Abstract
Accurate estimation and prediction of monthly rainfall are fundamental for effective water resources management, agricultural planning, drought preparedness, and flood risk mitigation, particularly in data-scarce, semi-arid regions. In Erbil, Iraq, characterized by a hot-summer Mediterranean climate and high interannual rainfall variability, reliable monthly precipitation forecasting is essential for sustainable water resources management. However, traditional statistical approaches, such as linear regression models, often fail to adequately represent the complex, nonlinear interactions among meteorological variables that govern rainfall processes. Consequently, more accurate and robust modeling techniques are required to support informed hydrometeorological decision making.This study systematically evaluates the performance of five machine learning (ML) models: Artificial Neural Network (ANN), Multiple Linear Regression (MLR), Extreme Gradient Boosting (XGB), Random Forest (RF), and Support Vector Regression (SVR) for monthly rainfall prediction using a 20-year dataset (October 2004–September 2023). Six input scenarios (M1–M6) were developed, with the final scenario incorporating comprehensive lagged meteorological variables to better capture temporal dependencies. Model performance was assessed using the coefficient of determination (R²), root mean squared error (RMSE), normalized RMSE (nRMSE), and mean bias error (MBE).The results reveal two key findings. First, the weak performance of the MLR model (validation R² < 0.60) confirms the inherently nonlinear nature of monthly rainfall dynamics in the study region. Second, although ensemble tree-based models (XGBoost and Random Forest) exhibit near-perfect calibration performance (R² ≈ 0.99), their generalization ability deteriorates during validation, indicating pronounced overfitting due to high variance and the limited temporal depth of the climatological record. For example, the XGB model showed poor predictive capability under Scenario M5, with validation R² = 0.37 and RMSE = 59.13. In contrast, the ANN model particularly under the full lagged input configuration (M6) demonstrated the most effective balance between model complexity and generalization, achieving the highest validation accuracy (R² = 0.80, RMSE = 22.10) along with minimal systematic bias (MBE = −1.15)Overall, this study establishes ANN as a reliable and robust nonlinear modeling approach for monthly rainfall prediction in Erbil, highlighting its suitability for operational water resources management in semi-arid environments. This framework demonstrates that shallow neural networks can serve as a reliable and computationally efficient decision-support tool for baseline water resource planning in semi-arid Mediterranean climates.
