Resources

Google's goal is to provide accurate and actionable flood alerts covering all affected by floods globally. Currently the focus is on riverine floods.

At Google, our mission has always been to organize the world’s information and make it universally accessible and useful. During a crisis, that is often when users need the information most. People come to Google to learn about what's going on and how to remain safe, and we have an opportunity and responsibility to help.

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 organizations. Our goal with this program is to provide accurate and actionable flood alerts covering all affected by floods globally.

Our Flood Forecasting tool is a part of our AI for Social Good portfolio. Products within this portfolio are meant to provide social impact to users and partners. All of our forecasts and products associated with Flood Forecasting are free of charge to partners and users.

Google has been using artificial intelligence to forecast floods since we started this project in 2017. Over the past few years our model has developed significantly. When we first started, we relied on real time river data from each country we forecasted in. Thanks to advancements in AI, we can now forecast in most countries in the world.

Our models are currently covering more than 5,000 locations across river basins in over 100 countries. Our research teams are working tirelessly to develop ways to expand our coverage.

Flood Hub is showing forecasts for locations in the following countries:

Angola, Argentina, Australia, Austria, Azerbaijan, Bangladesh, Belarus, Belgium, Belize, Bolivia, Botswana, Brazil, Bulgaria, Burkina Faso, Cambodia, Cameroon, Canada, Central African Republic, Chad, Chile, Colombia, Costa Rica, Croatia, Czech Republic, Democratic Republic of Congo, Denmark, Ecuador, Estonia, Finland, France, Gambia, Georgia, Germany, Ghana, Greece, Guatemala, Guinea, Guinea Bissau, Guyana, Honduras, Hungary, India, Indonisia, Ireland, Italy, Ivory Coast, Kazakhstan, Kyrgyzstan, Laos, Latvia, Lesotho, Liberia, Lithuania, Madagascar, Malawi, Malaysia, Mali, Mexico, Mozambique, Myanmar, Namibia, Nepal, Netherlands, New Zealand, Nicaragua, Nigeria, Norway, Pakistan, Papua New Guinea, Paraguay, Peru, Philippines, Poland, Portugal, Republic of the Congo, Romania, Senegal, Serbia, Sierra Leone, Slovakia, Slovenia, Somalia, South Africa, South Sudan, Spain, Sri Lanka, Suriname, Sweden, Switzerland, Tajikistan, Thailand, Turkey, Ukraine, United Kingdom, United States, Uruguay, Uzbekistan, Venezuela, Vietnam, Zambia, Zimbabwe.

go to g.co/floodhub for a free and public access

you can see more unverified reporting points by entering the Flood Hub Expert Mode

If you are an organization interested in implementing our flood forecasts in your workflows, please see the API bookmark.

• Warning and Danger and Extreme Flood Status levels are set in most areas of the world based on 2-year and 5-year and 20-year return periods, respectively. We are able to override these defaults in locations where local water agencies have communicated to us warning and danger levels for specific locations that are more commensurate with river levels that have potential to affect humans.
• In some cases, depending on data availability, a map of current or expected floods
• In some cases, depending on data availability, an illustration of water depth compared to the human body, e.g. ankle height or waist height.

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

The predictions are still a work in progress and are for informational purposes only. You should not use them as a sole source of data in the event of an emergency, but as potentially one data point in addition to other sources of data (e.g., the government, weather services, etc.).

Our researchers are evaluating a minimum level of quality before launching our model in a new location.

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.

One limiting factor is the availability of historical discharge records in certain parts of the world. If more discharge data (historical or real-time) would be publicly available, it would speed up our process to cover more rivers globally.

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.

Various versions of LSTM-based streamflow models have been benchmarked extensively in peer-reviewed publications. For example, this paper compares the performance of an LSTM-based simulation (not forecast) model over watersheds in the continental US with several commonly-used research models. This paper compares a simulation LSTM against the two hydrology models that are run operationally for flood forecasting in the United States.

The most up-to-date peer-reviewed model performance statistics can be found in this paper. That paper compares the skill of the Global Flood Awareness System (GloFAS) with Google’s hydrology model at predicting the timing of extreme events (1, 2, 5 ,and 10-year return period events). The Google model shows better performance, on average, in all continents, and we achieve either better or statistically indistinguishable performance at five-day lead times as GloFAS achieves over nowcasts (zero-day lead times) – see figure below.

We recently published our paper “Global prediction of extreme floods in ungauged watersheds” published in Nature, showing that AI generated global flood forecasts outperform existing models.

A longstanding challenge in hydrology is the problem of Prediction in Ungauged Basins (PUB). This refers to making streamflow predictions in river reaches where there is no streamflow data for calibrating models. Machine Learning hydrology models are able to transfer (learned) information between different watersheds and are significantly more reliable in ungauged catchments than other types of hydrology models. This is one of the main reasons why we use an ML-based modeling approach in the Google flood forecasting system. More information, including benchmarks against two hydrology models used operationally in the United States, can be found in this paper.

The AI model has higher scores than GloFAS in all continents and return periods (p < 1e − 2, 0.10 < d < 0.68) with three exceptions where there is no statistical difference: Africa over 1-year return period events (p = 0.07, d = 0.03) and Asia over 5-year (p = 0.04, d = 0.12) and 10-year (p = 0.18, d = 0.12) return period events.

Over 5-year return period events, GloFAS has a 54% difference between mean F1 scores in the lowest scoring continent (South America: f1 = 0.15) and the highest scoring continent (Europe: f1 = 0.32), meaning that, on average, true positive predictions are twice as likely (at a proportional rate). The AI model also has a 54% difference between mean F1 scores in the lowest scoring continent (South America: f1 = 0.21) and the highest scoring continent (South West Pacific: f1 = 0.46), which is due mostly to a large increase in skill in the South West Pacific relative to GloFAS (d = 0.68).

All of the data that we currently use to train and test the hydrology model is publicly available, and all data except meteorological forecasts are freely available (i.e., without monetary charge). This means, in principle, all of our modeling results are reproducible and verifiable. The meteorological forecast data that we use comes from the ECMWF Integrated Forecast System, GraphCast and obtaining this data in real time requires a paid license from ECMWF.

Target data: The hydrology model is trained on publicly available historical data from the Global Runoff Data Center, the Caravan dataset and BANDAS gauge data from Mexico for training.

We are currently integrating public data from the Caravan dataset into our model training.

Static attributes: Geophysical and geographical data are provided to the model as inputs. These come largely from the HydroAtlas dataset, which is part of the HydroSheds project. These include variables about climate, land use and land cover, soils, and human impacts. A full list of the static attributes that we use is available on request. These data are scalar (single) values representing aggregation (fractional coverage, mean, mode, max, etc.) over a given (sub)watershed.

Meteorological data: For the vast majority of our forecasts we rely on a variety of publicly available weather products, including ECMWF forecasts, IMERG precipitation, NOAA precipitation data, CPC rain gauge measurements, GraphCast forecasts by Google 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.

In a small number of cases we rely on local historical and real time data provided by the following governments:

• Bangladesh: BWDB - Bangladesh Water Development Board,
• Sri Lanka: Department of Irrigation
• Brazil: ANA- Agência Nacional de Águas e Saneamento Básico, Civil Defense and SGB/CPRM - Serviço Geológico do Brasil
• India: CWC - Central Water Commission

Unless otherwise stated, the models do not use real-time data provided by governmental entities, nor is Google Flood Forecasting project affiliated with, sponsored by, or endorsed by any governmental entity.

Share through the Caravan Dataset - Caravan is an open source global streamflow dataset, where global streamflow data providers (e.g., hydromet agencies) contribute daily. Several countries are currently represented in this dataset, including (but not limited to) Australia, Brazil, Canada, Chile, Denmark, Mexico, the UK, the US, and much of central Europe. Moreover, the WMO-supported Global Runoff Data Center recently contributed streamflow data for 25 additional countries.

The hydrology model is trained using streamflow data from the Global Runoff Data Center over the time period 1980 - 2023. We use training data from approximately 5,000 streamflow gauges, illustrated in the figure below. These locations represent watersheds ranging from 2 km2 to over 4,000,000 km2. We use data from all sizes of watersheds to train a single model.

Notice that not all input data is available for the whole training period (1980 - 2023). In cases where a certain data product is missing we either mask or impute the missing data. ECMWF HRES forecast data is imputed with the corresponding ERA5-Land variable during the time period 1980 - 2016. Missing data from IMERG or CPC is masked, meaning that the actual missing value is imputed with a dummy value (either zero or the overall mean of the dataset), and then a binary flag is fed as an extra input into the model that signals whether the data is real or masked.

As mentioned above, the model is trained using a negative log-likelihood loss function. Other hyperparameters of the model (being quite numerous) are available on request. No part of the model architecture or training procedure is proprietary.

Various versions of LSTM-based streamflow models have been benchmarked extensively in peer-reviewed publications. For example, this paper compares the performance of an LSTM-based simulation (not forecast) model over watersheds in the continental US with several commonly-used research models. This paper compares a simulation LSTM against the two hydrology models that are run operationally for flood forecasting in the United States.

The most up-to-date peer-reviewed model performance statistics can be found in this paper. That paper compares the skill of the Global Flood Awareness System (GloFAS) with Google’s hydrology model at predicting the timing of extreme events (1, 2, 5 ,and 10-year return period events). The Google model shows better performance, on average, in all continents, and we achieve either better or statistically indistinguishable performance at five-day lead times as GloFAS achieves over nowcasts (zero-day lead times) – see figure below.

Our objective is to understand the reliability of forecasts of extreme events, so we report precision, recall, and F1 scores (F1 scores are the harmonic mean of precision and recall) over different return period events.

The inundation models are ML based on models, trained based on past flood events.

The past events are Satellite-based flood inundation maps: The synthetic aperture radar ground range detected (SAR GRD) data from the Sentinel-1 satellite constellation are used to determine flood inundation maps at known timepoints and locations . At any AOI (area of interest), a SAR image is available once every several days, from which an inundation map was inferred using a binary classifier (Torres, 2012). Every pixel within a SAR image is classified as wet/dry via a Gaussian mixture based classification algorithm. In order to calibrate and evaluate the classification algorithm, we have collected a dataset of Sentinel 2 multispectral images of flood events that coincide with the SAR image dates and locations. Reference Sentinel-2 flood maps were created by calculating per-pixel Normalized Difference Water Index (NDWI=(B3-B8)/(B3 + B8), B3 and B8 are green and near infrared bands, respectively) and applied a threshold of 0.

These flood images are the ground truth for the ML model, while the inputs are the gauge value, the latest available satellite image, and flood history of the region.

The inundation models are trained and validated based on historical flood events, where flood inundation extent maps from satellite data, along with the corresponding gauge water stage measurements, are available. Similar to the stage forecast models, a 1-year leave out cross validation scheme is used for training and validation.

Lumped Catchment Modeling

Our river forecast model is a lumped catchment model. This means that it directly predicts streamflow (discharge) at a given river reach. We do not use any type of routing model to route water downstream or through a river network. The main limitation of this approach is that we are currently unable to assimilate near-real time streamflow data to improve downstream predictions in real time. We are currently developing a graph modeling approach that simulates both rainfall-runoff processes and routing processes.

Dams and Reservoirs

We do not currently explicitly model dam and reservoir operations. In gauged locations, our ML models are able to learn some of the dam and reservoir operation signal that is present in downstream gauge data, however we do not expect that any learned patterns of regular operations will necessarily be similar to operational procedures during times of flooding. This is an area of ongoing research, and we are looking into a dam and reservoir model in the future.

Physical Realism

A common concern with ML-based hydrology modeling is that ML models are not physically constrained, and predictions therefore might not be physically realistic. After many years of working with these models in both research and operational environments, we have not observed cases where this happens.

Mass balance or other physical constraints

We have two peer-reviewed publications on the topic of introducing mass balance constraints into the ML models to ensure that water is not lost or created within the modeled systems. We approached this question by developing an ML model that is constrained by mass conservation.

This paper looks at the effect of enforcing mass balance constraints on peak flow estimates, and found that ML models that do not include mass balance constraints have less bias when predicting peak flow events. Similarly, this paper looks at the effect of enforcing mass balance constraints on biases in the overall water balance, and again found that ML models which don’t include explicit mass balance constraints result in less biased predictions. The reason for this is because ML models are able to learn local, heterogeneous, and heteroscedastic biases in precipitation data, and can mitigate these input biases to result in less biased streamflow forecasts.

Snow accumulation and snowmelt

Using explainable AI techniques, we found that our LSTM-based streamflow models are able to model snow accumulation and snow melt processes effectively without being trained on snow data explicitly. The LSTM models use certain cell states (portions of the state vector that is recursive in time) to track snow accumulation and snow melt as a function of temperature. Certain LSTM states correlate strongly with snow accumulation, and the model learns to release water from those states when temperature is higher than freezing. Again, this is done without the model ever needing to see snow data explicitly. We published a peer-reviewed book chapter on this, and a subsequent paper from researchers at the University of Oxford confirmed these findings.

Geomorphological Changes and River Topologies

Our models do not simulate geomorphological change, however we work with the HydroSheds group, partially funding the development of updated global hydrography (river network) datasets. We are able to retrain our global model in less than 2 days, meaning that we can re-adapt to any hydrography changes quickly. As a point of reference, it takes weeks to months to re-calibrate a traditional hydrology model on a similar sized dataset.

Changes in Climate or Land Use and Land Cover

Although land use and land cover (LULC) is changing rapidly, we approximate land cover in watersheds as static. This is an approximation, and we are exploring to use satellite imagery as input data instead of static catchment attributes related to LULC. However, we have done extensive research on this and found that using dynamic land cover indexes does not generally improve the quality of streamflow predictions.

We retrain our models frequently, and climate change occurs at much longer timescales than our retraining cycles.