[Paper Review] ML-based Flood Forecasting: Advances in Scale, Accuracy and Reach

Sella Nevo, Gal Elidan+5 more|arXiv (Cornell University)|Nov 29, 2020

Flood Risk Assessment and Management9 references8 citations

TL;DR

This paper presents two machine learning-based approaches to scale global flood forecasting: water level-based hydrologic models that bypass the need for discharge data, and a morphological flood modeling method that combines physics and ML to reduce calibration effort and computational costs. The methods achieved high accuracy (R² = 0.99, precision 76.2%) and enabled real-time warnings for 75 million people in India and Bangladesh, significantly expanding access to actionable flood alerts.

ABSTRACT

Floods are among the most common and deadly natural disasters in the world, and flood warning systems have been shown to be effective in reducing harm. Yet the majority of the world's vulnerable population does not have access to reliable and actionable warning systems, due to core challenges in scalability, computational costs, and data availability. In this paper we present two components of flood forecasting systems which were developed over the past year, providing access to these critical systems to 75 million people who didn't have this access before.

Motivation & Objective

• Address the global scalability gap in flood warning systems due to high calibration costs, computational demands, and data scarcity.
• Overcome the limitation of relying on rare discharge measurements by developing models that use widely available water level (stage) data.
• Reduce manual calibration and computational costs in flood inundation modeling while maintaining or improving accuracy.
• Enable deployment of operational flood forecasting systems in data-scarce regions, particularly in South Asia.
• Scale flood warning systems to cover large geographic areas with minimal infrastructure and data requirements.

Proposed method

• Train machine learning models on 72-hour historical water level data from 52 gauges in India and Bangladesh to predict future water levels.
• Use regression models trained on non-public streamflow data from 2014–2019 to forecast water levels during the 2020 monsoon season.
• Develop a morphological flood modeling approach that infers river profiles from gauge measurements, elevation data, and historical SAR-derived inundation maps.
• Apply monotonicity and continuity constraints to learn a continuous, decreasing function from gauge measurement to river profile across the river network.
• Use a local search algorithm to optimize river profiles by minimizing over- or under-flooding relative to historical flood extents.
• Generate inundation maps efficiently using a flood-fill algorithm on the learned synthetic morphology, replacing computationally expensive finite-element hydrodynamic simulations.

Experimental results

Research questions

• RQ1 Can machine learning models trained on water level data alone achieve high-accuracy flood forecasts in regions without discharge measurements?
• RQ2 Can a hybrid physics-ML morphological model reduce manual calibration and computational costs compared to traditional finite-element hydrodynamic models?
• RQ3 How does the performance of the morphological model compare to classic hydrodynamic models in data-scarce regions with limited field data?
• RQ4 To what extent can learned river profiles generalize across different river systems and flood events?
• RQ5 Can these methods be deployed operationally at scale to deliver timely, accurate flood warnings to vulnerable populations?

Key findings

• The water level-based models achieved an average lead time of 20.7 hours and an R² of 0.99 on water level predictions across 52 gauges in India and Bangladesh.
• The deployed models had a mean absolute error of 0.067 meters and a mean squared error of 0.011 meters², indicating high predictive accuracy.
• The morphological model outperformed the classic hydrodynamic model in precision (75.8% vs. 72.7%) and reduced manual calibration effort from 30 to 4 hours per site.
• The morphological model reduced CPU costs from 4–131 CPU years to 100–1,200 CPU hours per site, enabling broader scalability.
• In real-time operations, the system achieved 76.2% precision and 77.6% recall at 64-meter resolution across 429 flood events during the 2020 monsoon season.
• The system successfully delivered over 39 million flood alerts to individuals and authorities in India and Bangladesh, demonstrating real-world impact.