Flood Forecasting

At Google, we’re investing in technologies that can help communities prepare for and respond to climate-related disasters and threats. As part of our Crisis Response efforts, we're working to bring trusted information to people in critical moments to keep them safe and informed. To do so, we rely on the research and development of our AI-powered technologies and longstanding partnerships with frontline emergency workers and organizations. Our goal with this program is to provide accurate and actionable flood alerts covering all affected by floods globally.

We currently focus only on riverine floods, as opposed to flash and coastal floods. We also do not generate flood maps for urban areas.

The Google hydrology model is the first operational model that uses machine learning to generate improved hydrology forecasts in more locations globally.

The Google inundation maps model is one of the few flood models that generate maps that allow to identify which villages or areas are going to be flooded given a specific hydrology forecast.

Forecasts are updated daily based on the most up-to-date meteorological data available. The forecasts are also in daily resolution.

We do not use any of the countries' proprietary data, but a variety of weather products, including ECMWF forecasts, CPC rain gauge measurements, IMERG precipitation and Copernicus Sentinel-1 satellites of the European Space Agency. Sentinel-1 is a C-band SAR satellite. We then use our algorithms to calculate the flooded area based on the SAR image.

We are working on gradually rolling out forecasts in more regions. Google only approves the release of data and forecasts in locations that have been thoroughly evaluated and are deemed to be of sufficient quality. We are continuously evaluating new locations and improving the quality of our forecasts as necessary to launch flood forecasts in more areas of the world.

The availability of local historical discharge records dramatically improves model quality. If you would like to contribute discharge data to facilitate coverage in your area, please consider looking at the Caravan project, which facilitates publishing streamflow data, combined with meteorological forcing data and catchment attributes.

As is common in riverine flood forecasting systems, there are two types of models involved in producing these alerts: a hydrologic model that calculates the one-dimensional flow in the river, and an inundation model that uses the flow to calculate a flood map.

The model uses different sets of features in different parts of the model, namely the hindcast LSTM (the part of the model that summarizes the past until the point of the issued forecast) and the forecast LSTM.

• Input features to the hindcast portion of the model are CPC precipitation, IMERG precipitation, ECMWF IFS nowcasts, including precipitation, temperature, and other surface (single-level) variables.

• Input features to the forecast portion of the model are various ECMWF IFS-HRES bands from the most recently issued IFS forecast, including precipitation, temperature, solar radiation, windspeed, and surface pressure.

• Additionally, both parts of the model (hindcast and forecast) use static geophysical attributes derived from HydroATLAS, alongside climate indices derived from long-term records from ERA5-Land.

More meteorological data products will be added to the model to improve the prediction quality (see for example this publication for details on the effect of using multiple forcing products).

We archive most of our forecasts in a publicly available dataset, see documentation here.

We evaluate the hydrologic and inundation models separately. In some regions, we see that our hydrologic models are very accurate when compared to ground truth information, but the inundation models do not show clear and regular inundation patterns. In these regions we decide to only share the hydrologic information.

Since our flood forecasting models are based on machine learning algorithms (as opposed to classic physics-based models), our models can incorporate dam behavior implicitly (i.e. we don’t need to explicitly code that in). We’ve seen very good performance of our models including in basins that are heavily instrumented and affected by dams. 

We are also working towards more explicit incorporation of dams, and also using our models to provide recommendations for dam management. This is currently in pilot research phase (in collaboration with the Indian government), and may become a service we provide in the future.

The benchmarking was done in collaboration with the GloFAS team itself. It involved testing the accuracy of our models relative to their models on all points in the world where we have gauge measurements (and therefore can know what the “correct prediction” is). We are currently working on submitting these results, together with GloFAS, to Nature.

If a country provides historical or real-time data we can use this to further train (or “calibrate”) the model, leading to a significant improvement in prediction quality at the location of the data (could even be 10% in NSE scores or more), and a more modest improvement in other locations. In the future we will also be able to incorporate real-time data as an input to the model, which will provide an even larger improvement in accuracy.

The data needed is discharge measurements at a daily resolution or better (e.g. hourly), with timestamps and the location of the gauge.