AI Boosts Reliability of Flood Forecasts

Introduction

Accurate and reliable flood forecasting is crucial for reducing disaster risks and adapting to climate change. Floods can have devastating consequences, causing loss of life, damage to property, and disruption to communities. To address this challenge, researchers have developed an advanced artificial intelligence (AI) model to improve short-term (0-7 day) forecasts of extreme riverine events.

Existing global flood forecasting systems, such as the

, have limitations in their ability to accurately predict high-flow events. The researchers saw an opportunity to leverage the power of AI and machine learning to enhance flood forecasting capabilities, potentially saving lives and property.

Research Purpose and Motivation

The study aimed to develop an AI-based model that could outperform the reliability of the GloFAS benchmark in predicting extreme riverine events. The researchers were motivated by the need to address the limitations of current global flood forecasting systems, which often struggle to provide accurate and timely warnings, especially in data-scarce regions.

By harnessing the potential of AI and machine learning, the researchers hoped to improve the precision and recall of flood forecasts, extending the reliability of global nowcasts from 0 to 5 days and enhancing the skill of forecasts in regions like Africa to match the level currently available in Europe.

Methodology and Study Design

The researchers developed an AI model using

, a type of

recurrent neural network

A type of artificial neural network that is designed to process sequential data, such as text or time series, by taking into account the relationships between elements in the sequence.

 

well-suited for processing sequential data, such as time series of streamflow. The model was trained and tested on a comprehensive dataset spanning 152,259 years of data from 5,680 watersheds around the world.

The input data for the model came from various sources, including the European Centre for Medium-Range Weather Forecasts (ECMWF), the National Oceanic and Atmospheric Administration (NOAA), and the National Aeronautics and Space Administration (NASA). The researchers employed rigorous cross-validation techniques to split the data in both time and space, ensuring that the model's performance was evaluated on unseen data to avoid data leakage.

The AI model's performance was then compared to the GloFAS global flood forecasting system. The researchers assessed the models' reliability by evaluating their precision and recall in predicting high-flow events defined by

of 1, 2, 5, and 10 years. They also compared the models' performance using standard

hydrograph metrics

Measurements that describe the shape and characteristics of a graph showing the flow of water in a river or stream over time.

 

, such as bias,

Nash-Sutcliffe efficiency

A statistic that measures how well a model's predictions match the observed data, with a value of 1 indicating a perfect match.

 

, and

Kling-Gupta efficiency

A statistic that evaluates the performance of a model by considering its ability to match the observed data in terms of correlation, variability, and bias.

 

, in both

gauged and ungauged basins

Gauged basins are areas where there are measurements of water flow, while ungauged basins are areas where there is no direct measurement of water flow.

 

.

Results and Significance

The results of the study showed that the AI model demonstrated improved reliability compared to the GloFAS benchmark. The AI model had higher precision and recall scores across all return period events, indicating that it was better at correctly identifying and predicting high-flow events.

Specifically, the AI model outperformed or matched the reliability of GloFAS forecasts up to a 5-day lead time for 1-year, 2-year, and 5-year return period events. This means that the AI model was able to provide more accurate and reliable flood warnings in the short-term, which is crucial for disaster risk reduction and climate change adaptation efforts.

However, the researchers also found that both the AI model and GloFAS exhibited regional differences in performance. The AI model showed a 54% difference in mean

(a metric that combines precision and recall) between the lowest-scoring continent (South America) and the highest-scoring continent (Southwest Pacific) for 5-year return period events.

The study also explored the challenge of predicting which forecasting model will perform better in a given location based on

. The researchers found that it is difficult to use catchment attributes to reliably predict model performance, but they were able to develop models that can predict, with reasonable accuracy, whether a particular forecasting model will perform well or poorly in a given watershed.

The key factors that determine model performance are related to aridity (as measured by

) and basin size, with the AI-based model generally performing better in smaller, less arid basins. This suggests ways that machine-learning-based streamflow modeling could be improved, such as by focusing training on larger basins or implementing explicit routing or graph models to better capture the hydrology of smaller watersheds.

Conclusions and Implications

The researchers have made significant strides in leveraging AI and open datasets to enhance short-term flood forecasting and early warning systems. By developing an advanced AI-based streamflow forecasting model, they have demonstrated the potential for AI to improve the reliability and accuracy of global flood predictions, especially in data-scarce regions.

To ensure timely dissemination of these accurate flood warnings, the researchers have made the forecasts publicly available in real-time through various channels, including the

, push alerts to smartphones, and an open online portal, without any cost or access barriers. This is a crucial step in protecting the lives and property of millions of people worldwide who are at risk of flooding.

The study's key contributions include the development of a robust and scalable deep learning-based streamflow forecasting model, the public release of the dataset and trained models to support further research and development, and the demonstration of the potential for AI-based approaches to improve global-scale flood forecasting.

The researchers emphasize the importance of continued research and collaboration to address the "grand challenges" in big data and Earth system science, with the ultimate goal of enhancing disaster risk reduction and climate change adaptation efforts worldwide. By leveraging the power of AI and open data, the scientific community can work towards more accurate and reliable flood forecasting, ultimately saving lives and protecting communities from the devastating impacts of extreme weather events.