In late 2018, in collaboration with the Central Water Commission of India, Israel Institute of Technology, and Bar-Ilan University, Google piloted a flood-predicting model in portions of Patna, India as a part of its Flood Forecasting Initiative. The company and its partners claimed the model could predict riverine floods — that is, floods from overrun riverbanks — with 75% accuracy during monsoon seasons. And after expanding it last year to target more than 11,000 square kilometers along the Ganga and Brahmaputra rivers, Google today announced the model will inform alerts for all of India, covering 500,000 square kilometers and over 240 million people.
Floods are among the most common — and deadly — natural disasters in the world. Every year, they’re responsible for tens of thousands of fatalities and hundreds of millions of displaced homeowners. And they’re extraordinarily costly — in the U.S. alone from 2005 to 2014, the average flood claim was $42,000, and total flood insurance claims averaged more than $3.5 billion per year.
Accurate flood forecasting is a desirable goal, needless to say. According to some studies, early warning systems can reduce deaths and economic damages by over a third.
Google says its expanded model, which will launch outside of India for the first time in Bangladesh thanks to an agreement with the Bangladesh Water Development Board, will provide localized depth information including the expected water depth and timing of floods in particular villages and areas. Tens of millions of people will see the depth information in the course of pilots across eight districts in collaboration with local governments, Google’s philanthropic arm Google.org, and the International Federation of Red Cross and Red Crescent Societies, and they’ll benefit from a new, experimental model that can double the lead time of alerts from 24 hours to 48 hours.
Beyond the improved modeling, Google says it reached 30 million people in flood-affected areas with notifications and added better support for local languages with alerts that are more visual in nature. On Google Search and notifications sent to a network of volunteers with the nonprofit SEEDS who spread warnings to people without phones, flood alerts now support nine local Indic languages including Hindi and Bengali. They also contain infographics emphasizing how much water will rise, along with textual explanations and local maps.
One of the biggest challenges in building a flood prediction model is parameter calibration, an optimization process aimed at matching the algorithm’s predictions to certain baseline measurements. The standard approach involves significant manual work and often results in models that aren’t generalizable.
Google overcame a few of those barriers by drawing on real-time measurements and short-term forecasts of river water levels, from which their model generates an inundation map estimating the extent of the predicted flood. (Inundation maps show where flooding may occur over a range of water levels.) Leveraging correlated and aligned large map image batches adjusted to solve for coarse terrain, some from the European Space Agency Sentinel-1 satellite constellation, the model creates per-image depth maps that it then uses to produce elevation maps for locations.
Forecast methodologies like a newly developed approach to creating Digital Elevation Models (DEMs) optimize the model to work with Google’s custom tensor processing units (TPU), which supply predictions 85 times faster than with processors alone. Recently, machine learning replaced some of the model’s physics-based hydraulic algorithms, enabling it to support larger regions covering more people.