River floods are among the most devastating natural disasters, threatening population, infrastructures and economy. The goal of this work is to adopt an end-to-end strategy to directly predict river levels in the short-term, up to six hours ahead. Unlike commonly used approaches based on multistep forecasting chains or complex physical models, radar frames are directly used as input to perform predictions. The architecture combines convolutional 3D layers to extract spatiotemporal features, with a dense linear encoder for past hydrologic trends, and provide probabilistic confidence intervals via quantile regression. The results of this methodology, validated on regional Piedmont data, show superior performance compared to the persistence model, with a reduction of almost 80% for MSE and improved reliability during extreme events.