Modeling and Identification of Key Factors Affecting Groundwater Level in the Ravansar-Sanjabi Plain Using CMIP6 Climate Data

Document Type : Research Paper

Authors

• Kobra Soltani ¹
• Jafar Masoompour ²
• Seyed Ehsan Fatemi ³
• Maryam Hafezparast Mavadat ³

¹ Department of Geography, Faculty of Literature and Humanities, Razi University, Kermanshah, Iran

² Department of Geography, Razi University

³ Department of Water Engineering, Campus of Agriculture and Natural Resources, Razi University, Kermanshah, Iran

Abstract

Keywords

• CMIP6
• GWL prediction
• K-means clustering
• Key variables
• Random Forest algorithm

Main Subjects

• منابع آب

Introduction:

The depletion of groundwater resources, particularly in the arid and semi-arid regions of Iran, has emerged as a major environmental and socio-economic crisis. In recent decades, excessive groundwater extraction, population growth, agricultural expansion, and climate change have placed immense pressure on aquifers. This situation underscores the urgent need for innovative approaches to forecasting and sustainable groundwater management. Among these, machine learning techniques have proven effective due to their capacity to analyze large datasets, detect complex patterns, and generate accurate predictions. The Random Forest (RF) algorithm, in particular, is favored for its robustness against overfitting and its ability to model nonlinear relationships. This study integrates RF with K-means clustering, employing climate data from CMIP6 models to analyze trends and forecast future groundwater level (GWL) changes in the Ravansar-Sanjabi plain.

Materials and Methods:

The study area (Ravansar-Sanjabi plain), located in western Iran (Kermanshah Province), covers approximately 457 km², situated between latitudes 34°25′00″N and 34°46′50″N, and longitudes 46°34′50″E and 46°50′00″E. Geologically, the plain is predominantly limestone-based, fed by karst springs and shallow aquifers.

Input data included precipitation, minimum and maximum temperature, relative humidity, and soil moisture from 98 CMIP6 climate models, as well as groundwater level observations from 23 monitoring wells (1992–2014). K-means clustering, with the Elbow method, identified three optimal clusters: Kouzaran, TamTam, and Lori. After the formation of the clusters, the center of each cluster was designated as the representative (including the average GWL time series from the wells in that area), and the Random Forest algorithm was used for groundwater level modeling.

Simulations were conducted under three climate scenarios—SSP1-2.6, SSP2-4.5, and SSP5-8.5—across three future periods: near- future (2026–2050), mid-future (2051–2075), and far- future (2076–2100). Model performance was evaluated using R, NSE, MBE, SI, and NRMSE in both training and testing phases. Recursive Feature Elimination (RFE) was applied to reduce dimensionality and enhance model precision, while model validation used 5-fold cross-validation.

Results and Discussion:

K-means clustering revealed a clear spatial structure in GWL data, delineating three distinct regions. RF modeling showed the highest accuracy in the Kouzaran cluster (R = 0.92, NSE = 0.83), while the Lori cluster showed relatively lower performance (R = 0.78, NSE = 0.60).

Among the 98 CMIP6 models, NorESM2-MM consistently outperformed others across all clusters. Soil moisture was identified as the most influential variable in predicting GWL. Incorporating additional models (ACCESS-CM2, INM-CM5-0, MIROC6, and CESM2-WACCM) further improved model accuracy, reaching up to 98%.

Historical trend analysis (1992–2014), using the Mann-Kendall test, indicated significant declining trends in all clusters: 77 cm/year in Kouzaran, 68 cm/year in TamTam, and approximately 56 cm/year in Lori. Future projections showed that under SSP1-2.6, a slight GWL improvement is expected in the near future, followed by a significant decline in the mid-term (e.g., −19 cm/year in Kouzaran). Under SSP2-4.5, moderate short-term recovery is followed by continued declines through the mid- and far- future periods (e.g., −22 cm/year in Kouzaran). Under SSP5-8.5, the near- and mid- future periods show substantial GWL increases (e.g., +24 cm/year in Kouzaran), while far- future trends point to gradual declines. Seasonally, fall is projected to show the greatest GWL recovery (up to 4 meters in Kouzaran under SSP5-8.5, far- future), whereas spring may experience the steepest decline (e.g., −1.83 meters in TamTam under SSP2-4.5, mid- future).

Conclusion:

This study demonstrates that integrating RF and K-means algorithms with CMIP6 climate data provides an effective framework for accurately predicting future GWL trends. NorESM2-MM, with soil moisture as the dominant predictor, was identified as the most reliable model. RF achieved strong performance across all clusters, particularly in Kouzaran, which exhibited greater fluctuations in groundwater levels.

Despite potential near-future improvements, long-term projections—especially during spring—suggest substantial declines in GWL and increased aquifer instability. These findings highlight the urgent need for climate-adaptive groundwater management strategies in the Ravansar-Sanjabi plain.

Author Contributions

Conceptualization, Kobra Soltani, Seyed Ehsan Fatemi and Jafar Masoompour Samakosh; methodology, Seyed Ehsan Fatemi and Jafar Masoompour Samakosh; software, Seyed Ehsan Fatemi and Kobra Soltani; validation, Jafar Masoompour Samakosh and Maryam Hafezparast Mavadat; analysis, Kobra Soltani, Seyed Ehsan Fatemi, and Jafar Masoompour Samakosh; investigation, Kobra Soltani and Seyed Ehsan Fatemi; resources, Kobra Soltani and Maryam Hafezparast Mavadat; writing—original draft preparation, Kobra Soltani;writing—review and editing, Jafar Masoompour Samakosh, Seyed Ehsan Fatemi and Maryam Hafezparast.All authors have read and agreed to the published version of the manuscript.

Data Availability Statement

Data can be sent from the corresponding author by email upon request.

Acknowledgements

The authors would like to thank all participants of the present study.

Ethical considerations

The authors avoided data fabrication, falsification, plagiarism, and misconduct.

Conflict of interest

The author declares no conflict of interest.