What is the new urban flash flood prediction feature?
Google is expanding its flood forecasting capabilities with a new AI model to predict urban flash floods. This new feature will provide early warnings for rapid-onset floods in urban areas, helping individuals and authorities stay informed and safe. This is a new and separate model from the one Google uses for riverine flood predictions. While riverine models are trained on historical data from river gauges, the new urban flash flood model is designed to predict floods that can occur anywhere in an urban area, not just near a river.
How was the urban flash flood model trained?
The AI flash floods forecasting model processes a historical 7-day hindcast sequence of meteorological (and geophysical) input data and a forecast for the next 24 hours with inputs from meteorological forecasts. The model predicts, at each time step the probability of a flash flood occurring in a specific urban region within the next 24 hours, with spatial resolution of 20km x 20km.
Unlike riverine floods, robust historical data does not exist for flash floods, which may not occur near a river. To address the lack of historical data, for flash flood model training we constructed a global dataset of past flood events.
How is the urban flash flood model evaluated?
The urban flash flood model is evaluated on flood events from the dataset described above (with the appropriate cross-validation time and spatial splits). Due to the current limitations of our dataset, we limit our coverage to urban areas, or densely populated areas. Specifically, we only consider pixels with a population density of more than approximately 100 people per kilometer squared (which translates to roughly 40,000 population per pixel).
What are the inputs to the urban flash flood model?
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.
Meteorological data: We rely on a variety of publicly available weather products, including ECMWF forecasts, IMERG precipitation, NOAA CPC precipitation data, Google DeepMind’s medium-range global weather forecasting model.
All input data were area-weighted averaged over 20km x 20km pixels.
Where are Google’s urban flash flood predictions available?
Google’s urban flash flood model provides predictions for urban areas across numerous countries around the globe. Currently, within each country, the model is focused on predicting impact in urban areas (primarily cities) with population densities greater than 100 people per square kilometer. Google plans to expand this coverage to other areas beyond urban areas over time.
What are the limitations of Google's urban flash flood predictions?
It's important to be aware of the model's current limitations, the model currently can only predict urban flash floods caused by weather events, not those resulting from man-made incidents like dam or levee failures. In addition initially the focus is on urban areas. Google plans to expand coverage to other areas over time.