Enhanced Prediction of Persistent Earthquake-Induced Groundwater Level Changes with Advanced Feature Engineering and Machine Learning

Enhanced Prediction of Persistent Earthquake-Induced Groundwater Level Changes with Advanced Feature Engineering and Machine Learning | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Enhanced Prediction of Persistent Earthquake-Induced Groundwater Level Changes with Advanced Feature Engineering and Machine Learning Amirhossein Tahmouresi, Sogand Basirian, Amirhossein Javanshir, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6649664/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 17 Oct, 2025 Read the published version in Multimedia Systems → Version 1 posted 10 You are reading this latest preprint version Abstract Forecasting groundwater level fluctuations induced by seismic activity presents a considerable challenge due to the inherent complexity and pronounced non-linearity of the underlying processes. The limited availability of predictive variables further complicates this task, with key factors such as seismic shaking intensity, geological characteristics of dams, and shear wave velocity serving as primary indicators. To address the scarcity of predictive features and the intricate non-linear dependencies between input variables and groundwater level responses, we introduce an innovative fusion of feature engineering and machine learning. Our methodology is applied to a comprehensive regional-scale, multi-site, multi-earthquake dataset from New Zealand aquifers. Utilizing a filter-based supervised feature selection technique, we extract novel feature sets with strong correlations to groundwater level dynamics. Subsequently, we develop a random forest classification model to predict earthquake-induced groundwater level changes. The proposed approach significantly enhances both predictive accuracy and interpretability compared to conventional probabilistic models, offering a robust framework for improved seismic hydrogeological forecasting. Earthquake-induced groundwater changes Machine learning Feature engineering Random forest classification Seismic hydrogeological forecasting Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 1. Introduction Hydrological responses to seismic events have been documented for millennia, encompassing variations in water level, temperature, chemical composition, streamflow, and spring discharge properties [ 1 ]. Among these, earthquake-induced groundwater level (GWL) fluctuations are the most extensively recorded in historical datasets [ 2 ]. The analysis of seismic-induced GWL changes typically involves investigating either the response of individual wells to multiple earthquakes or the collective response of multiple wells to a single seismic event [ 3 ]. These fluctuations have profound implications for water resource management, infrastructure stability, and environmental sustainability. Extensive research has been conducted to develop predictive models for earthquake-induced GWL variations, employing both experimental and numerical approaches. For instance, Shi et al. [ 4 ] utilized statistical methodologies to examine correlations between seismic activity and GWL fluctuations. Their study analyzed co-seismic responses in consolidated rock formations across extensive spatial scales and compared responses to multiple earthquakes. The findings revealed that co-seismic water level responses exhibit significant spatial heterogeneity in both magnitude and direction. The study underscored the challenges in attributing observed GWL variations to specific geophysical mechanisms, emphasizing the necessity for more comprehensive datasets and advanced modeling techniques to fully elucidate the governing processes of seismic-induced groundwater fluctuations. Similarly, Liu et al. [ 5 ] observed of co-seismic and post-seismic earthquake-induced GWL changes. The high-frequency data of water level were collected by a dense network of multiple-well monitoring stations. Furthermore, they analyzed the corresponding temporal variations and spatial distributions of co-seismic changes. The study acknowledges potential constraints, such as the availability and precision of GWL data, the complexity of subsurface geological conditions, and the need for more comprehensive models to fully capture the dynamics of earthquake-induced GWL changes. Zhang et al. [ 6 ] analyzed seismic-induced GWL variations to infer changes in aquifer permeability following multiple seismic events, utilizing long-term groundwater monitoring data from two wells. Their findings indicated that water level and streamflow responses to earthquakes are highly dependent on the hydrogeological properties of the aquifers. The study provided valuable insights into aquifer permeability and structural characteristics, yet its applicability may be constrained by the specific karst aquifers under investigation. Moreover, reliance on long-term monitoring data limits the generalizability of the results to other geological settings lacking such extensive records. In another study, Tsai et al. [ 7 ] explored earthquake-induced GWL variations by integrating well data with seismic strong motion recordings. To address the limitations of poroelastic theory in explaining persistent GWL changes, they examined volumetric strain-induced variations caused by seismic shaking. Their hybrid modeling approach demonstrated superior predictive performance compared to conventional probabilistic models, achieving higher accuracy and enhanced interpretability. This methodological advancement offers a robust tool for researchers and practitioners engaged in seismic hydrogeological forecasting. Despite the robustness of traditional predictive frameworks, they often fail to fully capture the intricate relationships governing earthquake-induced GWL variations. The accuracy and reliability of numerical models are frequently constrained by their dependence on extensive datasets and the challenges of handling high-dimensional, non-linear interactions. Consequently, proper data preprocessing and transformation into a structured, machine-readable format are imperative. In recent years, artificial intelligence (AI) and machine learning (ML) have gained prominence in addressing complex, non-linear geoscientific problems [ 8 , 9 ]. Their data-driven approach, coupled with reduced reliance on extensive calibration compared to conventional process-based models, has proven particularly advantageous in hydrogeology. ML techniques have been employed for forecasting groundwater dynamics [ 10 ], assessing groundwater availability [ 11 ], and modeling groundwater quality [ 12 ]. These advancements underscore the pivotal role of integrating machine learning with feature engineering to enhance predictive accuracy in hydrogeological studies. To address the challenges posed by limited predictor availability and the non-linearity of relationships governing seismic-induced GWL changes, this study proposes an innovative methodology that synergizes feature engineering with machine learning to achieve enhanced predictive performance. The key contributions of this research are as follows: Developing novel features that exhibit strong correlations with earthquake-induced groundwater level responses. Employing machine learning techniques to systematically evaluate and optimize predictive models. Validating the proposed model on a comprehensive regional-scale, multi-site, multi-earthquake dataset from New Zealand aquifers. The remainder of this paper is structured as follows: Section 2 details the Methodology, Section 3 presents Model Performance, Section 4 discusses the findings, and Section 5 concludes with key insights and outlines directions for future research. 2. Groundwater Prediction Methodology 2.1. Method Overview The workflow for predicting earthquake-induced groundwater level (GWL) changes is illustrated in Fig. 1 . The process begins with data acquisition, where groundwater level data from New Zealand aquifers is collected. This is followed by data preprocessing, involving cleaning, normalization, and structuring to ensure consistency and quality. Feature engineering is then applied using the Autofeat Python library to extract and construct relevant features that enhance predictive performance. Next, a machine learning classification model is developed utilizing the PyCaret Python library. Finally, model evaluation and optimization are performed to assess predictive accuracy, fine-tune hyperparameters, and ensure robust performance. This structured approach leverages advanced machine learning techniques and feature engineering to improve the accuracy and interpretability of earthquake-induced GWL change predictions. 2.2 Data Acquisition and Preprocessing The dataset, sourced from Weaver et al. [ 13 ], comprises 2,500 groundwater level responses recorded across 495 wells in response to eleven major earthquakes that occurred in New Zealand between 2007 and 2016. In addition to groundwater level data, several key predictive variables, listed in Table 1 , have been measured to enhance the understanding of earthquake-induced groundwater fluctuations. This dataset is one of the most comprehensive and well-curated resources available on seismic-induced groundwater changes, making it particularly well-suited for machine learning applications. To ensure robust model training and evaluation, the dataset is divided into training (80%) and testing (20%) subsets. To mitigate disparities in feature scales and improve model performance, two data normalization techniques are applied. Numerical features are normalized to a [0,1] or [-1,1] range to facilitate model convergence, while features exhibiting Gaussian distributions undergo z-score normalization (standard scaling) to maintain statistical consistency. These preprocessing steps enhance comparability across different features, ultimately improving the model’s predictive accuracy. Table 1 Definition and characteristics of predictor and response variables. Category Detail Dataset Source Peak ground velocity Time Period 2007–2016 Number of Earthquakes 495 wells Total Responses 2500 water level responses Predictors Peak ground velocity, Peak ground acceleration, Seismic energy density, Average shear-wave velocity, Well depth, Water level response 2.3 Feature Engineering In our model, we utilized AutoFeat [ 14 ], a Python tool designed for automated feature engineering, selection, and classification. Selecting an optimal subset of features was critical to preventing overfitting while preserving essential information for accurate predictions. To achieve this, we employed a filter-based supervised feature selection technique, which is computationally more efficient than wrapper methods. The feature selection process implemented by AutoFeat is illustrated in Fig. 2 . This process involves iterative feature transformations and combinations to generate new predictive variables. Initially, only the seven original features were used, with no additional transformations. In the subsequent phase, mathematical operations such as logarithm, square root, and polynomial transformations were applied to the original features, resulting in 35 newly generated features. To refine the feature set, correlated and noisy features were systematically removed through filtering, ensuring that only the most relevant features were retained. At each iteration, AutoFeat conducted five rounds of supervised feature selection, further optimizing the predictive power of the model. This systematic approach enhances model efficiency by reducing redundancy while improving the interpretability and generalization of earthquake-induced groundwater level change predictions. 2.4 Random Forest We developed a Random Forest (RF) classification model to forecast changes in GWLs caused by earthquakes. RF is a widely used machine learning technique that functions by building several decision trees during the training phase and providing the most frequent class as the output [ 15 ]. This method integrates bagging (bootstrap aggregating) and random feature selection, resulting in enhanced accuracy and resilience, as shown in Fig. 3 . To evaluate our model's performance, we divided the original dataset into training and testing subsets, adhering to the supervised learning framework of random forests. Two critical hyperparameters were considered: the number of trees ( \(\:n\_estimators\) ) and the number of features to evaluate at each split ( \(\:max\_features\) ). When \(\:max\_features\:=\:n\_features\) (the total number of features), the subsample size matches the original dataset size, with samples drawn with replacement. We experimented with various values for these hyperparameters. Additionally, the \(\:random\_state\) parameter, which controls the sample bootstrapping randomness, was kept constant to ensure the reproducibility of the model. In our final model configuration, we selected \(\:n\_estimators\) = 100, a standard value for remote sensing research [ 16 ], set \(\:random\_state\) to 0, and \(\:max\_features\) to 7. The remaining hyperparameters, such as the maximum leaf size and minimum number of splits, were left at their scikit-learn default settings [ 17 ]. Algorithm 1 depicts the pseudo-code of the feature selection and groundwater level prediction using random forest. Algorithm 1 Pseudocode for Feature Selection and GWL Prediction using Random Forest Input : Dataset D with features F and target variable Y N = Number of bootstrap samples M = Total number of features m = Number of randomly selected features at each node (m < M) k = Next node to be split Output : Trained Random Forest Model (RF) Selected Features Steps : 1. Load and Preprocess Data a. Load dataset D (2,563 rows, 7 initial features) b. Handle missing values using median imputation c. Normalize continuous features using Min-Max scaling d. Split dataset into training ( \(\:D\_train\) ) and testing ( \(\:D\_test\) ) sets 2. Apply Feature Construction and Selection a. Use filter-based supervised feature selection b. Apply AutoFeat to generate new transformed features c. Select top K features based on importance ranking 3. Train Random Forest Model a. Initialize an ensemble of decision trees (100 trees) b. FOR each tree in RF (from 1 to 100) Bootstrap sample from \(\:D\_train\) Randomly select m features from M at each node Determine the best split using Gini impurity Recursively split the tree until leaf nodes are reached Store trained decision tree in RF c. Aggregate predictions from all trees using majority voting 4. Evaluate Model Performance a. Predict outcomes on \(\:D\_test\) using trained RF b. Compute evaluation metrics (precision, recall, and F1-score) c. Identify feature contributions from RF feature importance 5. Deploy and Interpret Results a. Store trained RF model for future predictions b. Generate feature importance visualization c. Store selected features for further optimization End 2.5 Performance Metrics The accuracies of the above machine-learning models are evaluated using precision, recall, and F1-score. Precision measures the proportion of true positive predictions among all positive predictions made by the model. A high precision indicates that the model is making few false positive predictions. The precision is computed as follow: $$\:\frac{tp}{\:(tp\:+\:fp)}$$ 1 where \(\:tp\) is the number of true positives and \(\:fp\) the number of false positives. Recall (Sensitivity) measures the proportion of true positive predictions among all actual positive instances in the dataset. A high recall indicates that the model is capturing most of the positive instances in the dataset. The recall is computed as follow: $$\:\frac{tp}{\:(tp\:+\:fn)}$$ 2 where \(\:tp\) is the number of true positives and \(\:fn\) the number of false negatives. F1-score is the harmonic mean of precision and recall. It provides a single metric that combines both precision and recall, balancing the trade-off between the two. It is calculated as: $$\:F1-score=\frac{2\times\:Precision\times\:Recall}{Precision+Recall}$$ 3 Summary of model development and evaluation for Earthquake-Induced Groundwater Level Change Prediction are provide in Table 2 . Table 2 Summary for model development and evaluation Study area Duration of input data Type of classifier Programming language Performance evaluation Feature scaling method New Zealand 2007–2016 Random Forest Python Precision, recall, F1-score and support Filter-based supervised feature selection technique 3. Performance Results This work employed a Random Forest classifier to predict groundwater level (GWL) changes induced by earthquakes using various seismic and aquifer features. The key objectives were to develop novel feature sets that effectively capture the relationship between seismic activities and GWL responses and apply a machine learning model to systematically evaluate and predict the behavior of these changes. The following discussion provides a detailed analysis of the model’s performance based on the generated results. 3.1 Model Performance Overview The classification report (Fig. 4 ) indicates balanced performance across both classes (0: no change, 1: change). For class 0, the model achieved a precision of 0.882, recall of 0.903, and F1-score of 0.892, while for class 1, it recorded a precision of 0.661, recall of 0.610, and F1-score of 0.634, with macro-average precision, recall, and F1-score of 0.77, 0.76, and 0.76, respectively. The confusion matrix (Fig. 5 ) reflects the model’s prediction performance, where Class 0 (true negatives) is well predicted with 530 true negatives and 57 false positives. However, for Class 1 (true positives), there are 111 correct predictions and 71 false negatives. This supports the finding from the classification report, emphasizing that Class 1 (earthquake-induced changes) is more challenging to classify correctly, potentially due to class imbalance in the dataset. 3.2 Feature Importance and Model Interpretation The feature importance plot (Fig. 6 ) ranks the contributions of the features used in the model. Features F1 and sSED stand out as the most influential, with derived seismic features contributing more than traditional seismic measurements like PGA (Peak Ground Acceleration) and epicentral distance. This suggests that the developed features, which were designed to capture subtle interactions between seismic activity and groundwater levels, play a crucial role in predicting GWL changes. These results validate the project's approach of leveraging feature engineering to enhance model performance. The SHAP chart (Fig. 7 ) further corroborates this, showing the impact of individual features on the model’s output. Higher values of sSED and F1 tend to push predictions toward Class 1 (earthquake-induced GWL change), whereas lower values push predictions toward Class 0. This insight into feature behavior confirms the effectiveness of the engineered features in capturing earthquake-induced changes and supports the decision to prioritize feature development in this study. 3.3 Decision Boundary and Model Challenges The boundary plot (Fig. 8 ) visualizes the distribution of data points in a two-dimensional feature space (Feature One vs. Feature Two), with class 0 (blue) and class 1 (yellow) points. It reveals a dense clustering of class 0 points at lower values of Feature One (0 to 4) and Feature Two (-4 to 2), while class 1 points are more sparsely distributed at higher values of Feature One (4 to 14), indicating a non-linear decision boundary. This distribution suggests that the random forest model effectively captures the separation between classes, though the overlap in the feature space contributes to the observed false negatives and positives. 3.4 ROC, Precision-Recall, and Gain Curves The ROC curve (Fig. 9 a) with an AUC of 0.89 for both classes indicates that the model performs reasonably well in classifying both classes, though it still falls short of optimal performance, especially for Class 1. The micro- and macro-average AUCs of 0.92 suggest that on average, the model is good at distinguishing between the classes. The precision-recall curve (Fig. 9 b) reveals an average precision of 0.68, which indicates the model’s difficulty in maintaining high precision as recall increases. The cumulative gains curve (Fig. 9 c) highlights that the model ranks Class 1 instances well. The KS statistic plot (Fig. 10), with a KS value of 0.999, shows significant class separation at a threshold of 0.43. 4. Discussion Our proposed work significantly improved the shortcomings of previous works by leveraging machine learning and feature engineering techniques to better capture the non-linear relationships between seismic and aquifer variables. Below is a comparison and discussion of how this work addressed specific gaps in the existing literature and improves the predictive accuracy of GWL changes. Our findings extend the insights from Weaver et al. [ 13 ], who developed a probabilistic model using binary logistic regression with random effects (LRRE) to assess aquifer susceptibility to persistent groundwater level changes across 495 monitoring wells in response to 11 New Zealand earthquakes between 2008 and 2017. 4.1 Advancing Feature Engineering Weaver et al. [ 13 ] work relies heavily on standard seismic parameters such as PGA, earthquake magnitude, and distance from the epicenter to predict GWL changes. While effective to some extent, these features are limited in their ability to fully capture complex interactions in groundwater systems. To address this issue, this work introduced novel feature sets, such as seismic energy dissipation and derived aquifer-specific features, which offer deeper insights into how seismic forces interact with aquifers. By utilizing machine learning algorithms, specifically Random Forest classifiers, our proposed model systematically evaluated a broader range of features. This improvement addresses the issue of a limited number of predictors in the existing model by introducing a more comprehensive set of high-signal features, resulting in better model performance and more accurate predictions​. 4.2 Handling Non-Linearity The probabilistic model in [ 13 ] assumes linear or near-linear relationships between seismic activity and GWL changes. This assumption limits the accuracy of predictions, especially when non-linear interactions between variables exist, as often seen in complex groundwater systems. By employing a Random Forest model, our approach captured non-linear dependencies between features and GWL changes, enabling the model to identify intricate patterns that would be missed by linear models. This approach directly addresses the non-linearity issue, as your model can better represent the complexity of the groundwater system’s response to seismic forces​. 4.3 Model Performance and Class Imbalance The work in [ 13 ] reports an overall acceptable model performance but struggles with predicting instances of significant GWL changes (Class 1), likely due to class imbalance in the data. This is evident from lower precision, recall, and F1 scores for Class 1. While our study also faced challenges with class imbalance, the use of Random Forest classifiers and advanced feature engineering has improved the model's ability to predict Class 1. This improvement is reflected in higher precision and recall for this class compared to the baseline probabilistic approach, as shown by the results in the classification report and confusion matrix​. 4.4 Interpretability and Feature Importance The lack of interpretability in previous studies hinders understanding of which factors most influence the model’s predictions. Our model enhanced interpretability through the use of feature importance rankings and SHAP (SHapley Additive exPlanations) analysis. These techniques not only reveal which features (e.g., seismic energy dissipation and aquifer-specific parameters) contribute most to the model’s performance but also explain how these features push predictions toward different classes. This makes our model more transparent and valuable for decision-making in water resource management and seismic hazard assessment. 4.5 Accuracy and Model Optimization While our model has similar overall ROC performance with work in [ 13 ], the precision-recall curve provides a better perspective of how the model balances precision and recall, especially for Class 1. This work suggested that by adjusting the model threshold and focusing on precision-recall trade-offs, the predictive accuracy for earthquake-induced GWL changes can be significantly improved​. Moreover, this study shows significant improvement in model optimization through the use of learning curves and validation curves. The gains curve demonstrates better model ranking, and the learning curve analysis suggests that our model generalizes well to new data without overfitting. This is an important step in ensuring that the model can be deployed for real-time applications​. In overall, our proposed work substantially enhanced the predictive accuracy and interpretability of earthquake-induced GWL changes compared to the baseline probabilistic model. By incorporating feature engineering, non-linear models, and systematic evaluation tools, our approach addresses key limitations in the existing model. As a result, this work offers a more robust and reliable solution for predicting GWL changes, which is critical for earthquake preparedness and groundwater management. The use of advanced tools like SHAP, Random Forest, and precision-recall curves further strengthens the model's applicability for real-world scenarios. 5. Future work While the Random Forest classifier offers reasonably good performance, it faces challenges in correctly predicting GWL changes due to earthquakes. This is largely attributable to class imbalance, and the inherent complexity of the data. The results validate the importance of the novel feature engineering efforts, which significantly contributed to the model’s predictive power. However, further efforts are needed to refine the model, improve its generalizability, and increase its accuracy, especially in predicting earthquake-induced groundwater changes (Class 1). Moreover, the feature importance and SHAP analysis (Figs. 6 and 7 ) provide actionable insights for future work. Since features like sSED and F1 contribute significantly to model performance, future iterations should focus on refining these and potentially creating new features that capture more of the underlying dynamics between seismic events and groundwater responses. Conclusion In this study, we developed a novel approach that combines feature construction and machine learning to enhance the prediction of earthquake-induced groundwater level (GWL) changes. By applying a filter-based supervised feature selection technique, we identified a refined set of predictors with strong correlations to GWL responses. The implementation of a random forest classification model significantly enhanced the forecasting accuracy and interpretability of GWL changes, outperforming traditional probabilistic models with a robust accuracy of 0.825, strong discriminative power, and effective feature utilization. Our findings highlight the potential of machine learning-based approaches in addressing the challenges posed by the complex and non-linear nature of earthquake-induced GWL variations. Future research could focus on expanding the dataset, incorporating additional geophysical predictors, and testing alternative machine learning frameworks to further enhance prediction reliability. Declarations Conflicts of Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Author Contribution Conceptualization: A.T., M.R., and A.J., ; methodology: A.T., M.R., and A.J., ; investigation: A.T., M.R., and A.J.,; implementation: A.T.; validation: A.T.; writing—original draft preparation: A.T.; writing—review and editing: A.T., M.R., A.J., and Y.C.; supervision: M.R., and A.J.,; project administration: A.T. References Xiang, Y., Sun, X., & Gao, X. (2019). Different coseismic groundwater level changes in two adjacent wells in a fault-intersected aquifer system. Journal of Hydrology , 578 , 124123. Rashidi Gooya, H., Katibeh, H., & Maleki, A. (2024). Forecasting groundwater fluctuations caused by earthquakes using fuzzy logic and AHP Method: A case study from Iran. Earth Science Informatics , 17 (3), 2143-2158. Rohde, M. M., Biswas, T., Housman, I. W., Campbell, L. S., Klausmeyer, K. R., & Howard, J. K. (2021). A machine learning approach to predict groundwater levels in California reveals ecosystems at risk. Frontiers in Earth Science , 9 , 784499. Shi, Z., Wang, G., Manga, M., & Wang, C. Y. (2015). Mechanism of co-seismic water level change following four great earthquakes–insights from co-seismic responses throughout the Chinese mainland. Earth and Planetary Science Letters , 430 , 66-74. Liu, C. Y., Chia, Y., Chung, P. Y., Lee, T. P., & Chiu, Y. C. (2023). Temporal variation and spatial distribution of groundwater level changes induced by large earthquakes. Water , 15 (2), 357. Zhang, H., Shi, Z., Wang, G., Yan, X., Liu, C., Sun, X., ... & Wen, D. (2021). Different sensitivities of earthquake‐induced water level and hydrogeological property variations in two aquifer systems. Water Resources Research , 57 (5), e2020WR028217. Tsai, C. C., Huang, L. Y., & Chen, C. J. (2023). Earthquake-induced persistent and instantaneous groundwater variations caused by volumetric strain of soil in Taiwan from 1999 to 2020. Soil Dynamics and Earthquake Engineering , 164 , 107586. Zhao, T., Wang, S., Ouyang, C., Chen, M., Liu, C., Zhang, J., ... & Wang, L. (2024). Artificial intelligence for geoscience: Progress, challenges and perspectives. The Innovation . Takaew, P., Xia, J. C., & Doucet, L. S. (2024). Machine learning and tectonic setting determination: Bridging the gap between Earth scientists and data scientists. Geoscience Frontiers , 15 (1), 101726. Gholami, V. C. K. W., Chau, K. W., Fadaee, F., Torkaman, J., & Ghaffari, A. (2015). Modeling of groundwater level fluctuations using dendrochronology in alluvial aquifers. Journal of hydrology , 529 , 1060-1069. Hussein, E. A., Thron, C., Ghaziasgar, M., Bagula, A., & Vaccari, M. (2020). Groundwater prediction using machine-learning tools. Algorithms , 13 (11), 300. Koch, J., Stisen, S., Refsgaard, J. C., Ernstsen, V., Jakobsen, P. R., & Højberg, A. L. (2019). Modeling depth of the redox interface at high resolution at national scale using random forest and residual gaussian simulation. Water Resources Research , 55 (2), 1451-1469. Weaver, K. C., Arnold, R., Holden, C., Townend, J., & Cox, S. C. (2020). A probabilistic model of aquifer susceptibility to earthquake‐induced groundwater‐level changes. Bulletin of the Seismological Society of America , 110 (3), 1046-1063. Horn, F., Pack, R., & Rieger, M. (2020). The autofeat python library for automated feature engineering and selection. In Machine Learning and Knowledge Discovery in Databases: International Workshops of ECML PKDD 2019, Würzburg, Germany, September 16–20, 2019, Proceedings, Part I (pp. 111-120). Springer International Publishing. Parmar, A., Katariya, R., & Patel, V. (2019). A review on random forest: An ensemble classifier. In International conference on intelligent data communication technologies and internet of things (ICICI) 2018 (pp. 758-763). Springer International Publishing. Belgiu, M., & Drăguţ, L. (2016). Random forest in remote sensing: A review of applications and future directions. ISPRS journal of photogrammetry and remote sensing , 114 , 24-31. scikit‐learn developers. (2019). 3.2.4.3.2. sklearn.ensemble. RandomForestRegressor—scikit‐learn 0.23.1 documentation. https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestRegressor.html Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 17 Oct, 2025 Read the published version in Multimedia Systems → Version 1 posted Editorial decision: Revision requested 20 Jul, 2025 Reviews received at journal 19 Jul, 2025 Reviewers agreed at journal 09 Jul, 2025 Reviewers agreed at journal 07 Jul, 2025 Reviews received at journal 03 Jul, 2025 Reviewers agreed at journal 20 Jun, 2025 Reviewers invited by journal 17 Jun, 2025 Editor assigned by journal 16 Jun, 2025 Submission checks completed at journal 14 May, 2025 First submitted to journal 12 May, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6649664","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":472328782,"identity":"1e9bfdb7-2143-4e25-9d9e-0c3a7f142a6b","order_by":0,"name":"Amirhossein Tahmouresi","email":"","orcid":"","institution":"Kerman University of Medical Science","correspondingAuthor":false,"prefix":"","firstName":"Amirhossein","middleName":"","lastName":"Tahmouresi","suffix":""},{"id":472328783,"identity":"c5f98069-58dc-4310-9cc4-ba15562e0bcf","order_by":1,"name":"Sogand Basirian","email":"","orcid":"","institution":"Azad University","correspondingAuthor":false,"prefix":"","firstName":"Sogand","middleName":"","lastName":"Basirian","suffix":""},{"id":472328784,"identity":"23cfd61f-c3bf-4ed9-bf97-129fdf9fd5ad","order_by":2,"name":"Amirhossein Javanshir","email":"","orcid":"","institution":"Deakin University","correspondingAuthor":false,"prefix":"","firstName":"Amirhossein","middleName":"","lastName":"Javanshir","suffix":""},{"id":472328785,"identity":"282fe9bc-5fca-4bb7-b22d-67d9a446be3f","order_by":3,"name":"YoungJin Cha","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA/UlEQVRIiWNgGAWjYBACgwNgyi6Bgb0BxJAAI2K0JCcw8IBZEhIEtUiCzWY4mMAgkcAAsYaQFn72HrOHXxgO5PHPfGO64ecOizp+6eYHDD9qcGth4zljbizDcKBY4naO2c3eMxISknOOGTD2HMOjRSLHTFqC4UBiA1DLDd42CQmDGwkGzAxsRGiZf/OM2c2/YC3pH5gZ/uHXIvmB4WDihhs8ZrchtuQYMDO24fPLsTJpBoPkxI1n0spuy7ZJSM6cc6bgYG8fHi3szdskf1TYJc47fnjbzbdtdfz80u0bH/z4hlsLCDDzGKCJHMCvgYGB8QchFaNgFIyCUTCyAQC4KlD+sGbbAQAAAABJRU5ErkJggg==","orcid":"","institution":"University of Manitoba","correspondingAuthor":true,"prefix":"","firstName":"YoungJin","middleName":"","lastName":"Cha","suffix":""}],"badges":[],"createdAt":"2025-05-12 21:08:12","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6649664/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6649664/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s00530-025-02043-6","type":"published","date":"2025-10-17T15:58:16+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":84989656,"identity":"f576e2f7-8b7f-4bca-b7a5-271ffbf1441b","added_by":"auto","created_at":"2025-06-19 15:02:29","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":53638,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart showing the implementation process.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6649664/v1/34bcad23e6b1e9bddc5e497f.png"},{"id":84989369,"identity":"efd0d578-c3bd-4bf3-96d6-7ff128d71869","added_by":"auto","created_at":"2025-06-19 14:54:29","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":106828,"visible":true,"origin":"","legend":"\u003cp\u003eFeature selection process.\u003c/p\u003e","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6649664/v1/ed468a3fe978d77a8c877c14.jpeg"},{"id":84990586,"identity":"41ccc9c4-e65e-46d0-8d2e-76c416570c8d","added_by":"auto","created_at":"2025-06-19 15:10:29","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":302418,"visible":true,"origin":"","legend":"\u003cp\u003eRandom forest process.\u003c/p\u003e","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6649664/v1/2ea342bf80ce75b73cad2123.jpeg"},{"id":84989662,"identity":"23068083-d794-4139-9872-c6eaa83b3567","added_by":"auto","created_at":"2025-06-19 15:02:29","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":16956,"visible":true,"origin":"","legend":"\u003cp\u003eclassification report\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6649664/v1/2d4ac54b590fab4042679174.png"},{"id":84990587,"identity":"822c217f-821c-4260-9942-e07bec896c8f","added_by":"auto","created_at":"2025-06-19 15:10:29","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":10804,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrix\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-6649664/v1/0e83e90463a94690ba8d99b8.png"},{"id":84989660,"identity":"3d94dc49-f716-40a2-9000-b8835d9c5508","added_by":"auto","created_at":"2025-06-19 15:02:29","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":24774,"visible":true,"origin":"","legend":"\u003cp\u003eFeature importance\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-6649664/v1/75bf11d9559e8f6971bf27fa.png"},{"id":84990589,"identity":"9b957042-495d-429c-8141-899bd50478f8","added_by":"auto","created_at":"2025-06-19 15:10:29","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":69065,"visible":true,"origin":"","legend":"\u003cp\u003eSHAP chart\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-6649664/v1/6d063fcea9b7946074ca6271.png"},{"id":84989664,"identity":"4828dc35-5812-4695-b209-bb04fc146a45","added_by":"auto","created_at":"2025-06-19 15:02:30","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":45341,"visible":true,"origin":"","legend":"\u003cp\u003eBoundary plot\u003c/p\u003e","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-6649664/v1/12c3c8b06384a3756889fe12.png"},{"id":84989395,"identity":"ed41c0c9-3b0c-4ffd-ac27-c88508e5df40","added_by":"auto","created_at":"2025-06-19 14:54:30","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":129955,"visible":true,"origin":"","legend":"\u003cp\u003eROC curve, precision-recall curve and The KS statistic plot\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-6649664/v1/2a21af29106e19bbdc0e89aa.png"},{"id":93956071,"identity":"9081e534-9822-41f5-b39f-8b252997473d","added_by":"auto","created_at":"2025-10-20 16:10:06","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1542769,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6649664/v1/340686c1-00f4-418f-b7e6-cd7ee61a9d8e.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Enhanced Prediction of Persistent Earthquake-Induced Groundwater Level Changes with Advanced Feature Engineering and Machine Learning","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eHydrological responses to seismic events have been documented for millennia, encompassing variations in water level, temperature, chemical composition, streamflow, and spring discharge properties [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Among these, earthquake-induced groundwater level (GWL) fluctuations are the most extensively recorded in historical datasets [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. The analysis of seismic-induced GWL changes typically involves investigating either the response of individual wells to multiple earthquakes or the collective response of multiple wells to a single seismic event [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. These fluctuations have profound implications for water resource management, infrastructure stability, and environmental sustainability.\u003c/p\u003e \u003cp\u003eExtensive research has been conducted to develop predictive models for earthquake-induced GWL variations, employing both experimental and numerical approaches. For instance, Shi et al. [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e] utilized statistical methodologies to examine correlations between seismic activity and GWL fluctuations. Their study analyzed co-seismic responses in consolidated rock formations across extensive spatial scales and compared responses to multiple earthquakes. The findings revealed that co-seismic water level responses exhibit significant spatial heterogeneity in both magnitude and direction. The study underscored the challenges in attributing observed GWL variations to specific geophysical mechanisms, emphasizing the necessity for more comprehensive datasets and advanced modeling techniques to fully elucidate the governing processes of seismic-induced groundwater fluctuations.\u003c/p\u003e \u003cp\u003eSimilarly, Liu et al. [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e] observed of co-seismic and post-seismic earthquake-induced GWL changes. The high-frequency data of water level were collected by a dense network of multiple-well monitoring stations. Furthermore, they analyzed the corresponding temporal variations and spatial distributions of co-seismic changes. The study acknowledges potential constraints, such as the availability and precision of GWL data, the complexity of subsurface geological conditions, and the need for more comprehensive models to fully capture the dynamics of earthquake-induced GWL changes.\u003c/p\u003e \u003cp\u003eZhang et al. [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e] analyzed seismic-induced GWL variations to infer changes in aquifer permeability following multiple seismic events, utilizing long-term groundwater monitoring data from two wells. Their findings indicated that water level and streamflow responses to earthquakes are highly dependent on the hydrogeological properties of the aquifers. The study provided valuable insights into aquifer permeability and structural characteristics, yet its applicability may be constrained by the specific karst aquifers under investigation. Moreover, reliance on long-term monitoring data limits the generalizability of the results to other geological settings lacking such extensive records.\u003c/p\u003e \u003cp\u003eIn another study, Tsai et al. [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e] explored earthquake-induced GWL variations by integrating well data with seismic strong motion recordings. To address the limitations of poroelastic theory in explaining persistent GWL changes, they examined volumetric strain-induced variations caused by seismic shaking. Their hybrid modeling approach demonstrated superior predictive performance compared to conventional probabilistic models, achieving higher accuracy and enhanced interpretability. This methodological advancement offers a robust tool for researchers and practitioners engaged in seismic hydrogeological forecasting. Despite the robustness of traditional predictive frameworks, they often fail to fully capture the intricate relationships governing earthquake-induced GWL variations. The accuracy and reliability of numerical models are frequently constrained by their dependence on extensive datasets and the challenges of handling high-dimensional, non-linear interactions. Consequently, proper data preprocessing and transformation into a structured, machine-readable format are imperative.\u003c/p\u003e \u003cp\u003eIn recent years, artificial intelligence (AI) and machine learning (ML) have gained prominence in addressing complex, non-linear geoscientific problems [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Their data-driven approach, coupled with reduced reliance on extensive calibration compared to conventional process-based models, has proven particularly advantageous in hydrogeology. ML techniques have been employed for forecasting groundwater dynamics [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], assessing groundwater availability [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e], and modeling groundwater quality [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. These advancements underscore the pivotal role of integrating machine learning with feature engineering to enhance predictive accuracy in hydrogeological studies.\u003c/p\u003e \u003cp\u003eTo address the challenges posed by limited predictor availability and the non-linearity of relationships governing seismic-induced GWL changes, this study proposes an innovative methodology that synergizes feature engineering with machine learning to achieve enhanced predictive performance. The key contributions of this research are as follows:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eDeveloping novel features that exhibit strong correlations with earthquake-induced groundwater level responses.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eEmploying machine learning techniques to systematically evaluate and optimize predictive models.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eValidating the proposed model on a comprehensive regional-scale, multi-site, multi-earthquake dataset from New Zealand aquifers.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eThe remainder of this paper is structured as follows: Section \u003cspan refid=\"Sec2\" class=\"InternalRef\"\u003e2\u003c/span\u003e details the Methodology, Section \u003cspan refid=\"Sec8\" class=\"InternalRef\"\u003e3\u003c/span\u003e presents Model Performance, Section \u003cspan refid=\"Sec13\" class=\"InternalRef\"\u003e4\u003c/span\u003e discusses the findings, and Section \u003cspan refid=\"Sec19\" class=\"InternalRef\"\u003e5\u003c/span\u003e concludes with key insights and outlines directions for future research.\u003c/p\u003e"},{"header":"2. Groundwater Prediction Methodology","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Method Overview\u003c/h2\u003e \u003cp\u003eThe workflow for predicting earthquake-induced groundwater level (GWL) changes is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The process begins with data acquisition, where groundwater level data from New Zealand aquifers is collected. This is followed by data preprocessing, involving cleaning, normalization, and structuring to ensure consistency and quality. Feature engineering is then applied using the Autofeat Python library to extract and construct relevant features that enhance predictive performance. Next, a machine learning classification model is developed utilizing the PyCaret Python library. Finally, model evaluation and optimization are performed to assess predictive accuracy, fine-tune hyperparameters, and ensure robust performance. This structured approach leverages advanced machine learning techniques and feature engineering to improve the accuracy and interpretability of earthquake-induced GWL change predictions.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Data Acquisition and Preprocessing\u003c/h2\u003e \u003cp\u003eThe dataset, sourced from Weaver et al. [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], comprises 2,500 groundwater level responses recorded across 495 wells in response to eleven major earthquakes that occurred in New Zealand between 2007 and 2016. In addition to groundwater level data, several key predictive variables, listed in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, have been measured to enhance the understanding of earthquake-induced groundwater fluctuations. This dataset is one of the most comprehensive and well-curated resources available on seismic-induced groundwater changes, making it particularly well-suited for machine learning applications. To ensure robust model training and evaluation, the dataset is divided into training (80%) and testing (20%) subsets. To mitigate disparities in feature scales and improve model performance, two data normalization techniques are applied. Numerical features are normalized to a [0,1] or [-1,1] range to facilitate model convergence, while features exhibiting Gaussian distributions undergo z-score normalization (standard scaling) to maintain statistical consistency. These preprocessing steps enhance comparability across different features, ultimately improving the model\u0026rsquo;s predictive accuracy.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDefinition and characteristics of predictor and response variables.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCategory\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDetail\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDataset Source\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePeak ground velocity\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTime Period\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2007\u0026ndash;2016\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNumber of Earthquakes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e495 wells\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal Responses\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2500 water level responses\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePredictors\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePeak ground velocity, Peak ground acceleration, Seismic energy density, Average shear-wave velocity, Well depth, Water level response\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Feature Engineering\u003c/h2\u003e \u003cp\u003eIn our model, we utilized AutoFeat [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], a Python tool designed for automated feature engineering, selection, and classification. Selecting an optimal subset of features was critical to preventing overfitting while preserving essential information for accurate predictions. To achieve this, we employed a filter-based supervised feature selection technique, which is computationally more efficient than wrapper methods. The feature selection process implemented by AutoFeat is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. This process involves iterative feature transformations and combinations to generate new predictive variables. Initially, only the seven original features were used, with no additional transformations. In the subsequent phase, mathematical operations such as logarithm, square root, and polynomial transformations were applied to the original features, resulting in 35 newly generated features. To refine the feature set, correlated and noisy features were systematically removed through filtering, ensuring that only the most relevant features were retained. At each iteration, AutoFeat conducted five rounds of supervised feature selection, further optimizing the predictive power of the model. This systematic approach enhances model efficiency by reducing redundancy while improving the interpretability and generalization of earthquake-induced groundwater level change predictions.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Random Forest\u003c/h2\u003e \u003cp\u003eWe developed a Random Forest (RF) classification model to forecast changes in GWLs caused by earthquakes. RF is a widely used machine learning technique that functions by building several decision trees during the training phase and providing the most frequent class as the output [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. This method integrates bagging (bootstrap aggregating) and random feature selection, resulting in enhanced accuracy and resilience, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eTo evaluate our model's performance, we divided the original dataset into training and testing subsets, adhering to the supervised learning framework of random forests. Two critical hyperparameters were considered: the number of trees (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:n\\_estimators\\)\u003c/span\u003e\u003c/span\u003e) and the number of features to evaluate at each split (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:max\\_features\\)\u003c/span\u003e\u003c/span\u003e). When \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:max\\_features\\:=\\:n\\_features\\)\u003c/span\u003e\u003c/span\u003e (the total number of features), the subsample size matches the original dataset size, with samples drawn with replacement. We experimented with various values for these hyperparameters. Additionally, the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:random\\_state\\)\u003c/span\u003e\u003c/span\u003e parameter, which controls the sample bootstrapping randomness, was kept constant to ensure the reproducibility of the model. In our final model configuration, we selected \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:n\\_estimators\\)\u003c/span\u003e\u003c/span\u003e = 100, a standard value for remote sensing research [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e], set \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:random\\_state\\)\u003c/span\u003e\u003c/span\u003e to 0, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:max\\_features\\)\u003c/span\u003e\u003c/span\u003e to 7. The remaining hyperparameters, such as the maximum leaf size and minimum number of splits, were left at their scikit-learn default settings [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Algorithm 1 depicts the pseudo-code of the feature selection and groundwater level prediction using random forest.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e \u003ccolgroup cols=\"1\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAlgorithm 1 Pseudocode for Feature Selection and GWL Prediction using Random Forest\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eInput\u003c/b\u003e: \u003cem\u003eDataset D with features F and target variable Y\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003eN\u0026thinsp;=\u0026thinsp;Number of bootstrap samples\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003eM\u0026thinsp;=\u0026thinsp;Total number of features\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003em\u0026thinsp;=\u0026thinsp;Number of randomly selected features at each node (m\u0026thinsp;\u0026lt;\u0026thinsp;M)\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003ek\u0026thinsp;=\u0026thinsp;Next node to be split\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eOutput\u003c/b\u003e:\u003c/p\u003e \u003cp\u003e\u003cem\u003eTrained Random Forest Model (RF)\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003eSelected Features\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eSteps\u003c/b\u003e:\u003c/p\u003e \u003cp\u003e1. Load and Preprocess Data\u003c/p\u003e \u003cp\u003ea. \u003cem\u003eLoad dataset D (2,563 rows, 7 initial features)\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003eb. Handle missing values using median imputation\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003ec. Normalize\u003c/em\u003e continuous \u003cem\u003efeatures using Min-Max scaling\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003ed. Split dataset into training (\u003c/em\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:D\\_train\\)\u003c/span\u003e\u003c/span\u003e\u003cem\u003e) and testing (\u003c/em\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:D\\_test\\)\u003c/span\u003e\u003c/span\u003e\u003cem\u003e) sets\u003c/em\u003e\u003c/p\u003e \u003cp\u003e2. Apply Feature Construction and Selection\u003c/p\u003e \u003cp\u003ea. \u003cem\u003eUse filter-based supervised feature selection\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003eb. Apply AutoFeat to generate new transformed features\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003ec. Select top K features based on importance ranking\u003c/em\u003e\u003c/p\u003e \u003cp\u003e3. Train Random Forest Model\u003c/p\u003e \u003cp\u003e\u003cem\u003ea. Initialize an ensemble of decision trees (100 trees)\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003eb. FOR each tree in RF (from 1 to 100)\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003eBootstrap sample from\u003c/em\u003e \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:D\\_train\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003eRandomly select m features from M at each node\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003eDetermine the best split using Gini impurity\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003eRecursively split the tree until leaf nodes are reached\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003eStore trained decision tree in RF\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003ec. Aggregate predictions from all trees using majority voting\u003c/em\u003e\u003c/p\u003e \u003cp\u003e4. Evaluate Model Performance\u003c/p\u003e \u003cp\u003e\u003cem\u003ea. Predict outcomes on\u003c/em\u003e \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:D\\_test\\)\u003c/span\u003e\u003c/span\u003e \u003cem\u003eusing trained RF\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003eb. Compute evaluation metrics (precision, recall, and F1-score)\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003ec. Identify feature contributions from RF feature importance\u003c/em\u003e\u003c/p\u003e \u003cp\u003e5. Deploy and Interpret Results\u003c/p\u003e \u003cp\u003e\u003cem\u003ea. Store trained RF model for future predictions\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003eb. Generate feature importance visualization\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cem\u003ec. Store selected features for further optimization\u003c/em\u003e\u003c/p\u003e \u003cp\u003e\u003cb\u003eEnd\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Performance Metrics\u003c/h2\u003e \u003cp\u003eThe accuracies of the above machine-learning models are evaluated using precision, recall, and F1-score. Precision measures the proportion of true positive predictions among all positive predictions made by the model. A high precision indicates that the model is making few false positive predictions. The precision is computed as follow:\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:\\frac{tp}{\\:(tp\\:+\\:fp)}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:tp\\)\u003c/span\u003e\u003c/span\u003e is the number of true positives and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:fp\\)\u003c/span\u003e\u003c/span\u003e the number of false positives.\u003c/p\u003e \u003cp\u003eRecall (Sensitivity) measures the proportion of true positive predictions among all actual positive instances in the dataset. A high recall indicates that the model is capturing most of the positive instances in the dataset. The recall is computed as follow:\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:\\frac{tp}{\\:(tp\\:+\\:fn)}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:tp\\)\u003c/span\u003e\u003c/span\u003e is the number of true positives and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:fn\\)\u003c/span\u003e\u003c/span\u003e the number of false negatives.\u003c/p\u003e \u003cp\u003eF1-score is the harmonic mean of precision and recall. It provides a single metric that combines both precision and recall, balancing the trade-off between the two. It is calculated as:\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$\\:F1-score=\\frac{2\\times\\:Precision\\times\\:Recall}{Precision+Recall}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eSummary of model development and evaluation for Earthquake-Induced Groundwater Level Change Prediction are provide in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSummary for model development and evaluation\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStudy area\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDuration of input data\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eType of classifier\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eProgramming\u003c/p\u003e \u003cp\u003elanguage\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePerformance evaluation\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eFeature scaling method\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNew Zealand\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2007\u0026ndash;2016\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePython\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePrecision, recall, F1-score and support\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eFilter-based supervised feature selection technique\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"3. Performance Results","content":"\u003cp\u003eThis work employed a Random Forest classifier to predict groundwater level (GWL) changes induced by earthquakes using various seismic and aquifer features. The key objectives were to develop novel feature sets that effectively capture the relationship between seismic activities and GWL responses and apply a machine learning model to systematically evaluate and predict the behavior of these changes. The following discussion provides a detailed analysis of the model\u0026rsquo;s performance based on the generated results.\u003c/p\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Model Performance Overview\u003c/h2\u003e \u003cp\u003eThe classification report (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e) indicates balanced performance across both classes (0: no change, 1: change). For class 0, the model achieved a precision of 0.882, recall of 0.903, and F1-score of 0.892, while for class 1, it recorded a precision of 0.661, recall of 0.610, and F1-score of 0.634, with macro-average precision, recall, and F1-score of 0.77, 0.76, and 0.76, respectively.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe confusion matrix (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e) reflects the model\u0026rsquo;s prediction performance, where Class 0 (true negatives) is well predicted with 530 true negatives and 57 false positives. However, for Class 1 (true positives), there are 111 correct predictions and 71 false negatives. This supports the finding from the classification report, emphasizing that Class 1 (earthquake-induced changes) is more challenging to classify correctly, potentially due to class imbalance in the dataset.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Feature Importance and Model Interpretation\u003c/h2\u003e \u003cp\u003eThe feature importance plot (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e) ranks the contributions of the features used in the model. Features F1 and sSED stand out as the most influential, with derived seismic features contributing more than traditional seismic measurements like PGA (Peak Ground Acceleration) and epicentral distance. This suggests that the developed features, which were designed to capture subtle interactions between seismic activity and groundwater levels, play a crucial role in predicting GWL changes. These results validate the project's approach of leveraging feature engineering to enhance model performance.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe SHAP chart (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e) further corroborates this, showing the impact of individual features on the model\u0026rsquo;s output. Higher values of sSED and F1 tend to push predictions toward Class 1 (earthquake-induced GWL change), whereas lower values push predictions toward Class 0. This insight into feature behavior confirms the effectiveness of the engineered features in capturing earthquake-induced changes and supports the decision to prioritize feature development in this study.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Decision Boundary and Model Challenges\u003c/h2\u003e \u003cp\u003eThe boundary plot (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e) visualizes the distribution of data points in a two-dimensional feature space (Feature One vs. Feature Two), with class 0 (blue) and class 1 (yellow) points. It reveals a dense clustering of class 0 points at lower values of Feature One (0 to 4) and Feature Two (-4 to 2), while class 1 points are more sparsely distributed at higher values of Feature One (4 to 14), indicating a non-linear decision boundary. This distribution suggests that the random forest model effectively captures the separation between classes, though the overlap in the feature space contributes to the observed false negatives and positives.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.4 ROC, Precision-Recall, and Gain Curves\u003c/h2\u003e \u003cp\u003eThe ROC curve (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003ea) with an AUC of 0.89 for both classes indicates that the model performs reasonably well in classifying both classes, though it still falls short of optimal performance, especially for Class 1. The micro- and macro-average AUCs of 0.92 suggest that on average, the model is good at distinguishing between the classes. The precision-recall curve (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003eb) reveals an average precision of 0.68, which indicates the model\u0026rsquo;s difficulty in maintaining high precision as recall increases. The cumulative gains curve (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003ec) highlights that the model ranks Class 1 instances well. The KS statistic plot (Fig.\u0026nbsp;10), with a KS value of 0.999, shows significant class separation at a threshold of 0.43.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eOur proposed work significantly improved the shortcomings of previous works by leveraging machine learning and feature engineering techniques to better capture the non-linear relationships between seismic and aquifer variables. Below is a comparison and discussion of how this work addressed specific gaps in the existing literature and improves the predictive accuracy of GWL changes. Our findings extend the insights from Weaver et al. [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], who developed a probabilistic model using binary logistic regression with random effects (LRRE) to assess aquifer susceptibility to persistent groundwater level changes across 495 monitoring wells in response to 11 New Zealand earthquakes between 2008 and 2017.\u003c/p\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Advancing Feature Engineering\u003c/h2\u003e \u003cp\u003eWeaver et al. [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] work relies heavily on standard seismic parameters such as PGA, earthquake magnitude, and distance from the epicenter to predict GWL changes. While effective to some extent, these features are limited in their ability to fully capture complex interactions in groundwater systems. To address this issue, this work introduced novel feature sets, such as seismic energy dissipation and derived aquifer-specific features, which offer deeper insights into how seismic forces interact with aquifers. By utilizing machine learning algorithms, specifically Random Forest classifiers, our proposed model systematically evaluated a broader range of features. This improvement addresses the issue of a limited number of predictors in the existing model by introducing a more comprehensive set of high-signal features, resulting in better model performance and more accurate predictions​.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Handling Non-Linearity\u003c/h2\u003e \u003cp\u003eThe probabilistic model in [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] assumes linear or near-linear relationships between seismic activity and GWL changes. This assumption limits the accuracy of predictions, especially when non-linear interactions between variables exist, as often seen in complex groundwater systems. By employing a Random Forest model, our approach captured non-linear dependencies between features and GWL changes, enabling the model to identify intricate patterns that would be missed by linear models. This approach directly addresses the non-linearity issue, as your model can better represent the complexity of the groundwater system\u0026rsquo;s response to seismic forces​.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Model Performance and Class Imbalance\u003c/h2\u003e \u003cp\u003eThe work in [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] reports an overall acceptable model performance but struggles with predicting instances of significant GWL changes (Class 1), likely due to class imbalance in the data. This is evident from lower precision, recall, and F1 scores for Class 1. While our study also faced challenges with class imbalance, the use of Random Forest classifiers and advanced feature engineering has improved the model's ability to predict Class 1. This improvement is reflected in higher precision and recall for this class compared to the baseline probabilistic approach, as shown by the results in the classification report and confusion matrix​.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Interpretability and Feature Importance\u003c/h2\u003e \u003cp\u003eThe lack of interpretability in previous studies hinders understanding of which factors most influence the model\u0026rsquo;s predictions. Our model enhanced interpretability through the use of feature importance rankings and SHAP (SHapley Additive exPlanations) analysis. These techniques not only reveal which features (e.g., seismic energy dissipation and aquifer-specific parameters) contribute most to the model\u0026rsquo;s performance but also explain how these features push predictions toward different classes. This makes our model more transparent and valuable for decision-making in water resource management and seismic hazard assessment.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e4.5 Accuracy and Model Optimization\u003c/h2\u003e \u003cp\u003eWhile our model has similar overall ROC performance with work in [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], the precision-recall curve provides a better perspective of how the model balances precision and recall, especially for Class 1. This work suggested that by adjusting the model threshold and focusing on precision-recall trade-offs, the predictive accuracy for earthquake-induced GWL changes can be significantly improved​. Moreover, this study shows significant improvement in model optimization through the use of learning curves and validation curves. The gains curve demonstrates better model ranking, and the learning curve analysis suggests that our model generalizes well to new data without overfitting. This is an important step in ensuring that the model can be deployed for real-time applications​.\u003c/p\u003e \u003cp\u003eIn overall, our proposed work substantially enhanced the predictive accuracy and interpretability of earthquake-induced GWL changes compared to the baseline probabilistic model. By incorporating feature engineering, non-linear models, and systematic evaluation tools, our approach addresses key limitations in the existing model. As a result, this work offers a more robust and reliable solution for predicting GWL changes, which is critical for earthquake preparedness and groundwater management. The use of advanced tools like SHAP, Random Forest, and precision-recall curves further strengthens the model's applicability for real-world scenarios.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Future work","content":"\u003cp\u003eWhile the Random Forest classifier offers reasonably good performance, it faces challenges in correctly predicting GWL changes due to earthquakes. This is largely attributable to class imbalance, and the inherent complexity of the data. The results validate the importance of the novel feature engineering efforts, which significantly contributed to the model’s predictive power. However, further efforts are needed to refine the model, improve its generalizability, and increase its accuracy, especially in predicting earthquake-induced groundwater changes (Class 1). Moreover, the feature importance and SHAP analysis (Figs.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e and \u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e) provide actionable insights for future work. Since features like sSED and F1 contribute significantly to model performance, future iterations should focus on refining these and potentially creating new features that capture more of the underlying dynamics between seismic events and groundwater responses.\u003c/p\u003e "},{"header":"Conclusion","content":"\u003cp\u003eIn this study, we developed a novel approach that combines feature construction and machine learning to enhance the prediction of earthquake-induced groundwater level (GWL) changes. By applying a filter-based supervised feature selection technique, we identified a refined set of predictors with strong correlations to GWL responses. The implementation of a random forest classification model significantly enhanced the forecasting accuracy and interpretability of GWL changes, outperforming traditional probabilistic models with a robust accuracy of 0.825, strong discriminative power, and effective feature utilization. Our findings highlight the potential of machine learning-based approaches in addressing the challenges posed by the complex and non-linear nature of earthquake-induced GWL variations. Future research could focus on expanding the dataset, incorporating additional geophysical predictors, and testing alternative machine learning frameworks to further enhance prediction reliability.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eConflicts of Interest\u003c/h2\u003e \u003cp\u003eThe authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eConceptualization: A.T., M.R., and A.J., ; methodology: A.T., M.R., and A.J., ; investigation: A.T., M.R., and A.J.,; implementation: A.T.; validation: A.T.; writing\u0026mdash;original draft preparation: A.T.; writing\u0026mdash;review and editing: A.T., M.R., A.J., and Y.C.; supervision: M.R., and A.J.,; project administration: A.T.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eXiang, Y., Sun, X., \u0026amp; Gao, X. (2019). Different coseismic groundwater level changes in two adjacent wells in a fault-intersected aquifer system. \u003cem\u003eJournal of Hydrology\u003c/em\u003e, \u003cem\u003e578\u003c/em\u003e, 124123.\u003c/li\u003e\n\u003cli\u003eRashidi Gooya, H., Katibeh, H., \u0026amp; Maleki, A. (2024). Forecasting groundwater fluctuations caused by earthquakes using fuzzy logic and AHP Method: A case study from Iran. \u003cem\u003eEarth Science Informatics\u003c/em\u003e, \u003cem\u003e17\u003c/em\u003e(3), 2143-2158.\u003c/li\u003e\n\u003cli\u003eRohde, M. M., Biswas, T., Housman, I. W., Campbell, L. S., Klausmeyer, K. R., \u0026amp; Howard, J. K. (2021). A machine learning approach to predict groundwater levels in California reveals ecosystems at risk. \u003cem\u003eFrontiers in Earth Science\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e, 784499.\u003c/li\u003e\n\u003cli\u003eShi, Z., Wang, G., Manga, M., \u0026amp; Wang, C. Y. (2015). Mechanism of co-seismic water level change following four great earthquakes\u0026ndash;insights from co-seismic responses throughout the Chinese mainland. \u003cem\u003eEarth and Planetary Science Letters\u003c/em\u003e, \u003cem\u003e430\u003c/em\u003e, 66-74.\u003c/li\u003e\n\u003cli\u003eLiu, C. Y., Chia, Y., Chung, P. Y., Lee, T. P., \u0026amp; Chiu, Y. C. (2023). Temporal variation and spatial distribution of groundwater level changes induced by large earthquakes. \u003cem\u003eWater\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e(2), 357.\u003c/li\u003e\n\u003cli\u003eZhang, H., Shi, Z., Wang, G., Yan, X., Liu, C., Sun, X., ... \u0026amp; Wen, D. (2021). Different sensitivities of earthquake‐induced water level and hydrogeological property variations in two aquifer systems. \u003cem\u003eWater Resources Research\u003c/em\u003e, \u003cem\u003e57\u003c/em\u003e(5), e2020WR028217.\u003c/li\u003e\n\u003cli\u003eTsai, C. C., Huang, L. Y., \u0026amp; Chen, C. J. (2023). Earthquake-induced persistent and instantaneous groundwater variations caused by volumetric strain of soil in Taiwan from 1999 to 2020. \u003cem\u003eSoil Dynamics and Earthquake Engineering\u003c/em\u003e, \u003cem\u003e164\u003c/em\u003e, 107586.\u003c/li\u003e\n\u003cli\u003eZhao, T., Wang, S., Ouyang, C., Chen, M., Liu, C., Zhang, J., ... \u0026amp; Wang, L. (2024). Artificial intelligence for geoscience: Progress, challenges and perspectives. \u003cem\u003eThe Innovation\u003c/em\u003e.\u003c/li\u003e\n\u003cli\u003eTakaew, P., Xia, J. C., \u0026amp; Doucet, L. S. (2024). Machine learning and tectonic setting determination: Bridging the gap between Earth scientists and data scientists. \u003cem\u003eGeoscience Frontiers\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e(1), 101726.\u003c/li\u003e\n\u003cli\u003eGholami, V. C. K. W., Chau, K. W., Fadaee, F., Torkaman, J., \u0026amp; Ghaffari, A. (2015). Modeling of groundwater level fluctuations using dendrochronology in alluvial aquifers. \u003cem\u003eJournal of hydrology\u003c/em\u003e, \u003cem\u003e529\u003c/em\u003e, 1060-1069.\u003c/li\u003e\n\u003cli\u003eHussein, E. A., Thron, C., Ghaziasgar, M., Bagula, A., \u0026amp; Vaccari, M. (2020). Groundwater prediction using machine-learning tools. \u003cem\u003eAlgorithms\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(11), 300.\u003c/li\u003e\n\u003cli\u003eKoch, J., Stisen, S., Refsgaard, J. C., Ernstsen, V., Jakobsen, P. R., \u0026amp; H\u0026oslash;jberg, A. L. (2019). Modeling depth of the redox interface at high resolution at national scale using random forest and residual gaussian simulation. \u003cem\u003eWater Resources Research\u003c/em\u003e, \u003cem\u003e55\u003c/em\u003e(2), 1451-1469.\u003c/li\u003e\n\u003cli\u003eWeaver, K. C., Arnold, R., Holden, C., Townend, J., \u0026amp; Cox, S. C. (2020). A probabilistic model of aquifer susceptibility to earthquake‐induced groundwater‐level changes. \u003cem\u003eBulletin of the Seismological Society of America\u003c/em\u003e, \u003cem\u003e110\u003c/em\u003e(3), 1046-1063.\u003c/li\u003e\n\u003cli\u003eHorn, F., Pack, R., \u0026amp; Rieger, M. (2020). The autofeat python library for automated feature engineering and selection. In \u003cem\u003eMachine Learning and Knowledge Discovery in Databases: International Workshops of ECML PKDD 2019, W\u0026uuml;rzburg, Germany, September 16\u0026ndash;20, 2019, Proceedings, Part I\u003c/em\u003e (pp. 111-120). Springer International Publishing.\u003c/li\u003e\n\u003cli\u003eParmar, A., Katariya, R., \u0026amp; Patel, V. (2019). A review on random forest: An ensemble classifier. In \u003cem\u003eInternational conference on intelligent data communication technologies and internet of things (ICICI) 2018\u003c/em\u003e (pp. 758-763). Springer International Publishing.\u003c/li\u003e\n\u003cli\u003eBelgiu, M., \u0026amp; Drăguţ, L. (2016). Random forest in remote sensing: A review of applications and future directions. \u003cem\u003eISPRS journal of photogrammetry and remote sensing\u003c/em\u003e, \u003cem\u003e114\u003c/em\u003e, 24-31.\u003c/li\u003e\n\u003cli\u003escikit‐learn developers. (2019). 3.2.4.3.2. sklearn.ensemble. RandomForestRegressor\u0026mdash;scikit‐learn 0.23.1 documentation. https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestRegressor.html\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"multimedia-systems","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"mmsj","sideBox":"Learn more about [Multimedia Systems](http://link.springer.com/journal/530)","snPcode":"530","submissionUrl":"https://submission.nature.com/new-submission/530/3","title":"Multimedia Systems","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Earthquake-induced groundwater changes, Machine learning, Feature engineering, Random forest classification, Seismic hydrogeological forecasting","lastPublishedDoi":"10.21203/rs.3.rs-6649664/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6649664/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eForecasting groundwater level fluctuations induced by seismic activity presents a considerable challenge due to the inherent complexity and pronounced non-linearity of the underlying processes. The limited availability of predictive variables further complicates this task, with key factors such as seismic shaking intensity, geological characteristics of dams, and shear wave velocity serving as primary indicators. To address the scarcity of predictive features and the intricate non-linear dependencies between input variables and groundwater level responses, we introduce an innovative fusion of feature engineering and machine learning. Our methodology is applied to a comprehensive regional-scale, multi-site, multi-earthquake dataset from New Zealand aquifers. Utilizing a filter-based supervised feature selection technique, we extract novel feature sets with strong correlations to groundwater level dynamics. Subsequently, we develop a random forest classification model to predict earthquake-induced groundwater level changes. The proposed approach significantly enhances both predictive accuracy and interpretability compared to conventional probabilistic models, offering a robust framework for improved seismic hydrogeological forecasting.\u003c/p\u003e","manuscriptTitle":"Enhanced Prediction of Persistent Earthquake-Induced Groundwater Level Changes with Advanced Feature Engineering and Machine Learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-19 14:54:24","doi":"10.21203/rs.3.rs-6649664/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-07-20T12:58:05+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-07-19T07:22:00+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"82993598612514871930052304978851400392","date":"2025-07-09T07:47:42+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"144314572683158538737364180450034270641","date":"2025-07-07T08:45:07+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-07-03T08:54:10+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"163529535229366307839014423022132800942","date":"2025-06-20T08:40:01+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-06-17T05:40:31+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-06-17T02:06:25+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-05-14T06:06:10+00:00","index":"","fulltext":""},{"type":"submitted","content":"Multimedia Systems","date":"2025-05-12T20:53:07+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"multimedia-systems","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"mmsj","sideBox":"Learn more about [Multimedia Systems](http://link.springer.com/journal/530)","snPcode":"530","submissionUrl":"https://submission.nature.com/new-submission/530/3","title":"Multimedia Systems","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"8427ae21-3e92-4f6d-9d8a-a1104b99c259","owner":[],"postedDate":"June 19th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-10-20T16:04:02+00:00","versionOfRecord":{"articleIdentity":"rs-6649664","link":"https://doi.org/10.1007/s00530-025-02043-6","journal":{"identity":"multimedia-systems","isVorOnly":false,"title":"Multimedia Systems"},"publishedOn":"2025-10-17 15:58:16","publishedOnDateReadable":"October 17th, 2025"},"versionCreatedAt":"2025-06-19 14:54:24","video":"","vorDoi":"10.1007/s00530-025-02043-6","vorDoiUrl":"https://doi.org/10.1007/s00530-025-02043-6","workflowStages":[]},"version":"v1","identity":"rs-6649664","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6649664","identity":"rs-6649664","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}