Google Global Flood Forecast v2 Extends Reliable Forecast Horizon by 6 Days and Boosts Accuracy

Background and Motivation

Floods are among the most widespread and damaging natural hazards, and accurate river runoff predictions are critical for disaster mitigation, ecological safety, and socioeconomic stability. Over recent decades, machine learning has progressed from conceptual rainfall‑runoff models—useful for data‑scarce basins—to sophisticated approaches that combine data‑driven methods with physical mechanisms, uncertainty quantification, and data assimilation.

System Upgrade Overview (v2)

Google Research has deployed the second version of its global flood forecasting system (v2) as the core engine of Google FloodHub. The upgrade addresses three long‑standing constraints: insufficient training data, limited sequence length, and input data distribution shift. These changes markedly improve the stability and reliability of global runoff forecasts.

Dataset Release (GRRR)

Alongside v2, the team released the Google River Runoff Reanalysis and Reforecast dataset (GRRR), covering over one million river stations worldwide with decades of simulated and reforecast results, providing a valuable foundation for future research.

Input Features

The model ingests three categories of inputs:

Static basin attributes : 92 spatially averaged variables (e.g., mean elevation, aridity, precipitation seasonality, forest cover, soil hydraulic properties, population density) derived from HydroATLAS and ERA5‑Land.

Dynamic meteorological drivers : Multi‑source products including ECMWF HRES, NOAA CPC, GraphCast, and NASA IMERG, providing daily totals of precipitation, 2‑m temperature, and other key factors.

Target runoff observations : Combined training from Caravan, GRDC, and BANDAS datasets, expanding coverage beyond the GRDC‑only training of v1.

Multi‑source fusion mitigates regional and temporal errors, while a masking‑mean mechanism handles missing inputs.

Architectural Improvements

The core of v2 is the Mean‑Embedding LSTM (ME‑LSTM), replacing the encoder‑decoder LSTM (ED‑LSTM) used in v1. ME‑LSTM treats each meteorological product as a separate input stream, maps them through dedicated embedding networks into a shared latent space, and aggregates them with a mask‑mean operation, improving robustness to missing data and distribution shifts.

Unlike v1, which split the sequence into “reanalysis” and “forecast” phases with separate LSTMs, ME‑LSTM processes the entire time series with two stacked LSTM layers, eliminating the state‑transition discontinuity that caused early‑time instability.

Both versions use a mixture‑density output layer that predicts parameters of a count‑mixture asymmetric Laplace distribution (CMAL) to capture runoff uncertainty; deterministic forecasts are taken as the distribution mean.

Training Procedure

v2 is trained with the Adam optimizer and CMAL likelihood loss for 125 epochs. Regularization includes Gaussian noise injection, gradient clipping, and random dropping of input timesteps to enhance robustness to real‑world data gaps.

Evaluation Methodology

The evaluation uses 1,223 test basins, split into basins with in‑situ measurements and “no‑data” basins that rely entirely on cross‑basin generalization. Independent test sets replace the random ten‑fold cross‑validation of v1, better reflecting operational deployment. Metrics include Nash‑Sutcliffe Efficiency (NSE) and Kling‑Gupta Efficiency (KGE). Baselines comprise the Global Flood Awareness System (GloFAS) and the European Flood Awareness System (EFAS).

Results

Key findings:

v2 achieves higher NSE across all forecast horizons; reliable forecast length extends by 6 days for basins with stations and by 1 day for basins without.

Overall accuracy surpasses v1 and two third‑party benchmarks.

Architecture upgrades mainly benefit basins with observations, while the inclusion of GraphCast yields notable gains for medium‑ to long‑range forecasts and improves both observed and unobserved basins.

KGE decomposition shows that performance gains stem from better modeling of runoff temporal dynamics and flow variability.

Dry basins and those with extensive reservoir regulation remain challenging; however, v2 shows larger relative improvements in dry basins when observations are available.

Limitations and Outlook

Despite advances, the system still struggles in data‑scarce, heavily regulated, or arid regions, and its dependence on local observations has not been fully eliminated. The work demonstrates that high‑quality training data, global multi‑source meteorology, and task‑specific deep learning architectures can support truly global flood forecasting, offering a promising direction for flood risk management and climate resilience.