SEED-FD improves flood and drought forecasting by using artificial intelligence (AI) to reduce prediction errors. In a three-step process, the project enhances accuracy and reliability, strengthening forecasting systems for decision-makers and communities worldwide.
Why Forecasts Have Errors – and Why Reducing Them Matters
Even though forecasts for floods and droughts aim for maximum reliability, they always carry some level of uncertainty. This can arise from errors in input data, limitations in hydrological models, or unexpected environmental changes. Such errors may lead to false alarms or missed events – both of which can have serious real-world consequences. In SEED-FD, researchers from the European Centre for Medium-Range Weather Forecasts (ECMWF) are working on advanced AI-based post-processing methods to correct these errors and improve forecast accuracy.
Step 1 – Benchmarking with Quantile Mapping
In the first step, a statistical method called quantile mapping was used to set a benchmark. Quantile mapping adjusts forecasts to match observations, helping to correct systematic biases in the model – such as consistent over- or underpredictions.
As this method is straightforward and data-efficient, it serves as a useful starting point for improvement. However, it also has clear limitations: it cannot easily combine different types of data, it relies on local measurements and therefore cannot be used in regions without observation stations, and it typically does not incorporate Earth Observation data from satellites. It also cannot address dynamic errors, meaning those forecast inaccuracies that vary between forecasts. For these reasons, SEED-FD builds on this baseline by applying AI-based methods in the next steps.
Step 2 – Correcting Forecasts with LSTM Networks
To go beyond static corrections, SEED-FD uses a Long Short-Term Memory (LSTM) network framework – a type of deep learning algorithm designed to detect patterns in time series data.
The LSTM is trained on historical datasets that combine simulations of the past with the corresponding observed conditions. This form of training on reanalysis data enables the algorithm to identify typical errors and how they develop over time. Using the LSTM can reduce dynamic errors in future forecasts.
SEED-FD is currently working on this step, refining the LSTM-based approach in preparation for broader application.
Step 3 – Enriching Forecasts with Additional Data
The third step in SEED-FD’s approach integrates additional data sources to further improve forecast accuracy. Among these are Earth Observation (EO) data from satellites, which provide valuable information on variables such as precipitation, river discharge and water levels, and soil moisture – particularly valuable in regions where ground-based measurements are limited or unavailable.
To process these data, a two-stage LSTM framework is used:
- LSTM 1 processes historical data to improve the initial conditions of the forecast. It passes on its internal state to the second LSTM.
- LSTM 2 focuses on correcting forecast errors dynamically, ensuring better accuracy over different lead times.
Additional inputs used in this approach include meteorological forcings, such as precipitation and temperature forecasts, and catchment attributes, including land use, soil types, topography, and vegetation. By accounting for how different catchments respond to similar weather conditions, the forecasting model becomes more adaptable and accurate.
Outlook: SEED-FD Using AI to Improve Forecast Error-Correction
By combining statistical and AI-based methods, SEED-FD is improving the accuracy of flood and drought forecasts. While quantile mapping provides an initial benchmark for correcting systematic errors, the two-stage LSTM approach is designed to correct both systematic and dynamic errors – even in regions with limited data availability.
The third step – integrating additional data sources – marks a significant leap in forecast quality. Compared to the initial benchmark, this data-enriched AI approach is expected to deliver more reliable forecasts and better detection of extreme events. It will be able to adapt more effectively to local conditions, making forecasts more relevant and actionable. These advancements will be validated in hydrologically diverse regions using the operational forecasts from GloFAS, the Copernicus Emergency Management Service global flood forecasting system.
A wide range of users will benefit from SEED-FD’s improvements: decision-makers and public authorities will be able to use forecasts more confidently; humanitarian organisations will be able to plan ahead with better information; businesses and consultants in sectors like agriculture, infrastructure, or insurance will be able to reduce risks; and communities around the world will be better prepared for the impacts of floods and droughts.
Stay tuned for more updates about how SEED-FD’s innovations are helping to strengthen resilience to floods and droughts worldwide.