Using AI to expand global access to reliable flood forecasts

Posted by Yossi Matias, VP Engineering & Research, and Grey Nearing, Research Scientist, Google Research

Floods are nan most communal earthy disaster, and are responsible for astir $50 billion successful yearly financial damages worldwide. The rate of flood-related disasters has much than doubled since nan twelvemonth 2000 partially due to ambiance change. Nearly 1.5 cardinal people, making up 19% of nan world’s population, are exposed to important risks from terrible flood events. Upgrading early informing systems to make meticulous and timely accusation accessible to these populations can prevention thousands of lives per year.

Driven by nan imaginable effect of reliable flood forecasting connected people’s lives globally, we started our flood forecasting effort successful 2017. Through this multi-year journey, we precocious investigation complete nan years hand-in-hand pinch building a real-time operational flood forecasting strategy that provides alerts connected Google Search, Maps, Android notifications and done nan Flood Hub. However, successful bid to scale globally, particularly successful places wherever meticulous section information is not available, much investigation advances were required.

In “Global prediction of utmost floods successful ungauged watersheds”, published successful Nature, we show really instrumentality learning (ML) technologies tin importantly amended global-scale flood forecasting comparative to nan existent state-of-the-art for countries wherever flood-related information is scarce. With these AI-based technologies we extended nan reliability of currently-available world nowcasts, connected average, from zero to 5 days, and improved forecasts crossed regions successful Africa and Asia to beryllium akin to what are presently disposable successful Europe. The information of nan models was conducted successful collaboration pinch nan European Center for Medium Range Weather Forecasting (ECMWF).

These technologies besides alteration Flood Hub to supply real-time stream forecasts up to 7 days successful advance, covering stream reaches crossed complete 80 countries. This accusation tin beryllium utilized by people, communities, governments and world organizations to return anticipatory action to thief protect susceptible populations.

Flood forecasting astatine Google

The ML models that powerfulness nan FloodHub instrumentality are nan merchandise of galore years of research, conducted successful collaboration pinch respective partners, including academics, governments, world organizations, and NGOs.

In 2018, we launched a pilot early informing strategy successful nan Ganges-Brahmaputra stream basin successful India, pinch nan hypothesis that ML could thief reside nan challenging problem of reliable flood forecasting astatine scale. The aviator was further expanded nan pursuing twelvemonth via nan combination of an inundation model, real-time h2o level measurements, nan creation of an elevation representation and hydrologic modeling.

In collaboration pinch academics, and, successful particular, pinch nan JKU Institute for Machine Learning we explored ML-based hydrologic models, showing that LSTM-based models could produce much meticulous simulations than accepted conceptual and physics-based hydrology models. This investigation led to flood forecasting improvements that enabled nan expansion of our forecasting sum to see each of India and Bangladesh. We besides worked pinch researchers astatine Yale University to trial technological interventions that summation nan reach and impact of flood warnings.

Our hydrological models foretell stream floods by processing publically disposable upwind information for illustration precipitation and beingness watershed information. Such models must beryllium calibrated to agelong information records from streamflow gauging stations successful individual rivers. A debased percent of world stream watersheds (basins) person streamflow gauges, which are costly but basal to proviso applicable data, and it’s challenging for hydrological simulation and forecasting to supply predictions successful basins that deficiency this infrastructure. Lower gross home product (GDP) is correlated pinch accrued vulnerability to flood risks, and location is an inverse relationship betwixt nationalist GDP and nan magnitude of publically disposable information successful a country. ML helps to reside this problem by allowing a single exemplary to beryllium trained connected each disposable stream data and to beryllium applied to ungauged basins wherever no information are available. In this way, models tin beryllium trained globally, and tin make predictions for immoderate stream location.

There is an inverse (log-log) relationship betwixt nan magnitude of publically disposable streamflow information successful a state and nationalist GDP. Streamflow information from nan Global Runoff Data Center.

Our world collaborations led to ML investigation that developed methods to estimate uncertainty successful stream forecasts and showed really ML stream forecast models synthesize accusation from aggregate information sources. They demonstrated that these models tin simulate utmost events reliably, moreover erstwhile those events are not portion of nan training data. In an effort to contribute to unfastened science, successful 2023 we open-sourced a community-driven dataset for large-sample hydrology successful Nature Scientific Data.

The stream forecast model

Most hydrology models utilized by nationalist and world agencies for flood forecasting and stream modeling are state-space models, which dangle only connected regular inputs (e.g., precipitation, temperature, etc.) and nan existent authorities of nan strategy (e.g., ungraded moisture, snowpack, etc.). LSTMs are a version of state-space models and activity by defining a neural web that represents a azygous clip step, wherever input information (such arsenic existent upwind conditions) are processed to nutrient updated authorities accusation and output values (streamflow) for that clip step. LSTMs are applied sequentially to make time-series predictions, and successful this sense, behave likewise to really scientists typically conceptualize hydrologic systems. Empirically, we person recovered that LSTMs execute well connected nan task of stream forecasting.

A sketch of nan LSTM, which is simply a neural web that operates sequentially successful time. An accessible primer tin beryllium recovered here.

Our stream forecast exemplary uses 2 LSTMs applied sequentially: (1) a “hindcast” LSTM ingests humanities upwind information (dynamic hindcast features) up to nan coming clip (or rather, nan rumor clip of a forecast), and (2) a “forecast” LSTM ingests states from nan hindcast LSTM on pinch forecasted upwind information (dynamic forecast features) to make early predictions. One twelvemonth of humanities upwind information are input into nan hindcast LSTM, and 7 days of forecasted upwind information are input into nan forecast LSTM. Static features see geographical and geophysical characteristics of watersheds that are input into some nan hindcast and forecast LSTMs and let nan exemplary to study different hydrological behaviors and responses successful various types of watersheds.

Output from nan forecast LSTM is fed into a “head” furniture that uses mixture density networks to nutrient a probabilistic forecast (i.e., predicted parameters of a probability distribution complete streamflow). Specifically, nan exemplary predicts nan parameters of a substance of heavy-tailed probability density functions, called asymmetric Laplacian distributions, astatine each forecast clip step. The consequence is simply a substance density function, called a Countable Mixture of Asymmetric Laplacians (CMAL) distribution, which represents a probabilistic prediction of nan volumetric travel complaint successful a peculiar stream astatine a peculiar time.

LSTM-based stream forecast exemplary architecture. Two LSTMs are applied successful sequence, 1 ingesting humanities upwind information and 1 ingesting forecasted upwind data. The exemplary outputs are nan parameters of a probability distribution complete streamflow astatine each forecasted timestep.

Input and training data

The exemplary uses 3 types of publically disposable information inputs, mostly from governmental sources:

1. Static watershed attributes representing geographical and geophysical variables: From nan HydroATLAS project, including information for illustration semipermanent ambiance indexes (precipitation, temperature, snowfall fractions), onshore cover, and anthropogenic attributes (e.g., a nighttime lights scale arsenic a proxy for quality development).
2. Historical meteorological time-series data: Used to rotation up nan exemplary for 1 twelvemonth anterior to nan rumor clip of a forecast. The information comes from NASA IMERG, NOAA CPC Global Unified Gauge-Based Analysis of Daily Precipitation, and nan ECMWF ERA5-land reanalysis. Variables see regular full precipitation, aerial temperature, star and thermal radiation, snowfall, and aboveground pressure.
3. Forecasted meteorological clip bid complete a seven-day forecast horizon: Used arsenic input for nan forecast LSTM. These information are nan aforesaid meteorological variables listed above, and travel from nan ECMWF HRES atmospheric model.

Training information are regular streamflow values from nan Global Runoff Data Center complete nan clip play 1980 - 2023. A azygous streamflow forecast exemplary is trained utilizing information from 5,680 divers watershed streamflow gauges (shown below) to amended accuracy.

Location of 5,680 streamflow gauges that proviso training information for nan stream forecast exemplary from nan Global Runoff Data Center.

Improving connected nan existent state-of-the-art

We compared our stream forecast exemplary pinch GloFAS type 4, nan existent state-of-the-art world flood forecasting system. These experiments showed that ML tin supply meticulous warnings earlier and complete larger and much impactful events.

The fig beneath shows nan distribution of F1 scores erstwhile predicting different severity events astatine stream locations astir nan world, pinch positive aliases minus 1 time accuracy. F1 scores are an mean of precision and callback and arena severity is measured by return period. For example, a 2-year return play arena is simply a measurement of streamflow that is expected to beryllium exceeded connected mean erstwhile each 2 years. Our exemplary achieves reliability scores astatine up to 4-day aliases 5-day lead times that are akin to aliases better, connected average, than nan reliability of GloFAS nowcasts (0-day lead time).

Distributions of F1 scores complete 2-year return play events successful 2,092 watersheds globally during nan clip play 2014-2023 from GloFAS (blue) and our exemplary (orange) astatine different lead times. On average, our exemplary is statistically arsenic meticulous arsenic GloFAS nowcasts (0–day lead time) up to 5 days successful beforehand complete 2-year (shown) and 1-year, 5-year, and 10-year events (not shown).

Additionally (not shown), our exemplary achieves accuracies complete larger and rarer utmost events, pinch precision and callback scores complete 5-year return play events that are akin to aliases amended than GloFAS accuracies complete 1-year return play events. See nan paper for much information.

Looking into nan future

The flood forecasting inaugural is portion of our Adaptation and Resilience efforts and reflects Google's commitment to reside ambiance change while helping world communities go much resilient. We judge that AI and ML will proceed to play a captious domiciled successful helping beforehand subject and investigation towards ambiance action.

We actively collaborate pinch respective world assistance organizations (e.g., nan Centre for Humanitarian Data and nan Red Cross) to supply actionable flood forecasts. Additionally, successful an ongoing collaboration pinch nan World Meteorological Organization (WMO) to support early informing systems for ambiance hazards, we are conducting a study to thief understand really AI tin thief reside real-world challenges faced by nationalist flood forecasting agencies.

While nan activity presented present demonstrates a important measurement guardant successful flood forecasting, early activity is needed to further grow flood forecasting sum to much locations globally and different types of flood-related events and disasters, including flash floods and municipality floods. We are looking guardant to continuing collaborations pinch our partners successful nan world and master communities, section governments and nan manufacture to scope these goals.