Document Type : Original Article

Authors

• Sepide Zeraati Neyshabouri ¹
• Mohamad Fouladi Nasrabad ²
• mahna javaheri ³

¹ Department of Water science and Engineering, University of Birjand, Iran

² Department of Water Science and Engineering, Birjand University, Birjand, Iran

³ MSc Student, Department of Water Engineering, Faculty of Agriculture, University of Birjand.

Abstract

Arid and semi-arid regions of Iran, particularly the Qaen Plain in South Khorasan Province, face severe water resource limitations. Qanats, as traditional sustainable water supply systems with over 3,000 years of history, play a vital role in local agriculture and livelihoods. However, accurately estimating their discharge is challenging due to climate change impacts and data limitations. This study employs the Random Forest (RF) algorithm—a powerful machine learning method—to estimate qanat discharges in the Qaen Plain using 12 years of data (2007–2018 CE). Input variables included historical observed discharges, mean temperature, monthly precipitation, and elevation of each qanat. After preprocessing (normalization and missing data imputation), the dataset was split into training (70%) and testing (30%) sets. The model was trained with optimized hyperparameters (max_depth=10, max_features=0.5, min_samples_split=5, n_estimators=500). Results showed strong performance: RMSE = 4.27 L/s, NSE = 0.54, r = 0.76, KGE = 0.51, and NRMSE = 15.19%. Sensitivity analysis using SHAP and Permutation Importance identified elevation as the key variable (mean |SHAP| = 2.16), while precipitation exhibited a slight negative influence (SHAP = -0.013). These findings validate RF's efficacy and provide a practical framework for water resource management in arid regions.

Keywords

• Qanat discharge prediction
• climate data
• spatio-temporal variations
• sensitivity analysis

Main Subjects

• ًQanat and aquifer