Quantitative Precipitation Estimation and Mapping in Radar-Absent Regions: A Hybrid MSG SEVIRI and Machine Learning Framework Validated by Rain Gauge Reference Observations

Accurate Quantitative Precipitation Estimation (QPE) is a critical imperative for hydrological modeling, disaster risk reduction and optimized irrigation scheduling in smart agriculture, yet it remains a significant challenge in topographically complex regions lacking weather radar infrastructure. This study proposes a robust hybrid machine learning framework for near-real-time precipitation mapping over the Arta plain, Greece, a vital agricultural zone characterized by pre-orographic enhancement from the Pindus mountain range. Utilizing an 11-year record of Meteosat Second Generation (MSG) SEVIRI observations, topographic metadata, and dynamic rate-of-change features, we implement a two-stage “Hurdle” model architecture. By decoupling the detection of rain occurrence from the estimation of its intensity through diurnal stratification, the framework effectively mitigates the zero-inflation bias inherent in imbalanced Mediterranean rainfall datasets. The first stage employs an optimized Multi-Layer Perceptron (MLP) for binary rainfall detection, while the second stage utilizes an XGBoost regressor for precise volumetric quantification. To ensure spatial consistency within the study area and model rigor, the framework was evaluated via a Leave-One-Station-Out (LOSO) cross-validation strategy, using a dense network of automated agrometeorological stations as the exclusive ground truth. Results demonstrate high diagnostic performance across both diurnal regimes, with the framework achieving a mean F1-score of 0.786 (CSI: 0.652) for nighttime and 0.774 (CSI: 0.631) for daytime events, achieving performance levels comparable to established operational standards, although station-level skill varied with proximity to the Pindus foothills. Relative to a conventional single-stage regressor, the proposed hurdle formulation reduced the False Alarm Ratio by over 50% and improved the Critical Success Index by more than 65% in both diurnal regimes, confirming the necessity of decoupling rain detection from intensity quantification. Quantitatively, the model exhibited strong hydrological skill with a mean Kling-Gupta Efficiency (KGE) of 0.658 and 0.629 for night and day periods respectively, alongside near-zero Mean Error, confirming its ability to capture intense localized precipitation. This research provides a computationally efficient and cost-effective alternative for real-time hydrological monitoring in radar-absent basins, offering a viable solution where expensive infrastructure is unavailable.