KAL (Kernel Atmospheric Learner) is a hybrid machine learning framework, developed and trained on the European Weather Cloud (EWC), for high-resolution daily Quantitative Precipitation Estimation (QPE) in topographically complex regions that lack ground-based weather radar.
The framework combines a 10-year archive of Meteosat Second Generation (MSG) SEVIRI multispectral observations with Copernicus DEM topographic metadata and applies a two-stage "Hurdle" architecture: a Deep Multi-Layer Perceptron (Keras/TensorFlow) for binary rainfall detection, followed by an XGBoost regressor for precipitation intensity quantification. This design directly addresses the strong zero-inflation bias typical of Mediterranean rainfall data.
Ground truth was provided by rain gauge data from the agrometeorological station network of the University of Ioannina (UoI). The framework was validated using a Leave-One-Station-Out (LOSO) cross-validation strategy against six agrometeorological rain gauge stations in the Arta plain (Epirus, Greece), achieving a mean F1-score of 0.796 and a Critical Success Index of 0.663 for nighttime events, and a mean Kling-Gupta Efficiency of 0.658.
Estimation of resources used to train the model:
KAL is now running operationally within EWC, continuously processing near real-time MSG SEVIRI data, and has become an active asset for our Department (Agriculture, University of Ioannina).
An independent blind test on the extreme rainfall event of December 4, 2025 (fully excluded from training) showed strong structural agreement with the operational EUMETSAT H-SAF H90 product. KAL demonstrates a scalable, cost-effective approach for bridging precipitation-monitoring gaps in radar-void, mountainous regions across the wider Mediterranean basin.
