2024-03-21: Google Predicts Floods

on machine learning improving the emerging world

Author

Ate A Pi
March 20, 2024

šŸ”· Subscribe to get breakdowns of the most important developments in AI in your inbox every morning.

Hereā€™s today at a glance:

ā˜” Google Flood Precognition

This amazing paper from Google reveals that their flood predictions model is on par with a state-of-the-art nowcastā€¦ except a full day five days earlier.

Here we show that artificial intelligence-based forecasting achieves reliability in predicting extreme riverine events in ungauged watersheds at up to a five-day lead time that is similar to or better than the reliability of nowcasts (zero-day lead time) from a current state-of-the-art global modelling system (the Copernicus Emergency Management Service Global Flood Awareness System). In addition, we achieve accuracies over five-year return period events that are similar to or better than current accuracies over one-year return period events.

In essence:

  • Hydrological models take in weather data like precipitation and physical watershed formation

  • But, they also need to be calibrated to long data records from streamflow gauging stations in individual rivers

  • These are expensive, which creates huge issues as the most flood-insecure countries are also the poorest

  • So the countries most likely to have disastrous floods are also most likely to have flood prediction models that donā€™t work

  • Googleā€™s machine learning model is a train once, use anywhere model, which can make predictions for rivers where no data is available

Namely the model only uses some broad aggregate globally available data along with 7 day weather forecast:

The model uses three types of publicly available data inputs, mostly from governmental sources:

1. tatic watershed attributes representing geographical and geophysical variables: From the HydroATLAS project, including data like long-term climate indexes (precipitation, temperature, snow fractions), land cover, and anthropogenic attributes (e.g., a nighttime lights index as a proxy for human development).

2. Historical meteorological time-series data: Used to spin up the model for one year prior to the issue time of a forecast. The data comes from NASA IMERG, NOAA CPC Global Unified Gauge-Based Analysis of Daily Precipitation, and the ECMWF ERA5-land reanalysis. Variables include daily total precipitation, air temperature, solar and thermal radiation, snowfall, and surface pressure.

3. Forecasted meteorological time series over a seven-day forecast horizon: Used as input for the forecast LSTM. These data are the same meteorological variables listed above, and come from the ECMWF HRES atmospheric model.

The model is better than state-of-the-art up to 5 days in advance, but F-scores on SOTA are only in the 0.25 range (this is a hard problem!)

F-scores: Blue is SOTA. Orange is days days-ahead prediction.

What strikes me as beautiful is that theyā€™ve managed to extend the model to use less precise data. Because data is expensive!

šŸŒ  Enjoying this edition of Emergent Behavior? Send this web link with a friend to help spread the word of technological progress and positive AI to the world!

Or send them the below subscription link:

šŸ—žļø Things Happen

  • Deepmind releases a football manager. Football is a game of territorial aggression, so this is, of course, a future warmaking tool.

šŸ–¼ļø AI Artwork Of The Day

Thatā€™s it for today! Become a subscriber for daily breakdowns of whatā€™s happening in the AI world:

Reply

or to participate.