ERA5–NASA Ensembles for Daily Rain Prediction Supporting Irrigation in Konya, Türkiye

Reliable daily precipitation prediction is pivotal for agricultural water management in semi-arid regions, where water scarcity and climate variability raise operational risk. This study develops and evaluates machine-learning approaches for precipitation occurrence in Konya, Türkiye - one of the country’s key agricultural basins - using 44 years (1980–2024) of meteorological data. We examine three data strategies: (i) European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis v5 (ERA5) reanalysis features alone, (ii) National Aeronautics and Space Administration (NASA) Prediction Of Worldwide Energy Resources (POWER) observations alone, and (iii) a combined strategy that merges ERA5 predictors with the NASA rain/no-rain target. Five model families are compared— Random Forest (RF), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Categorical Boosting (CatBoost), and a three-layer Long Short-Term Memory (LSTM) on analysis-ready tabular datasets. All approaches use advanced feature engineering (cyclical seasonality, lags, rolling windows) and a deployment-minded post-processing step that scans decision thresholds from 0.10 to 0.90 to maximize F1‑score. Performance is assessed with a chronological train/test split; F1‑score is the primary metric, complemented by Area Under the Receiver Operating Characteristic Curve (AUC-ROC), precision, recall, confusion matrices, and calibration. Results show that the combined strategy with CatBoost delivers the highest skill (F1‑score: 84.57%; AUC-ROC: 94.37%), confirming that pairing rich ERA5 features with the quality-controlled NASA target improves performance relative to single-source approaches. Ensemble tree-based methods consistently outperform the LSTM baseline on this daily, tabular classification task. Threshold optimization raises F1‑score by about 1–5% across models, and calibration indicates that predicted probabilities closely match observed frequencies. For semi-arid farming in Konya, calibrated daily probabilities of precipitation can be converted into simple decision rules (e.g., skip irrigation when the predicted probability exceeds a locally selected threshold), supporting irrigation scheduling, fertilizer timing, and harvest planning. The workflow is computationally efficient and reproducible, built on globally available ERA5 and NASA POWER resources, and readily transferable to other semi-arid basins with minor re-tuning. Findings highlight a practical path to trustworthy, probability-based rainfall guidance for agricultural water management. Feature importance analysis highlights seasonality, humidity, temperature ranges, pressure tendencies, and wind extremes as leading signals.