Sustainable Soil Moisture Prediction in the Eslamabad Plain: A Machine Learning Approach to Climate Model Selection and Downscaling

Climate change and soil moisture reduction are significant challenges in water resource management, food security, and environmental sustainability. This study investigates the changes in the average monthly soil moisture values at the geographical location of the Eslamabad Synoptic Station in Kermanshah province utilizing MERRA2 data, historical (1991–2014), future scenarios data (SSP1-2.6, SSP2-4.5, and SSP5-8.5) from all CMIP6 models (r1i1p1f1 run) across three future periods: near (2026–2050), mid (2051–2075), and far (2076–2100). The Random Forest (RF) machine learning algorithm was utilized for downscaling and identifying the most effective soil moisture models from CMIP6 (achieving a total accuracy of 98%) by employing a Recursive Feature Elimination method in combination with k-fold validation. Out of 12 available models, 5 achieved an accuracy of 98% in predicting soil moisture. The MRI-ESM2-0 model was identified as the top model, with an accuracy of 87%. The other top 4 models collectively increased the final accuracy to 98%. RF evaluation confirmed its precise performance with metrics such as the Nash-Sutcliffe (0.98 for training and 0.78 for testing) and low error values (MAD = 0.01, MAE and MBE equal to zero). The results indicated a significant decrease in soil moisture during the future periods. A short reduction of 0.02 to 0.04 mm is projected for August and winter. In the mid and far future, a decrease of 0.03 to 0.05 mm under SSP5-8.5 is projected for most months. The annual trend indicates a decrease with rates of -0.0002 mm per year shortly under SSP2-4.5 and SSP5-8.5. At the same time, the SSP1-2.6 scenario predicts a trend of -0.0007 mm per year for the mid future. The results of this study are of particular environmental sustainability significance, as future soil moisture reduction could lead to increased water stress, decreased agricultural productivity, and pose a threat to regional food security.
 
Extended Abstract
1-Introduction
Soil moisture prediction is crucial for agricultural productivity, water management, and drought resilience, but it is challenging due to spatial and temporal variations. Climate change and extreme climate events significantly affect soil moisture, with data indicating a global trend of increasing aridity. In this context, machine learning algorithms, especially Random Forest (RF), have demonstrated high effectiveness in predicting soil moisture, particularly when combined with CMIP6 data. This study employs RF and the best climate models to predict soil moisture at the Eslamabad synoptic station, offering a reliable approach to improve prediction accuracy for potential applications in both Iran and globally.
 
2-Materials and Methods
Eslamabad city, located in Kermanshah province, has a semi-arid, cold climate with an altitude of approximately 1,300 meters above sea level. The area receives an average annual precipitation of 463 mm, and its average temperature is 13.8°C. This research utilized historical data, future CMIP6 scenarios, and MERRA2 reanalysis data to predict soil moisture using RF. Data from the Earth System Grid Federation (ESGF) were employed, along with socio-economic SSP scenarios and RF for downscaling. The models selected in this study, with 98% accuracy, were analyzed and categorized for prediction in the near, mid, and far future periods (2026-2100). Random Forest, a robust machine learning algorithm, was utilized with k-fold cross-validation for soil moisture prediction, demonstrating its efficacy. This method, resistant to overfitting, was evaluated using criteria such as Nash-Sutcliffe Efficiency (NSE), correlation coefficient (CC), and normalized root mean square error (NRMSE), providing more accurate results compared to MERRA2 and CMIP6 data. The RF algorithm, in combination with climate models and reanalysis data, allowed for better analysis of soil moisture dynamics in response to climate changes, proving to be a valuable tool for agricultural and environmental management in the region.
 
3- Results and Discussion
This study aimed to analyze the monthly mean soil moisture at the Eslamabad station using historical CMIP6 and MERRA2 data combined with the RF machine learning algorithm. The RF model successfully identified effective models for soil moisture prediction, with the MRI-ESM2-0 model showing the best performance at 87% accuracy. The combination of other top models improved the overall accuracy to 98%. Furthermore, changes in monthly soil moisture were examined under three scenarios: SSP1-2.6, SSP2-4.5, and SSP5-8.5, revealing significant reductions in soil moisture, particularly under high greenhouse gas emission scenarios such as SSP5-8.5. The results indicated a decrease in soil moisture during warmer months and even in fall and winter, especially in the near and far future. This reduction was attributed to increasing temperatures, reduced precipitation, and higher evaporation rates. These findings align with previous studies and highlight RF’s high accuracy in simulating and predicting climate variables. In contrast, the annual trend of soil moisture under the SSP1-2.6 scenario indicated a relatively stable or slight increase, while under SSP5-8.5, a more gradual decrease was observed. The results of this study are of particular environmental sustainability significance, as future soil moisture reduction could lead to increased water stress, decreased agricultural productivity, and pose a threat to regional food security. These results can aid in better decision-making for water resource management and climate change adaptation.
 
4- Conclusion
This study analyzed changes in the mean soil moisture at the Eslamabad synoptic station under three climate scenarios (SSP1-2.6, SSP2-4.5, and SSP5-8.5). Among the CMIP6 models, MRI-ESM2-0 was identified as the most accurate, with 87% precision. The addition of four other top models contributed only 11% to the overall accuracy, bringing it to 98%, underscoring the importance of selecting key models for climate predictions. The RF machine learning algorithm also demonstrated excellent performance, predicting soil moisture with high accuracy, as validated by a Nash-Sutcliffe Efficiency (NSE) of 0.98 during training and 0.78 during testing, along with MAD, 0.01 and MAE and MBE equal to zero. A futuristic analysis (2026-2100) revealed a significant decrease in soil moisture for most months. For example, in August of the near future, soil moisture is expected to decrease by 0.02 to 0.04 mm compared to the observed (0.087 mm), with a further decrease to 0.03 to 0.05 mm under the SSP5-8.5 scenario in the far future. The annual trend analysis also showed a decreasing soil moisture pattern under the SSP2-4.5 and SSP5-8.5 scenarios shortly, while under the SSP1-2.6 scenario, a similar pattern is observed in the mid-future. These findings emphasize the need for adaptive measures in water resource management and agricultural practices to mitigate the impacts of climate change in the Eslamabad region.