The Next-Generation Google Flood Forecasting Model & Community Resources

Accurate global flood forecasting, particularly in ungauged basins, remains a primary challenge for operational hydrology and disaster risk reduction. While data-driven approaches using Long Short-Term Memory (LSTM) networks have set new benchmarks in global streamflow prediction, limitations regarding forecast lead time, temporal consistency, and operational robustness to missing data persist.

In this work, we present the next-generation Google global hydrologic model, which introduces three major advancements over previous state-of-the-art systems. First, we integrated AI-based medium-range weather forecasts as additional meteorological forcing, alongside traditional deterministic products. Second, leveraging recent contributions to the Caravan community dataset, we expanded the training dataset three-fold to include nearly 16,000 streamflow gauges globally. Third, we implemented a novel masked mean embedding LSTM architecture. This design eliminates the traditional encoder-decoder state hand-off issue (which introduces temporal inconsistencies or forecast hairs) and enables the model to remain operational during weather data outages by dynamically averaging embeddings from available input sources.

Our results demonstrate a significant extension of the reliable forecast horizon: the new model achieves accuracy at a 7-day lead time comparable to the 5-day lead time performance of its predecessor. Furthermore, the model continues to outperform other global operational systems, such as GloFAS and GeoGlows, across both gauged and ungauged basins. These advancements represent a significant step toward providing more timely and reliable flood warnings in regions where traditional monitoring infrastructure is scarce.

In conjunction with this update, we released two new community resources. The Google Runoff Reanalysis & Reforecast (GRRR) dataset provides a comprehensive, multi-decade reforecast archive generated by the current operational global model. Additionally, we have launched the GoogleHydrology GitHub repository, which provides an open-source research implementation that closely approximates our operational environment. This release is intended to facilitate the reproduction of our findings and provide the scientific community with a robust baseline for future global hydrologic modeling research.