing epicentral distance during the earthquake preparation process, directly influencing the development of thermal water seepage pathways and the intensity of water–rock interactions (Wang and Manga, 2010; Ingebritsen and Manga, 2019). To account for distance-related effects, the seismic moment requires correction using the following empirical formula (Piersanti et al., 2016): 4M0cor=M0/Δω where ω is the weighting factor to modulate the relative contribution of distant and local earthquakes to the observed hydrochemical variations. A sensitivity analysis was performed by testing multiple values of ω (including 0, 0.5, 1, 1.5, 2, 2.5, and 3). It was observed that the cross-correlations peak (the methodology for which is presented in Sect. 3.3.2) consistently emerged within the same lag range across all ω values. Among these, the cross-correlations for ω=0, 0.5, and 1 were relatively significant and exhibited minimal differences (Fig. S1). Considering the physical interpretability, in this study, ω takes the value of 1.

The geochemical behaviours of different components in thermal spring water exhibit significant variations, with each element displaying distinct characteristics with respect to anomaly amplitude, temporal evolution, and precursor response sensitivity. Regulated by unique hydrogeological conditions, the hydrochemical variations of each thermal spring also display spatial differences in response to tectonic activity. To effectively extract anomalous signals, the anomaly responses of different components and springs in the study area are compared, and the algorithm's generalisability across springs is validated. This study establishes independent time series for each component at different springs (Figs. 2 and S2).

The average water temperature at QJ is approximately 60 °C, with a pH of 7.5 and an EC of 1148 µS cm⁻¹, while WN has a higher temperature of 80 °C, a pH of 7.9, and a lower EC of 579 µS cm⁻¹. Both QJ (sandstone) and WN (mylonite, gneiss, etc.) exhibit similar hydrochemical types (HCO₃-Na), due to the comparable lithology of the surrounding rocks. The δ18O values at QJ range from -13.19 ‰ to -11.81 ‰, while the δD values range from -94.93 ‰ to -90.59 ‰. At WN, the δ18O values range from -13.22 ‰ to -12.01 ‰, and the δD values range from -91.26 ‰ to -88.09 ‰. The narrow fluctuation range of stable isotopes in both thermal springs, coupled with their proximity to the local and global meteoric water lines (Fig. S4), indicates that the thermal spring water originates from atmospheric precipitation. Overall, the two springs exhibit similar hydrochemical characteristics, which minimises the impact of compositional differences on the evaluation of algorithm effectiveness across the different springs. For detailed hydrochemical ion concentrations and isotope values, please refer to the supplementary materials.

The cross-correlation analysis findings (Fig. 5) show that after the 15-day moving average treatment, the correlation (blue dotted lines) between rainfall and hydrochemical components are reduced to a weak level (within ±0.2). This result indicates that the moving average treatment effectively mitigates rainfall-induced noise. To validate the robustness of the denoising process, the results obtained from the moving average were compared with those derived from Fast Fourier Transform low-pass filtering and wavelet-based denoising techniques (Fig. S5). The approximately consistent outcomes across all methods confirm the suitability of the moving average approach for suppressing high-frequency noise. Furthermore, this method is better suited to the model's real-time anomaly detection framework. Notably, Na⁺, Ca²⁺, and δ18O exhibit minor response peaks at lags of 18–30 days, suggesting that these components remain continuously affected by rainwater infiltration within 15–30 days after precipitation. Similarly, the correlations between the distance-corrected M0 and the denoised hydrochemical component time series (red dotted lines) remain low (around ±0.2). However, K⁺, Ca²⁺, Cl⁻, SO42-, F⁻, δD, and δ18O exhibit weak response peaks at lags of -39 to -12 days, with varying peak directions for each component. This observation suggests that seismic activity (12–39 days before seismic moment release) may influence hydrochemical components, causing their concentrations to fluctuate (either increasing, decreasing, or remaining stable) due to different geochemical mechanisms.

Cross-correlation analysis revealed the hydrochemical responses preceding seismic moment release, while BCP analysis further captured these anomalous signals and identified their timing. The results show that change points are successfully detected in all component time series from both thermal springs (Fig. 6). Consistent with geological precursor studies, where low-probability signals are often retained due to the weak intensity of hydrochemical anomalies. For example, before the M4.6 earthquake on 3 March 2023, the posterior probability for a Ca²⁺ change point at QJ was 0.20 at 20 days before the earthquake, while SO42- exhibited a posterior probability of 0.19 at 28 days before the earthquake. Similarly, before the M5.2 earthquake on 16 November 2021, Cl⁻ at WN showed posterior probabilities for change points of 0.15 and 0.51 at 3 and 17 days before the earthquake, respectively. Notably, most change points are identified within 45 days preceding the earthquakes. This observation suggests that component concentration changes are sensitive to earthquake preparation processes and occur before earthquakes, which provides critical empirical support for anomaly detection algorithm models.

Notably, the timing and posterior probabilities of change points exhibit significant uncertainty, reflecting the complexity of factors influencing variations in hydrochemical components. These factors include inhomogeneity in stress accumulation, the structural complexity of fault and aquifer systems, modulation by deep gas degassing, and the mixing effects of multi-source fluids (Skelton et al., 2014; Kim et al., 2019; Hosono et al., 2020). In particular, the variable dominance of different mechanisms (e.g., water-rock interaction and fluid mixing) in different earthquake cases leads to inconsistent timing and intensity of hydrochemical anomalies, thereby causing uncertainty in BCP-detected change points. Among these complex factors, well-established mechanisms like water-rock interaction and fluid mixing can effectively explain several detected anomalies. For instance, the rise in Na⁺ concentrations before earthquake may result from a switchover to nonstoichiometric dissolution of analcime at fresh rock surface with preferential release of Na⁺ into groundwater (Andrén et al., 2016). Similarly, the increase in δ18O before the M6.6 Tottori earthquake in southwestern Japan has been attributed to enhanced water–rock interaction due to rock strain during the earthquake preparation process (Onda et al., 2018). Furthermore, the mixing of fluids with significantly different isotopic and hydrochemical compositions can cause variations in δD, δ18O, and ion concentrations. Previous studies in Iceland have reported pre-earthquake increases in Na⁺ concentration, δD, and δ18O due to mixing with different groundwater (Skelton et al., 2014, 2024). A similar phenomenon is observed in the elevated EC and ion concentrations preceding the Mw5.0 Mudanya earthquake in western Turkey (Yakupoğlu et al., 2025).

Compared with the Na⁺ detection results at WN, the Cl⁻ time series produces two false positives for 2022 and misses three earthquakes for 2023 (Fig. 6c, d). This result suggests that the analysis requires a longer time series and is highly sensitive to previous distribution settings. Larger-amplitude component changes often overshadow smaller-amplitude variations, which makes the latter difficult to detect and more prone to missed detections or false positives. In addition, most BCP methods have a fundamental limitation: they inherently perform retrospective analysis on complete time series. Specifically, identifying a change point at time ti relies on data collected after ti (t>ti), making real-time, forward earthquake forecasting unfeasible with short-term data sequences (Piersanti et al., 2016). According to the BCP analysis results, change-point detection for earthquake forecasting should be viewed as a supplementary approach for anomaly detection model to locate anomaly timing, guide parameter optimization, and conduct comparative verification.

Figures 8 and 9 present the 15-day moving average time series of hydrochemical components, anomaly detection results, and earthquake events for the anomaly detection model at QJ and WN. For each component, the model successfully identifies varying numbers of pre-earthquake anomalies and triggered alarms. The model activates comprehensive alarms when anomalies are detected in three or more components simultaneously. Based on the sensitivity analysis results presented in Table S5, requiring three or more anomalous components was identified as the preferred criterion for balancing the POD and the FAR. Furthermore, this choice is hydrochemically justified by its ability to suppress random dual-component interferences stemming from a single source, as discussed previously, thereby leading to the highest TS among all tested numbers of components. As detailed in Table S6, the 1,000-iteration bootstrap resampling results provide a robust evaluation of each individual component's performance alongside the comprehensive alarm (DA). This analysis demonstrates that even at the lower bound of the 95 % confidence interval (CI), the model maintains a reliable detection capability. At QJ, the model provided 21 effective alarms for 22 earthquake events (POD = 0.95, 95 % CI: 0.85 ∼ 1.00), with 8 false alarms (FAR = 0.28, 95 % CI: 0.10 ∼ 0.45), and a TS of 0.70 with a 95 % CI of 0.53 ∼ 0.87. At WN, the model generated 10 accurate alarms for 12 events (POD = 0.83, 95 % CI: 0.60 ∼ 1.00), 5 false alarms (FAR = 0.33, 95 % CI: 0.12 ∼ 0.60), and a TS of 0.59 with a 95 % CI of 0.35 ∼ 0.82. To further evaluate the robustness of these results and exclude potential overfitting, the model performance was compared against a stochastic baseline model through 1,000 Monte Carlo iterations (Fig. S7). The actual TS values for QJ (0.70) and WN (0.59) are significantly higher than the mean null TS (0.39 and 0.41, respectively), with p-values of p < 0.001 for QJ and p = 0.025 for WN. This statistical evidence confirms that the model's forecasting performance is statistically significant and demonstrably superior to random chance. Although the obtained FAR remains relatively high, which is a common challenge in earthquake anomaly detection that prioritises detection sensitivity, the TS provides a comprehensive metric that balances both POD and FAR. Compared with the internal single-component anomaly detection results from our model, the multicomponent synergistic alarm results exhibit higher TS values (Figs. 8, 9, 10). Their CIs are relatively narrower than those of individual components (Table S6), indicating that the synergistic approach effectively enhances the stability and reduces the uncertainty. The multicomponent synergy mitigates the effects of geochemical behavior differences among components, reduces environmental interference on individual ions/ion pairs, and consequently enhances the accuracy of the anomaly detection model. These results indicate the practical value of the multicomponent model for anomaly identification, though its practical application would benefit from integration with other geophysical, geodetic, and geological data to further reduce the false alarm burden. The results from the anomaly detection model and BCP analysis are mutually corroborative; however, the anomaly detection model exhibits superior sensitivity in processing nonlinear time series data. Taking QJ as an example, the model achieves POD values of 0.70 and 0.59 for Ca²⁺ and SO42- detection results, respectively (Figs. 5 and 8), representing significant improvements over the BCP analysis results (0.50 and 0.41). The model also can accurately detect subtle anomalies that the BCP analysis may miss.

Owing to variations in the geochemical behaviors of hydrochemical components, their response direction and magnitudes to earthquakes differ. The direction of anomalous variations (increase/decrease) is crucial for understanding the response mechanisms and improving event identification. A detailed statistical analysis of the anomaly directions for components across the two hot springs was performed (Table S7). The results reveal that certain pre-seismic anomalies manifest as intense bidirectional fluctuations. However, a definitive criterion to define a “dominant direction” for such oscillatory signals is lacking. Initial statistical results indicate that the frequency of anomalous increases (57 % ∼ 77 %) is higher than that of decreases for most ions (e.g., Na⁺, Cl⁻, SO42-). While these findings provide tentative support for the assumption that “increases” possess a stronger correlation with seismic events (Skelton et al., 2014, 2024; Yakupoğlu et al., 2025), the current sample size and statistical depth remain exploratory. Additionally, as observed regarding δD and δ18O, statistics indicate that these components exhibit a higher frequency of decreases (60 % ∼ 75 %), and change directions can vary across different components for the same earthquake event. Such divergence among different components suggests that a simplistic, unidirectional criterion might overlook vital precursor information. Based on these findings, the existing detection model is primarily oriented toward identifying sustained positive anomalies (increases), while also accounting for recovery phases following a decline (i.e., the increase following a decrease). Nevertheless, effectively incorporating these specific directional variations into earthquake forecasting remains a challenge at this stage.

Given these complexities and the lack of academic consensus on response mechanisms, this study evaluates the forecasting efficacy of individual components through the anomaly detection model to identify the effective indicators for the study area. A comparison of the TS values of each component's alarm results in QJ and WN (Fig. 10) reveals that in the two thermal springs of the study area, the TS values for Na⁺, Ca²⁺, Cl⁻, SO42-, δD, and δ18O detection (around 0.50) are relatively high. This observation suggests that these components can serve as sensitive indicators for strong earthquake forecasting in the study area. In general, QJ in the study area exhibits a more sensitive response to earthquakes. In addition, the anomalies are categorised into multiple consecutive anomalies and single anomalies (Figs. 8 and 9). This phenomenon is more pronounced in the stable isotope time series, likely because isotopic changes are more sensitive and tend to trigger multiple alarm signals before an earthquake. The higher sensitivity of δD and δ18O may be attributed to the fact that stable water isotopes are more conservative than hydrochemical ions, meaning they typically track water sources and mixing without being significantly affected by short-term water-rock interactions (Skelton et al., 2014, 2024; Su et al., 2025). The hydrochemical ions are easily altered by short-term chemical processes, such as dissolution or precipitation, cation exchange, adsorption or desorption, and redox reactions, making their signals more complex.

Among the earthquakes for which the earthquake preparation zone covers both thermal springs, only two earthquakes (represented by gray vertical bars in Figs. 8 and 9) fail to induce multicomponent anomalies before the earthquake. Earthquake E1 causes no synchronous anomalies at either spring, suggesting that E1 has a limited impact on regional tectonic activity. For E2 (epicentral distance >∼600 km), WN shows no alarm response, while QJ reacts effectively. This discrepancy is likely related to WN's location on the eastern boundary of the SYB, where stress accumulation mainly affects QJ, which is also located on the eastern border. The muted response in WN is likely attributed to the blocking effects of the RRF (Li et al., 2024; Shao et al., 2024). The similar abnormal response sensitivity of different springs to the same earthquake demonstrates regional-scale hydrochemical impacts from earthquake preparation and confirms the stable and reliable performance of the anomaly detection model.

In Figs. 8 and 9, the orange-red boxes represent model results of successfully predicted earthquakes, identified through synchronised anomalies in six or more hydrochemical components. The width of the boxes, which indicates the interval between the appearance of anomalies and earthquake occurrence, shows no clear correlation with magnitude or epicentral distance. This observation highlights the complex dynamic mechanisms and regional structural differences that are involved in the earthquake preparation process. The geochemical anomalies often arise from the combined effects of multiple mechanisms, such as fluid mixing following aquifer breaching or fresh mineral surface exposure during micro fracturing, resulting in increased concentrations of hydrochemical component (Thomas, 1988). Spatially, the number of hydrochemical components (Y) exhibiting synchronous anomalies correlates with earthquake magnitude and epicentral distance (Fig. 11). Earthquakes that induce synchronous anomalies in six or more hydrochemical components have epicentral distances of less than 150 km for earthquakes with magnitudes less than 6.0 (M<6.0), while this distance extends to approximately 400 km for earthquakes with magnitudes greater than or equal to 6.0 (M≥6.0). Among these significant events (Y≥6), a robust linear correlation exists between magnitude and distance (n=9, R=0.85, P=0.004). This correlation provides a quantitative basis for seismic risk assessment, allowing for the estimation of a potential earthquake's minimum magnitude or its maximum likely distance based on observed hydrochemical precursors. Although it is difficult to quantify the exact impact of magnitude and distance on the number of components exhibiting synchronous anomalies, as magnitude increases or distance decreases, the number of components with synchronous anomalies detected by the model tends to increase. Especially, the result reveals a distinct delineation in trend at the M=6.0 threshold. For earthquakes where M<6.0, the number of anomalous components is primarily sensitive to epicentral distance; shorter distances yield a higher count of synchronous anomalous components due to proximity to the stress-release center. For earthquakes where M≥6.0, magnitude appears to be a more dominant driver than distance, although limited data renders this a preliminary trend rather than a statistically significant finding. This is potentially due to high-energy release triggering regional-scale crustal stress perturbations that induce hydrochemical anomalies even at remote distances. The overall findings align with the positive correlation between the scale of earthquake energy release and the number of anomalies, as confirmed by the hydrochemical monitoring results (Li et al., 2022). Therefore, a significant relationship is present between the temporal variation of hydrochemical components and earthquakes in the study region. The number of components exhibiting synchronous anomalies can be used as an effective criterion for forecasting, with higher count of anomalous components generally corresponding to larger earthquake magnitudes or shorter epicentral distances.

Furthermore, this study reveals that hot springs closer to the epicentre tend to exhibit a greater number of components with synchronous anomalies during the same earthquake. Pre-earthquake hydrochemical anomalies generally manifest on a regional scale, which means different thermal springs can not only validate each other in terms of anomaly timing for forecasting purposes but also help identify the closest springs to the epicentre based on the number of synchronous anomalous components. This approach helps define potential earthquake preparation zones. According to this approach, a dense thermal spring monitoring network provides more opportunities for spatial earthquake forecasting.

This study focuses on evaluating the performance of anomaly detection models in predicting the timing of earthquakes with magnitudes ≥4. One potential cause of false alarms could be anomalous fluctuations in hydrochemical components triggered by seismic activities with magnitudes <4 in areas near thermal springs. Four days after the second synchronised false alarm involving five components (Fig. 8, grayish-blue boxes), an M2.6 earthquake occurred 3 km from QJ. This occurrence suggests that high-frequency false alarms may not solely result from non-seismic fluid anomalies, but could also reflect the model's limited ability to distinguish anomalies caused by microseisms. According to this finding, it is recommended to establish observation station networks and optimise algorithms to enable hierarchical alarm systems. Approximately 30 days after the last synchronized false alarms at the two thermal springs (Figs. 8 and 9), an M4.1 earthquake occurred outside the expected preparation zone. This highlights another key limitation: the current algorithm relies on an effective and widely used isotropic underground structure assumption, which restricts model's adaptability to account for directional tectonic complexities and spatial heterogeneity.

Despite these constraints, the multicomponent anomaly detection model developed here demonstrates robust performance, achieving a POD exceeding 83 % and a TS ranging from 59 % to 70 %. Within the field of geochemical forecasting, these results are comparable to the positive predictive values of 62 % to 85 % derived from the statistical analysis reported by Skelton et al. (2024) in Iceland. While that study utilized the oscillatory behaviors of δD and δ18O, and deuterium excess to trace groundwater mixing associated with crustal dilation and fracture mineralization before earthquakes, the current approach seeks to complement these isotopic insights by integrating them with multiple major ions. Synchronous fluctuations in ion concentrations can provide additional anomaly signals when isotopic shifts are subtle, reflecting a wider range of water-rock interaction processes. The model results also align with the 70 % POD and R-score of 0.6 described by Zhu et al. (2024) in the southeast coastal region of China, where ion ratios and PCA indicators were combined with the best-performing LOF algorithm to identify outliers. In contrast, the present model focuses on refining anomaly identification performance and classifying anomaly intensity by assessing the number of components with synchronous anomalies across multiple components. By leveraging the differentiated responses of various components to crustal stress, this synergistic analysis significantly enhances the identification of subtle pre-earthquake signals. Mechanistically, this multi-indicator integration provides a more objective reflection of pre-earthquake stress changes and fluid mixing processes, while offering certain advantages in reducing environmental interference, consistent with the tectonic stress drivers discussed by Yakupoğlu et al. (2025).

Although the specific parameters in this study are local, the underlying method framework could be transferable when specific selection criteria for monitoring sites are met. Monitoring stations should be situated in active tectonic regions, such as western Türkiye (Yakupoğlu et al., 2025), or even at major fault intersections, such as XJF and RRF intersection zone (Shao et al., 2024). Priority should be given to thermal springs with high discharge temperatures, as such temperatures typically indicate deep circulation and their hydrochemical components may carry deep-seated information while minimizing interference from meteoric water. Furthermore, the hydrogeochemical background of the spring should be clearly defined to distinguish its hydrogeochemical characteristics from shallow water bodies and to facilitate the monitoring of fluid mixing processes. Springs with a long observation duration, where historical earthquakes have induced data oscillations, should be prioritized. According to findings in Iceland (Skelton et al., 2024), these oscillatory signals originate from the physical coupling of crustal dilation and fracture mineralization, representing a reliable geochemical signature during the stress build-up phase. Preference is also given to springs with bubbling gases to expand gas-chemical observations. Additionally, sites that could incorporate specific crustal deformation-sensitive isotopic monitoring indicators are ideal, such as He isotopes in the Noto Peninsula (Kagoshima et al., 2025) or H-O-C isotopes in the Xianshuihe fault (Yu et al., 2026). In such cases, the implementation of differentiated weighting based on indicator sensitivity offers potential for further refining the precision of anomaly detection. The model can be practically applied to platforms such as the China Seismic Experimental Site (CSES) and the European Plate Observing System (EPOS), where the future of earthquake forecasting benefits from multiple-method synergy. For instance, cross-validating hydrochemical anomalies with b-value evaluation (Chang et al., 2025) or ionospheric TEC (Baselga, 2024) establishes a robust, multi-layered verification process. Within this hierarchical system, broad-scale geophysical tools effectively flag regional risks. Conversely, site-specific, multicomponent hydrochemical model provides the confirmation necessary for small-scale, short-term forecasting.