Source-linked AI summary
Flood forecasting with machine learning models in an operational framework
Sella Nevo, Efrat Morin, Adi Gerzi Rosenthal, Asher Metzger, Chen Barshai, Dana Weitzner, Dafi Voloshin, Frederik Kratzert, Gal Elidan, Gideon Dror, Gregory Begelman, Grey Nearing, Guy Shalev, Hila Noga, Ira Shavitt, Liora Yuklea, Moriah Royz, Niv Giladi, Nofar Peled Levi, Ofir Reich, Oren Gilon, Ronnie Maor, Shahar Timnat, Tal Shechter, Vladimir Anisimov, Yotam Gigi, Yuval Levin, Zach Moshe, Zvika Ben-Haim, Avinatan Hassidim, Yossi Matias
TL;DR
Flood-warning systems need reliable operational evaluation, while machine-learning systems capable of timely and accurate warnings remain uncommon. This paper presents and evaluates Google's operational system, including machine-learning models for stage forecasting and inundation, and introduces the Manifold model. The LSTM outperformed the Linear model, the inundation models had similar extent performance, and the system delivered over 100 million alerts during the 2021 monsoon season.
Problem
Operational flood-warning systems are difficult to evaluate, and systems using machine-learning models as core components for timely, accurate warnings are rare.
Method
The paper presents Google's operational flood-warning system, evaluates machine-learning models for stage forecasting and inundation, and introduces the Manifold model for flood extent and depth.
Results
The LSTM model was significantly better than the Linear model, with Persistent-NSE medians of 0.67 versus 0.60, while over 100 million flood alerts were provided during monsoon 2021.
Takeaways & Limitations
The operational system was deployed in India and Bangladesh during the 2021 monsoon season and provided flood alerts to affected populations, authorities, and emergency organizations.
Takeaways & Limitations
Machine-learning models rely solely on data and require large, representative records to achieve sufficient performance.
Abstract
from arXiv · showhide
The operational flood forecasting system by Google was developed to provide accurate real-time flood warnings to agencies and the public, with a focus on riverine floods in large, gauged rivers. It became operational in 2018 and has since expanded geographically. This forecasting system consists of four subsystems: data validation, stage forecasting, inundation modeling, and alert distribution. Machine learning is used for two of the subsystems. Stage forecasting is modeled with the Long Short-Term Memory (LSTM) networks and the Linear models. Flood inundation is computed with the Thresholding and the Manifold models, where the former computes inundation extent and the latter computes both inundation extent and depth. The Manifold model, presented here for the first time, provides a machine-learning alternative to hydraulic modeling of flood inundation. When evaluated on historical data, all models achieve sufficiently high-performance metrics for operational use. The LSTM showed higher skills than the Linear model, while the Thresholding and Manifold models achieved similar performance metrics for modeling inundation extent. During the 2021 monsoon season, the flood warning system was operational in India and Bangladesh, covering flood-prone regions around rivers with a total area of 287,000 km2, home to more than 350M people. More than 100M flood alerts were sent to affected populations, to relevant authorities, and to emergency organizations. Current and future work on the system includes extending coverage to additional flood-prone locations, as well as improving modeling capabilities and accuracy.
2 Hebrew University of Jerusalem, Jerusalem, Israel
The passage is a fragment concerning capabilities and accuracy.
- The passage mentions capabilities.
- The passage mentions accuracy.
- The passage links capabilities with accuracy.
Key-words:
The keywords identify the paper's focus on riverine flooding, operational warning systems, machine learning, and flood modeling in India and Bangladesh.
- The paper focuses on riverine floods.
- The paper concerns operational flood warning systems using machine learning for stage forecasting and flood inundation.
- India and Bangladesh are identified as geographic contexts.
1. Introduction
Floods pose substantial risks, while operational warning systems remain difficult to evaluate and systems using machine learning as core components are uncommon. This study presents and evaluates Google's operational system in India and Bangladesh, including machine-learning methods for forecasting and inundation modeling.
- Floods cause fatalities and extensive economic damage, making operational warning systems important for reducing risks and damages.
- Evaluation studies of operational flood warning systems are scarce, partly because ground-truth data are difficult to access.
- Operational systems using machine-learning models as core components and producing timely, accurate warnings are rare.
- Google's system expanded from limited initial deployment in India to forecasts covering most of India and Bangladesh during the 2021 monsoon seasons.
- The study presents the end-to-end operational system, evaluates machine-learning models for stage forecasting and inundation, and introduces the Manifold model for flood extent and depth.
- The study aims to reduce the scarcity of information on operational performance by assessing the utility of the presented machine-learning models.
2. End-to-end real-time operational flood warning system
Google’s operational system combines validated real-time observations, machine-learning stage forecasts, inundation modeling, and alert dissemination for large gauged rivers. It was deployed across India and Bangladesh during the 2021 monsoon, covering more than 287,000 km2 and sending approximately 115M notifications.
- System architecture: The system ingests, validates, corrects, and preprocesses stage and precipitation data before forecasting future river stages and high-resolution inundation.Validated inputs include near-real-time gauge measurements, precipitation observations, and precipitation forecasts.
- System architecture: Target gauges define warning thresholds, maximal forecast lead times, and surrounding areas of interest where inundation models are applied.The system is designed for large gauged rivers with relatively slow response and potentially complex inundation areas.
- Stage forecast modeling: Stage forecasting uses multiple linear regression and LSTM models to predict future water stages from gauge and precipitation inputs.The LSTM uses hindcast and forecast networks, while the Linear model is trained separately for each target gauge and lead time.
- Inundation modeling: The Manifold model provides a machine-learning alternative to hydraulic inundation modeling and can produce both inundation extent and depth.The Thresholding model provides inundation extent, whereas the Manifold model computes physically reasonable inundated depths in addition to the map.
- Model evaluation: Machine-learning models achieved higher accuracy than the hydraulic model, and LSTM models showed improved forecast skills compared with Linear models.These model comparisons supported operational use of the selected models.
3. Machine learning model evaluation
The evaluation examines four stage-forecast models and two machine-learning inundation models using historical gauge and flood-event data. Stage forecasting favors LSTM models and precipitation inputs, while Thresholding generally outperforms Manifold for inundation evaluation.
- Stage forecast models evaluation: Four stage-forecast models combine Linear or LSTM architectures with past water-stage inputs, with or without past precipitation.The models are evaluated using NSE and Persistent-NSE across 167 target gauges.
- Stage forecast models evaluation: NSE values are very high across target gauges and models, indicating generally very good predictive skill.These high scores are consistent with slow responses in large rivers and the use of historic and upstream water stages as inputs.
- Stage forecast models evaluation: Persistent-NSE is lower than NSE because the last observed value provides a stronger baseline than the mean observation.Persistent-NSE compares forecasting against the last observed stage, whereas NSE uses the mean-observation baseline.
- Stage forecast models evaluation: The LSTM model outperforms the Linear model, and precipitation improves both models; all comparisons are statistically significant.LSTM has a better Persistent-NSE score at 95% of gauges; precipitation improves LSTM at 89% and Linear at 65% of gauges.
- Stage forecast models evaluation: Stage-model metric values are positively correlated with watershed area, while the smallest watersheds have the greatest variance.The smallest basin bin is below 10,000 km2; its median does not significantly differ from the middle-size bin, whereas comparisons with the largest bin are significant.
- Inundation models evaluation: Thresholding achieves better F1 values than Manifold in most AOIs under both validation schemes, with significant paired differences.Thresholding is better in 70% of AOIs under 1-year cross-validation and 57% under leave-extreme-out validation.
4. Discussion
The discussion presents Google’s operational flood warning system, its model performance, practical deployment, communication strategies, and boundaries for broader application. Machine-learning models support scalable forecasting and inundation modeling, but coverage remains focused on large gauged rivers and representative data.
- Operational system: Google’s system was deployed in India and Bangladesh and evaluated stage-forecasting and inundation models across operational settings.Accuracy metrics were reported for 167 stage gauges and 126 inundation models in India and Bangladesh.
- Operational system: The 2021 monsoon deployment covered 287,000 km2 and sent approximately 115M flood alerts to affected populations.The system provided alerts at relatively high temporal and spatial resolution, including hourly forecasts and 16-meter inundation information.
- Model performance: The LSTM achieved good stage-forecast performance, with median NSE and Persistent-NSE values of 0.99 and 0.69, respectively.The reported F1 values for cross-validation and extreme-leave-out evaluation were 69% and 76%, respectively.
- Model performance: The present system’s stage forecasts had median root mean square errors of 0.1 meter at 24 hours and 0.2 meter at 48 hours.The comparison with Zalenski et al. is indicative rather than fully comparable because the analyzed watersheds differ.
- Alert communication: Effective warnings depend on communication choices such as local-language messages, visual explanations, and targeted information for specific locations.The discussion identifies warning receipt, comprehension, and decision use as operational challenges.
- Model performance: Machine-learning inundation models achieved higher accuracy than the physics-based hydraulic model while requiring less manual per-gauge work.Reduced calibration effort supports applying the models to new basins and expanding coverage.
- Scope and future work: Good performance was achieved with six to seven years of training data, although machine-learning models require large, representative records.Precipitation uncertainty may significantly affect forecasts, especially for proposed ungauged-basin and flash-flood extensions.
- Scope and future work: The system currently targets large gauged rivers, while ungauged basins and flash floods remain planned extensions with additional uncertainty and input-data challenges.Ungauged applications would rely more heavily on precipitation and could not optimize models directly to target basins; flash floods require faster response.
5. Conclusions
The operational system combines machine-learning stage forecasting and inundation mapping, and was deployed across India and Bangladesh during the 2021 monsoon. Its conclusions report stronger LSTM stage forecasts, operational use of both inundation models, and more than 100 million alerts.
- System scope: The system provides operational flood forecasts and warnings for large gauged rivers, where effective warning delivery remains a challenge.Operational flood warning systems can save lives and reduce risks and damages.
- Stage forecasting: The LSTM model was significantly better than the Linear model, with Persistent-NSE medians of 0.67 versus 0.60.LSTM forecasts were used operationally in 2021 for 165 of 167 target gauges.
- Stage forecasting: Past precipitation data slightly improved stage-forecast performance when added to the observed-stage inputs.Inputs included past target-gauge stage, upstream-gauge stage, and IMERG precipitation data.
- Inundation mapping: Thresholding and Manifold models were developed for inundation mapping; Manifold additionally forecasts water depth using DEM data.Water depth is described as crucial information for affected individuals.
- Inundation mapping: The Thresholding model outperformed the physics-based hydraulic model, required less manual effort and computation, and supported large-scale application.Its F1 distribution was close to Manifold’s, with medians of 69% and 76% in the leave-1-year and leave-extreme-out schemes.
- Operational deployment: Both inundation models were applied operationally during 2021, when the system sent over 100 million alerts to affected people and relevant agencies.Feedback indicated that local-language alerts and expected stage changes improved alert effectiveness.
- Future work: The literature lacks performance results from operational flood warning systems, motivating continued assessment and improvement of such systems.The authors identify extending coverage and improving modeling capabilities and accuracy as ongoing work.
Code availability
The software was developed at Google, and its code is proprietary.
- Code availability: Google developed the software used by the operational flood forecasting system.The passage identifies Google as the software developer.
- Code availability: The system’s source code is proprietary rather than publicly distributed.The passage states that the code is proprietary.
- Code availability: Code availability is therefore restricted to the proprietary software developed at Google.This restates the availability status reported for the system’s software.
Data availability
The study uses government-provided hourly stage data, NASA IMERG precipitation data, and publicly available Sentinel-1 flood-extent data.
- Stage data: Hourly stage data came from India’s Central Water Commission and Bangladesh’s Water Development Board.The passage identifies both agencies as data providers.
- Stage data: The stage data are government property and cannot be distributed.This availability restriction applies to the CWC and BWDB data.
- Precipitation and flood extent: Precipitation data came from NASA’s GPM IMERG Early product, while historical flood extents were derived from publicly available Sentinel-1 SAR GRD data on Google Earth Engine.The cited sources specify the IMERG product and the Earth Engine Sentinel-1 collection.
Supplementary:
The supplementary target-gauge table lists locations across Bangladesh and India together with forecast-model assignments and associated drainage-area and population information.
- Target gauge table: The supplementary material is organized as a target gauge table covering Bangladesh and India.Rows identify individual gauge locations and countries.
- Geographic attributes: The table reports approximated AOI area, population, and drainage-area values for individual target gauges, with N.A. where watershed association was unavailable.The table note explains that drainage areas were estimated from HydroSHEDS and population from WorldPop data.
- Model assignments: Inundation-model assignments in the table include Thresholding, Manifold, or None.Examples include Bangladesh rows using both Thresholding and Manifold and India rows using either model.