Google now ready to give public ‘flood alerts’

In the pilot, implemented in partnership with the Central Water Commission in India, Google showed, via Public Alerts, a map that included areas designated as ‘high risk’ ‘medium risk’ and ‘low risk’. The pilot used an operational hydro-dynamic model, with the explicit goal of preparing the ground for integrating Machine Learning (ML) models into the process. Alerts were then sent out to individuals in the catchment area in the form of maps and Android notifications.

Patna was chosen for the pilot given its large population and one of the most frequent occurrences of severe riverine flooding. “After we sent out the alerts, this allowed us to look into how people experience and interact with the alerts we send. We’ve realised that many people prefer text that describes the same information our maps show,” Sella Nevo from Google’s Research and Machine Intelligence team said. “This allowed us to be relatively confident that our pilot will also be informative as we prepare to scale our efforts up… Our alerts, though limited in scope for our first pilot, achieved results we’re happy with,” Nevo said.

There are a number of new technologies and methods for the creation of flood forecast maps. For example, ‘LDAR’ (light detection and ranging) technology, which has the ability to get accurate elevations throughout the floodplain quickly and affordably, and TRIMR2D (Transient Inundation Model for Rivers 2 Dimensional), a numerical computer program, that can simulate flood flows across the floodplain and several miles downstream from the forecast point. Then, there is a spatial analysis software (GIS) that turns the model results into maps and overlays them on other maps, like a map of a locality, or even onto an aerial shot.

Effective riverine flood forecasting at scale, though, is hindered by a number of factors, such as the need to rely on human calibration in current methodology, the limited amount of data for a specific location, and the computational difficulty of building continent level models that are sufficiently accurate.

A Union Ministry for Water Resources official said the Government hopes the collaboration with Google will help in flood management efforts. “The initiative (with Google) could assist crisis management agencies to deal with extreme hydrological events in a better manner,” the official said. Google says its ML-based modeling provides a wide range of improvements over the traditional physics-based models. ML is primed to be advantageous in this scenario, with some models often exceeding human experts in complex high-dimensional scenarios, and the framework of transfer or multitask learning is an appealing solution for leveraging local signals to achieve improved global performance.”

“First, it can enable incorporating additional types of data, details and nuances, that are either neglected by physics-based models or modeled inaccurately. Second, it enables much more efficient models, critical for scaling this effort across India and eventually globally. Finally, it enables automating many of the processes that need to be implemented manually using existing methods – which drastically reduces costs and allows us to scale,” Nevo said.
In September last year, during the Kerala floods, a joint team of researchers from the Michigan Technological University and University of Kerala had come out with an inundation map using satellite images and data from the European Space Agency’s radar satellite Sentinel to map the inundated areas in Kerala during and following the floods. The team followed up the mapping with a field visit to validate the data.