Directing ML towards pure hazard mitigation via collaboration – Google AI Weblog

Floods are the most typical sort of pure catastrophe, affecting greater than 250 million people globally annually. As a part of Google’s Crisis Response and our efforts to address the climate crisis, we’re utilizing machine studying (ML) fashions for Flood Forecasting to alert folks in areas which might be impacted earlier than catastrophe strikes.

Collaboration between researchers within the business and academia is crucial for accelerating progress in direction of mutual targets in ML-related analysis. Certainly, Google’s present ML-based flood forecasting approach was developed in collaboration with researchers (1, 2) on the Johannes Kepler College in Vienna, Austria, the College of Alabama, and the Hebrew College of Jerusalem, amongst others.

Right this moment we talk about our current Machine Studying Meets Flood Forecasting Workshop, which highlights efforts to convey collectively researchers from Google and different universities and organizations to advance our understanding of flood conduct and prediction, and construct extra sturdy options for early detection and warning. We additionally talk about the Caravan venture, which helps to create an open-source repository for international streamflow information, and is itself an instance of a collaboration that developed from the earlier Flood Forecasting Meets Machine Studying Workshop.

2023 Machine Studying Meets Flood Forecasting Workshop

The fourth annual Google Machine Studying Meets Flood Forecasting Workshop was held in January. This 2-day digital workshop hosted over 100 members from 32 universities, 20 governmental and non-governmental businesses, and 11 personal corporations. This discussion board supplied a chance for hydrologists, pc scientists, and assist staff to debate challenges and efforts towards enhancing international flood forecasts, to maintain up with state-of-the-art know-how advances, and to combine area data into ML-based forecasting approaches.

The occasion included talks from six invited audio system, a collection of small-group dialogue classes centered on hydrological modeling, inundation mapping, and hazard alerting–associated subjects, in addition to a presentation by Google on the FloodHub, which gives free, public entry to Google’s flood forecasts, as much as 7 days upfront.

Invited audio system on the workshop included:

The shows will be considered on YouTube:

2023 Flood Forecasting Meets Machine Studying Talks Day 1

2023 Flood Forecasting Meets Machine Studying Talks Day 2

Among the prime challenges highlighted throughout the workshop have been associated to the combination of bodily and hydrological science with ML to assist construct belief and reliability; filling gaps in observations of inundated areas with fashions and satellite tv for pc information; measuring the talent and reliability of flood warning techniques; and enhancing the communication of flood warnings to various, international populations. As well as, members harassed that addressing these and different challenges would require collaboration between a lot of totally different organizations and scientific disciplines.

The Caravan venture

One of many major challenges in conducting profitable ML analysis and creating superior instruments for flood forecasting is the necessity for big quantities of information for computationally costly coaching and analysis. Right this moment, many nations and organizations acquire streamflow data (sometimes both water ranges or stream charges), however it isn’t standardized or held in a central repository, which makes it troublesome for researchers to entry.

Through the 2019 Machine Learning Meets Flood Forecasting Workshop, a gaggle of researchers recognized the necessity for an open supply, international streamflow information repository, and developed concepts round leveraging free computational assets from Google Earth Engine to handle the flood forecasting group’s problem of information assortment and accessibility. Following two years of collaborative work between researchers from Google, the college of Geography on the College of Exeter, the Institute for Machine Studying at Johannes Kepler College, and the Institute for Atmospheric and Local weather Science at ETH Zurich, the Caravan project was created.

In “Caravan – A global community dataset for large-sample hydrology”, printed in Nature Scientific Data, we describe the venture in additional element. Primarily based on a world dataset for the event and coaching of hydrological fashions (see determine beneath), Caravan gives open-source Python scripts that leverage important climate and geographical information that was beforehand made public on Google Earth Engine to match streamflow information that customers add to the repository. This repository initially contained information from greater than 13,000 watersheds in Central Europe, Brazil, Chile, Australia, the USA, Canada, and Mexico. It has additional benefited from community contributions from the Geological Survey of Denmark and Greenland that features streamflow information from many of the watersheds in Denmark. The aim is to proceed to develop and develop this repository to allow researchers to entry many of the world’s streamflow information. For extra data concerning contributing to the Caravan dataset, attain out to caravan@google.com.

Areas of the 13,000 streamflow gauges within the Caravan dataset and the distribution of these gauges in GEnS international local weather zones.

The trail ahead

Google plans to proceed to host these workshops to assist broaden and deepen collaboration between business and academia within the growth of environmental AI fashions. We’re trying ahead to seeing what advances may come out of the latest workshop. Hydrologists and researchers keen on collaborating in future workshops are inspired to contact flood-forecasting-meets-ml@google.com.