Ron Sarafian

Accurate real-time quantitative precipitation estimation (QPE) is essential for rain-dependent sectors such as agriculture, hydrology, and emergency management. Yet, it remains a significant challenge due to the complex spatiotemporal variability of precipitation. Precipitation radar observations provide high-resolution spatiotemporal data but often suffer from limitations such as beam elevation and spreading, or various empirical Z - R relationships, leading to biases in surface rainfall estimation. Integrating a precipitation radar with rain-gauge data improves accuracy but introduces additional biases, mostly due to interpolation assumptions. This study introduces RadarGaugeNet, a novel deep learning model designed to generate continuous, high-resolution, real-time QPEs by leveraging radar data supervised solely by real-time rain-gauge observations. RadarGaugeNet employs an encoder–decoder architecture with a mechanism that learns spatially varying radar error patterns without imposing assumptions about rainfall distribution or geographic factors. Trained and validated on a decade of rainfall data from Israel, RadarGaugeNet significantly improves precision, recall, and root-mean-squared error (RMSE) across diverse climatic and topographical regions, with greater improvement in remote regions where radar coverage is limited. RadarGaugeNet effectively corrects radar biases, avoids interpolation artifacts, and captures terrain-induced precipitation patterns without explicit orographic input. These advancements highlight RadarGaugeNet’s potential for scalable, accurate QPE in complex environments, addressing critical gaps in operational meteorology.