The Devastating Impact of Floods
Floods are the most prevalent natural disasters, and their occurrence rate has more than doubled since 2000. This increase is mainly due to the accelerated hydrological cycle caused by human – induced climate change. Developing countries bear the brunt of flood impacts as their populations are highly vulnerable. The severe consequences of floods highlight the urgent need for accurate and timely flood warnings to protect human lives and property.
The Current State of Flood Forecasting
Currently, flood forecasting faces significant challenges, especially in ungauged basins. Hydrological prediction models lack reliable data for calibration in such areas, which limits the accuracy and lead – time of flood forecasts. The absence of dense streamflow gauge networks in developing countries further compounds the inaccuracy of flood warnings, emphasizing the need for improved global access to reliable flood forecasts.
A Ray of Hope: Google AI to the Rescue
Google’s Artificial Intelligence (AI) offers a promising solution to the global flood forecasting challenges. By utilizing AI and open datasets, there is a potential to greatly enhance the precision, recall, and lead – time of short – term forecasts for extreme riverine events. The development of an operational system that provides real – time, publicly available forecasts in over 80 countries shows the potential of AI to issue early and accurate flood warnings in ungauged basins, marking a significant step forward in improving global access to reliable flood forecasts and early warning systems.
Google Research Paper: AI Revolutionizes Flood Forecasting
The Google research paper presents a major advancement in flood forecasting. It uses AI trained on open and public datasets to evaluate the potential of revolutionizing global access to forecasts of extreme events in international rivers. An operational system has been developed to produce 7 – day short – term flood forecasts in over 80 countries, offering real – time forecasts without any access barriers such as fees or website registration.
Using AI for Global Flood Forecasts
The paper delves into the use of AI for global flood forecasts, introducing an AI streamflow forecasting model. This model builds on previous work on hydrological nowcast models and uses long short – term memory (LSTM) networks to predict daily streamflow over a 7 – day forecast horizon. Notably, it doesn’t use streamflow data as inputs, solving the real – time data availability problem, especially in ungauged locations. The model architecture includes an encoder – decoder model with separate LSTM units for historical and forecast meteorological input data.
From Open Data to Real – Time Forecasts
The operational system based on the AI model provides real – time flood forecasts in over 80 countries, a significant milestone in improving global access to reliable flood warnings. Its ability to produce short – term forecasts without access barriers, being available in real – time and free of charge, emphasizes the potential of AI to enhance flood early warning systems.
Beyond State – of – the – Art
The performance of the AI model outperforms the current state – of – the – art global modeling system, the Copernicus Emergency Management Service Global Flood Awareness System (GloFAS). The AI – based forecasting can predict extreme riverine events in ungauged watersheds with reliability up to five days in advance, which is comparable to or better than GloFAS nowcasts. Also, its accuracy for five – year return period events is similar to or better than current accuracies for one – year return period events, showing its potential in providing early and accurate flood warnings for larger and more impactful events in ungauged basins.
Under the Hood: The AI Model
Building the Brains
The AI streamflow forecasting model extends previous work on hydrological nowcast models using LSTM networks. It has an encoder – decoder architecture with one LSTM for historical meteorological data and another for the 7 – day forecast horizon with meteorological forecast inputs. Due to the lack of real – time streamflow data in ungauged locations, the model doesn’t use it as an input, and the benchmark (GloFAS) doesn’t use autoregressive inputs. The dataset is large, with model inputs and streamflow targets for 152,259 years from 5,680 watersheds, totaling 60 GB on disk.
The Data Timeline
The available data periods from each source for training and prediction with the AI model are shown in a figure. During training, missing data was handled by using similar variables from other sources or imputing with a mean value and adding a binary flag. The model uses a hindcast sequence length of 365 days and has a hidden size of 256 cell states for both encoder and decoder LSTMs.
How Well Does the AI Model Predict?
The performance of the AI model was evaluated through cross – validation experiments, splitting data from 5,680 gauges in time and space for out – of – sample predictions. The model predicts parameters of a single asymmetric Laplacian distribution over area – normalized streamflow discharge at each time step and forecast lead time. It was trained on 50,000 minibatches with a batch size of 256 and standardized inputs.
Putting the Model to the Test
The cross – validation experiments included splits across continents, climate zones, and hydrologically separated watershed groups. The AI model was evaluated out – of – sample in both location and time, and the results were reported over a hydrograph from averaging predicted hydrographs of three separately trained encoder – decoder LSTMs.
Evaluating the Model with Hydrograph Metrics
Hydrograph metrics for the AI model and GloFAS were assessed, with scores decreasing as the lead time increased. Results were calculated for 2014 – 2021 and listed in Extended Data Table 1. Metrics were also evaluated for the 1,144 gauges where GloFAS is calibrated, with similar trends.
What Makes the AI Tick?
Feature importance rankings from reliability classifiers were used to identify geophysical attributes that determine high or low reliability in the AI model. The most important features include drainage area, mean annual potential evapotranspiration (PET), mean annual actual evapotranspiration (AET), and elevation, which are correlated with reliability scores, indicating high nonlinearity and parameter interaction in the model.
Conclusion
Although hydrological modeling has advanced, many flood – prone regions still lack reliable forecasting and early warning systems. The Google research paper shows that leveraging AI and open data can significantly improve short – term forecasts for extreme riverine events. AI – based forecasting extends the reliability of current global nowcasts to 5 days and improves forecast skills in Africa to European levels. Real – time, publicly available forecasts without access barriers enable timely flood warning dissemination. However, further improvement is possible by increasing access to hydrological data and real – time updates through open – source initiatives. Enhancing global flood predictions and early alerts is crucial for protecting millions from flood devastation, and the combination of AI, open data, and collaborative efforts can help achieve this goal.
