Document Type : Original Article

Authors

• A. H. Ramezani Feriz ¹
• Yousef Ramezani ²

¹ M.Sc. Student, Department of Water Engineering, University of Birjand, Birjand, Iran.

² Associate Professor, Department of Water Engineering, University of Birjand, Birjand, Iran.

Abstract

In arid and semi-arid regions, qanats are essential sources of groundwater, and predicting their discharge is crucial for effective water resource management. Machine learning models can be an efficient tool for simulating and forecasting qanat discharge. This study evaluates and compares the performance of three machine learning models—Random Forest, Gaussian Process Regression, and K-Star—in predicting qanat discharge in these regions. The primary aim was to assess the accuracy of these models in simulating qanat discharge behavior and predicting water resources under various climatic and geographical conditions. The data were collected from an arid region in South Khorasan Province. For model training, 80% of the data were used, and 20% were allocated for testing. The models were evaluated using standard statistical metrics, including R², NSE, RMSE, and KGE. The results showed that the Random Forest model had the best performance, explaining 94% of the data variance in the training phase and 96% in the testing phase. This model achieved NSE = 0.94, RMSE = 1.5 (l/s), and KGE = 0.88 in training, and NSE = 0.94, RMSE = 1.1 (l/s), and KGE = 0.85 in testing. In contrast, the Gaussian Process Regression model performed poorly, with R² = 0.01 and 0.14. This study demonstrates that Random Forest, due to its ability to process complex data and simulate nonlinear relationships, is a more effective predictive model for qanat discharge in arid regions.

Keywords

• Climate change
• Groundwater resources
• Machine learning
• Qanat
• Semi-arid and arid regions

Main Subjects

• ًQanat and aquifer