Google’s AI Research: Part 1
Google recently held their AI research event where they present on all of the findings from their AI researchers. Here’s a few of the key takeaways
Flood Notification Alerts in India and Pakistan
As part of their efforts to help during natural disasters Google has launched a new site called FloodHub which tracks flooding, and helps local first responders organize themselves and the affected communities.
It looks rather basic – just a layer filter on top of Google Maps but the power is in the notifications that can be sent to anyone with the Google Maps app.
Google’s FloodHub Overlay
For people that are traveling or simply aren’t aware of how much water has reached a particular area these app and search notifications can be incredibly useful. Google is able to not only release these notifications but provide a guess of the water-depth in a given area.
So is Google just pulling in local weather data from these locations and then figuring out how much precipitation has actually come down? Not quite. They use a four-stage machine learning model that validates the data, forecasts, models the inundation, and distributes alerts.
Google takes their satellite images and builds a topographical map of a given zone then models a variety of different scenarios. River morphology and hydrology are extremely difficult to accurately model and predict so they will get a lot of things wrong with these models, but that can be fixed through manual intervention.
While there are a lot of satellites zipping around the planet it obviously gets somewhat difficult to measure how much water has come down if there’s a bunch of clouds in the way! Google is still reliant on local authorities to advise them about the water level in different areas.
Once a flood is predicted alerts can be sent out – as we’ve seen with numerous natural disasters early warning is critical since those people who take action first avoid the worst of these disasters. When an emergency is declared you’ll often have an above average number of people all trying to exit at the same time on roads that are designed for a much smaller amount of traffic.
This technology is critical in parts of the world that frequently flood like India and Bangladesh. Most of our major civilizations sprang up around reliable flood plains like the Nile, Ganges, and Mekong. The full list of countries Google has expanded this service to is available here:
Brazil
Colombia
Sri Lanka
Burkina Faso
Cameroon
Chad
Congo
Ivory Coast
Ghana
Guinea
Malawi
Sierra Leone
Angola
South Sudan
Namibia
Liberia
South Africa
Sample Alert from Google FloodHub
Now how well do these machine learning models work in comparison to traditional physics-based models? Recent models show better prediction accuracy, regional generalization, and scalability. They even outperformed regionally calibrated conceptual models, and individualized models.
What Google does is collect data on precipitation measurements, forecasts, and real-life stream gauge measurements. If the river is already above a certain threshold level on a stream gauge, an alert is automatically sent out.
Once the data is collected and verified it can then be fed into a forecasting model which tries to estimate the stage of flooding in various points. If it’s determined that the stage is higher than the warning threshold, an inundation model is then created for that particular stage to model the flooding and the affected area. Two different predictive models are used to simulate the stage of flooding in a given area and then these predictions are fed into the inundation model. The inundation model can then create a map showing which areas are underwater, and which areas will be flooded next based on historic data.
General Algorithm for the machine learning model used by Google for flood forecasting
The thresholding model is capable of producing an inundation model, but not computing the water depth. This is where the manifold model comes in, providing an estimate of water depth. Once the model decides an area is predicted to be underwater it will automatically push alerts to anyone using Google in those areas.
The code used to develop this model can be found here: https://github.com/google-research/google-research/blob/master/flood_forecasting/Inundation_Models.ipynb
Full paper can be read here: https://hess.copernicus.org/articles/26/4013/2022/
Additional Readings from Google can be found here: https://blog.google/technology/ai/flood-forecasts-india-bangladesh/
https://ai.googleblog.com/2020/09/the-technology-behind-our-recent.html