H216-04
Google’s flood forecasting framework: connecting data, models, risks and people
Wednesday, 16 December 2020: 20:42
Virtual
Sella Nevo1, Adi Gerzi Rosenthal1, Asher Metzger1, Dana Weitzner1, Efrat Morin2,3, Gal Elidan1,4, Gregory Begelman1, Guy Shalev1, Hila Noga1, Ira Shavitt1, Liora Yuklea5, Moriah Royz1, Niv Giladi1, Nofar Peled Levi6, Ofir Reich1, Oren Gilon1, Tal Shechter1, Vladimir Anisimov1, Yotam Gigi1, Zach Moshe1, Zvika Ben-Haim1, Avinatan Hassidim1 and Yossi Matias6, (1)Google, Research, Tel Aviv, Israel, (2)The Hebrew University of Jerusalem, The Fredy and Nadine Herrmann Institute of Earth Sciences, Jerusalem, Israel, (3)Google, Research, Mountain View, United States, (4)Hebrew University of Jerusalem, Department of Statistics, Jerusalem, Israel, (5)Google, Search, Tel Aviv, Israel, (6)Google, Tel Aviv, Israel
Abstract:
One of the major natural disasters is floods, which cause thousands of fatalities, affects the lives of hundreds of millions, and results in huge economic damages annually. Among the main challenges for an accurate and reliable global flood forecasting system are computational costs, handling diverse (and sometimes low quality) data, and reaching the affected individual with meaningful warnings. Google is well positioned to address these challenges. Accordingly, Google’s flood forecasting initiative aims at providing high-resolution, high accuracy, flood forecasts and timely warnings around the globe, while focusing first on developing countries where most of the fatalities occur.
The framework developed to achieve this goal is composed of several components, including data processing, machine learning based hydrologic models, machine learning and physics based flood inundation models, risk assessment, and a flood warning system that issues alerts to affected regions. The system utilizes state of the art computational resources (such as TPUs) and data (e.g., novel high resolution DEM). It is currently operational during the monsoon season in large rivers in India and Bangladesh. In the near future it is expected to grow and cover more flood-prone regions in these countries and others across the globe.
In this presentation, we describe the full flood forecasting framework, evaluate its performance according to different measures, considering also social impacts. We highlight special research aspects of interest:
- Machine learning techniques for hydrologic and flood inundation models
- Data and models uncertainties and their effect on flood forecasts
- Forecast horizon: tradeoffs between accuracy and relevance
- Flood alerts: how to make them understood and effective and methods to evaluate their effectiveness
