the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 17 Jul 2025
Decoding multicomponent hydrochemical anomalies: A synergy anomaly detection model for earthquake forecasting in active tectonic zone
Abstract. The intersection of the Xiaojiang Fault and the Red River Fault at the southeastern margin of the Tibetan Plateau experiences intense tectonic activity. At this intersection, frequent destructive earthquakes have induced hydrochemical variations in thermal springs. In this study, Bayesian change point analysis is applied, and a multicomponent synergy anomaly detection model is developed using five years of monitoring data (2019–2024) from two thermal springs in the region to achieve real-time forecasting of occurrence timing for M ≥ 4 earthquakes. Comprehensive analysis demonstrates that the anomaly detection model possesses reliable real-time anomaly detection capabilities. Tailored model parameters for specific hydrochemical components account for their differences in response characteristics to seismic activity. The model identifies Na+, Ca2+, Cl−, SO42−, δD, and δ18O as sensitive indicators for strong earthquake forecasting. The multicomponent synergy alarm mechanism for hydrochemistry overcomes the limitations of single-parameter methods, which significantly enhances the model’s overall performance in earthquake forecasting. The number of hydrochemical components with synchronous anomalies serves as a reliable criterion for determining alarm intensity, with higher intensity typically correlating with larger earthquake magnitudes or shorter epicentral distances.
- Preprint
(8514 KB) - Metadata XML
-
Supplement
(388 KB) - BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2025-2132', Anonymous Referee #1, 07 Aug 2025
The work presented by Shao et al. (2025) built a multi-component joint anomaly detection model by integrating continuous monitoring data of hydrochemical ions and hydrogen-oxygen isotopes with earthquake catalogs and applying Bayesian change point analysis. The results of the multicomponent synergy anomaly detection reveal a clear connection between hydrochemical variations and seismic activity, offer valuable insights for precursor identification in earthquake forecasting. This work is quite well written, though considerable issues require minor revision before publication.
General comments:
1. Abstract: It would be worth rephrasing to make the message clear and better reflect the key findings and the value of this study.2. Introduction: The Introduction is mostly well written. However, some minor issues should be state clearer and some relevant references are missing. Please see minor comments below.
3. Method and data:
a) Some results are presented and discussed in Section 3.3 and 3.4, which makes the structure unclear. The authors are suggested to reorganize some contents in 3.3 and 3.4, and move them into results and discuss them accordingly.
b) Also, some contents in this section are too lengthy. The authors are suggested to simplify some of the method (for example, the introduction of limitations of BCP method could go to later section or Supplementary information).4. Results and discussion: The results presented here are convincing; however, some lack in-depth discussion, causing some implications of the study to be obscured. It is recommended that the authors further discuss how some of these findings could be applied to other tectonically active regions around the world.
Specific comments:
Lines 26-27 Please specify how these isotopes changes before earthquake
Line 79 Please add relevant references for this statement
Line 98 what are the common machine learning algorithms
Line 172 Please provide the references for this equation and explain the meaning of each parameter
Line 186-187 Ambiguous. Consider rephrasing it to: '22 earthquakes with M ≥ 4.
Line 203 Please explain why you chose 𝑤=1. Have you conducted a sensitivity analysis?
Line 238 Please explain why a 15-day backward moving average is applied.
Line 250 Please cite references here about this definition.
Line 315 This paragraph is more like results and discussion (limitation). It is not appropriate to present here.
Line 577 Please describe this conclusion in more detail.Citation: https://doi.org/10.5194/egusphere-2025-2132-RC1 -
AC1: 'Reply on RC1', Ying Li, 22 Aug 2025
Publisher’s note: a supplement was added to this comment on 22 August 2025.
Dear Anonymous Referee #1,
We would like to thank you for reviewing our manuscript. Please find our responses to your comments in the attached supplement.
Best regards,
Ying Li on behalf of all co-authors -
AC2: 'Reply on RC1', Ying Li, 22 Aug 2025
Publisher’s note: this comment is a copy of AC1 and its content was therefore removed on 22 August 2025.
Citation: https://doi.org/10.5194/egusphere-2025-2132-AC2
-
AC1: 'Reply on RC1', Ying Li, 22 Aug 2025
-
RC2: 'Comment on egusphere-2025-2132', Anonymous Referee #2, 13 Aug 2025
Overview:
This study presents a multicomponent synergy anomaly detection model for real-time forecasting of M ≥ 4 earthquakes at the intersection of the Xiaojiang and Red River Faults on the southeastern margin of the Tibetan Plateau. Utilizing five years (2019-2024) of hydrochemical data from two thermal springs (Qujiang and Wana), the model uses Bayesian Change Point (BCP) analysis and an enhanced anomaly detection algorithm to identify precursors in Na⁺, Ca²⁺, Cl⁻, SO₄²⁻, δD, and δ¹⁸O. It achieves high accuracy, with a POD up to 0.95 and a TS up to 0.70, by leveraging synchronous anomalies, where a greater number of anomalies correlates with larger earthquake magnitudes or shorter epicentral distances, thereby enhancing forecasting reliability.
However, the study has notable limitations, including a restricted sample size (two springs, 24 events), which limits generalizability. The English prose occasionally lacks clarity, obscuring technical details. The manuscript’s structure, while generally coherent, includes redundant discussion sections. The methodology lacks sufficient clarity (e.g., elucidating the relationship between BCP analysis and the anomaly detection model). Mechanistic explanations for hydrochemical responses are underdeveloped, and figure captions lack precision, despite the figures’ informative nature. These shortcomings undermine the study’s scientific rigor, generalizability, and practical utility. Consequently, I recommend rejecting the manuscript pending major revisions, including broader validation, refined methodological assumptions, detailed mechanistic insights, improved figure clarity, and polished English to enhance accessibility and scientific impact.
Major Comments:
- The introduction offers an overview of earthquake precursor research, referencing relevant studies (e.g., Chen, 2009; Pritchard et al., 2020). However, it lacks a comprehensive review of competing methodologies, such as Bayesian Change Point (BCP) analysis or geophysical approaches, which would better contextualize the proposed hydrochemical method. Additionally, while it notes challenges in isolating seismic precursors from complex fluid monitoring data, it fails to clearly articulate specific gaps in existing hydrochemical anomaly detection research. For example, it mentions limitations of single-indicator methods but does not quantify their shortcomings (e.g., false positive rates) or directly compare them to the multicomponent approach. This limited scope results in an underdeveloped problem statement, rendering the study’s contribution ambiguous and weakening the justification for its methodological novelty.
- The earthquake selection method employs Dobrovolsky’s preparation zone radius formula, assuming isotropic subsurface structures. This oversimplification disregards the anisotropic complexities of faults and aquifers at the Xiaojiang-Red River Fault (XJF-RRF) intersection, potentially leading to inaccurate event selection. The manuscript neither justifies this assumption nor evaluates its sensitivity, which compromises the reliability of correlations between hydrochemical anomalies and seismic events.
- The 15-day moving average effectively reduces rainfall-induced noise, as evidenced by low cross-correlation coefficients (±0.2). However, the method lacks validation against other environmental factors, such as temperature or barometric pressure, or long-term trends that may influence hydrochemical signals. Without a control dataset or comparison with alternative filtering techniques (e.g., wavelet transforms), confidence in the denoising process is limited, undermining the robustness of the data preprocessing methodology.
- The optimization of parameters p1 and p3 is thoroughly described but lacks transparency regarding the selection of parameter ranges. The manuscript does not explore alternative optimization methods, such as grid search with cross-validation, nor does it assess the sensitivity of results to parameter variations beyond theTS presented in Figure 6. This omission raises concerns about the robustness and reproducibility of the anomaly detection model. Furthermore, using the entire dataset for parameter optimization introduces a significant risk of overfitting, which is not addressed, further undermining the model’s reliability.
- The study relies on data from only two thermal springs and a limited dataset (22 and 12 M ≥ 4 earthquakes, respectively), restricting the generalizability of findings to other tectonic settings or regional springs. The manuscript does not address how site-specific factors, such as lithology or fault geometry, might limit the model’s applicability, thus diminishing its broader scientific impact. Furthermore, the abstract claims that Na⁺, Ca²⁺, Cl⁻, SO₄²⁻, δD, and δ¹⁸O are sensitive indicators for earthquake forecasting, but this assertion is likely valid only for the studied springs. Additionally, the shorter time series for δD and δ¹⁸O in Figure 2 undermines their reliability as sensitive indicators.
- The discussion attributes hydrochemical anomalies to stress-induced fluid mixing and rock dissolution. However, it fails to propose specific geochemical pathways, such as mineral dissolution kinetics or isotopic fractionation, to explain the heightened sensitivity of certain components (e.g., δD, δ¹⁸O). This lack of mechanistic insight limits the study’s contribution to understanding earthquake preparation processes and weakens its theoretical foundation.
- The discussion asserts that the model outperforms single-component methods and compares favorably to Zhu et al. (2024). However, the reported false alarm rate (FAR, ~0.28–0.33) is relatively high for practical forecasting applications. The absence of statistical tests to confirm the model’s superiority, combined with an overemphasis on TS and POD without addressing the operational impact of false alarms, overstates the model’s practical utility.
- Line 239: The manuscript states that a 15-day moving average is applied to 3-day resolution hydrochemical data, implying only five measurements per 15-day period. The rationale for averaging over 15 days is not explained, which raises questions about the appropriateness of this window size for capturing tectonic signals while filtering noise. A justification or sensitivity analysis for this choice is needed to ensure methodological rigor.
- Figure 4: For the QJ station, the manuscript reports 22 M ≥ 4 earthquakes, yet Figure 4a displays only 11 earthquakes. This discrepancy is unacceptable and suggests incomplete data visualization. Similar inconsistencies appear in other subfigures, undermining the reliability of the visual representation of results and necessitating a comprehensive review of figure accuracy.
Mirror Comments:
- Line 25: The phrase “tailored model parameters for specific hydrochemical components” is imprecise. It should specify that parameters are optimized for individual components (e.g., Na⁺, δ¹⁸O) based on their distinct geochemical responses to seismic stress, as elaborated later (lines 428–431).
- Line 42: The term “physicochemical properties” is overly general. To align with the study’s focus on hydrochemical components, specify the properties primarily affected by crustal stress changes, such as ion concentrations and isotopic ratios.
- Line 78: The statement “thermal springs tend to exhibit high stability” may mislead readers, as stability is context-specific. Clarify that this refers to their low susceptibility to short-term environmental fluctuations (e.g., temperature) compared to other fluid systems.
- Line 96: The reference to Piersanti et al. (2016) is introduced abruptly without clarifying its relevance to hydrochemical data. Briefly note that the algorithm, originally developed for radon time series, was adapted for multicomponent hydrochemical analysis to enhance reader comprehension.
- Line 155: The description of the water quality analyzer (HQ40D, HACH, USA) and its measurement accuracies (0.1°C, 0.01 pH, 1 μS/cm) is tangential to the study’s primary focus. Omit or briefly summarize this detail to maintain emphasis on the hydrochemical data.
- Line 163: The list of analyzed ions is overly comprehensive. Specify only the ions used in the study to maintain focus and avoid extraneous detail.
- Line 172: The ion balance error equation is presented but not referenced or applied elsewhere in the study, rendering it disconnected from the analysis. Clarify its use or remove it to avoid confusion.
- Line 175: The claim that all earthquakes with M ≥ 4 are “destructive” is inaccurate, as destructiveness depends on depth, location, and infrastructure. Revise to reflect that M ≥ 4 earthquakes are the study’s focus without implying universal destructiveness.
- Line 188: The text states that QJ was within the preparation zones of 22 M ≥ 4 earthquakes, but Table S1 lists 24 events, creating a discrepancy. Clarify the correct number in the text to ensure consistency.
- Line 214: Figure 2 is referenced without specifying its content (e.g., time series of which components). Clarify that it illustrates hydrochemical component time series (e.g., Na⁺, Ca²⁺, Cl⁻) alongside rainfall and earthquake events for QJ spring to guide readers.
- Line 299: The evaluation metrics (FAR, POD, TS) are introduced in Figure 4’s caption but not defined until later (lines 370–380). Define them first before using them to avoid confusion for readers encountering the metrics early.
- Figures 7 and 8 effectively present anomaly detection results, but their captions and annotations lack sufficient detail. The figures omit scales or legends for posterior probabilities. Revise captions to include legends to enhance accessibility and enable independent verification of results.
- Line 520: The discussion references “multiple mechanisms” for anomalies (e.g., Thomas, 1988) without specifying examples, such as fracture dilation or fluid mixing. Briefly list one or two mechanisms to clarify the context.
- Line 574: The phrasing “tailored model parameters… account for their differences” is awkward and lacks clarity. Streamline for precision and readability.
Citation: https://doi.org/10.5194/egusphere-2025-2132-RC2
Data sets
Continuous monitoring data from thermal springs Weiye Shao https://doi.org/10.17632/xkd75cyfmb.1
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 232 | 53 | 13 | 298 | 18 | 6 | 5 |
- HTML: 232
- PDF: 53
- XML: 13
- Total: 298
- Supplement: 18
- BibTeX: 6
- EndNote: 5
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1