SEED-FD is enhancing flood and drought forecasting developing advanced artificial intelligence (AI) methods for post-processing. Quantile mapping serves as a first benchmark due to its simplicity, while the AI-driven approach, based on long short-term memory (LSTM) networks, aims to significantly improve prediction accuracy by addressing both systematic and dynamic errors. This innovative combination sets the stage for a new era in early warning systems.
The importance of post-processing in forecasts
Hydrological forecasts aim to predict events with maximum reliability, yet every forecast inherently comes with a level of uncertainty. These uncertainties can stem from limitations in the initial data, modelling processes, or changes in environmental conditions over time.
Post-processing plays a critical role in reducing these uncertainties by correcting errors in raw forecasts. It adjusts predictions to better align with observed data, improving both accuracy and usability. This is essential for decision-makers, as more reliable forecasts enable proactive responses to extreme weather events, minimizing risks to communities and infrastructure.
SEED-FD applies post-processing as a crucial step to improve the reliability of flood and drought predictions, creating a benchmark for innovative AI solutions. Building on this groundwork, the project focuses on enhancing precision and addressing complex forecasting challenges with dynamic, data-driven methods.
Establishing a benchmark with quantile mapping
In the first phase of the SEED-FD project, quantile mapping is being used as a baseline method for post-processing hydrological forecasts. This technique adjusts the statistical distribution of forecast data to better align with observed patterns, effectively correcting systematic biases such as consistent over- or underestimations.
Quantile mapping is simple, requires minimal data, and can be implemented rapidly across multiple locations. Its flexibility makes it an ideal starting point for benchmarking. However, it is limited in addressing dynamic errors, such as variations in timing or magnitude that change with forecast lead times.
Advancing forecast post-processing with artificial intelligence
SEED-FD aims to even surpass the capabilities of quantile mapping by developing an advanced AI-based post-processing method. This approach, leveraging long short-term memory (LSTM) networks, addresses both systematic and dynamic errors in hydrological forecasts.
The AI model currently being developed introduces a 2-stage LSTM framework:
- Pre-forecast LSTM: Processes historical data to refine initial conditions, improving the initial conditions of the forecast.
- Forecast LSTM: Focuses on correcting forecast errors dynamically, ensuring better accuracy over different lead times.
This innovative design, applied to GloFAS (Global Flood Awareness System) predictions, will address errors beyond static biases, such as those related to changing conditions over lead times.
Progress and application in real-world use cases
The development of SEED-FD’s AI model is currently focused on the Danube catchment, a region rich in observational data. Quantile mapping has already demonstrated its effectiveness in reducing systematic biases in this region. Next, the AI model will be tested to evaluate its ability to address more complex errors, setting a higher standard for post-processing.
Once validated, the AI method will be applied to additional regions, including the Bhima in India, and catchments in Africa (Juba-Shebelle, Niger), and South America (Paraná). These diverse contexts will test the scalability and adaptability of the method, ensuring its effectiveness across varied hydrological conditions.
Impact on communities and stakeholders
By integrating AI into hydrological post-processing, SEED-FD aims to deliver forecasts that are more aligned with reality, enabling better decision-making for disaster preparedness. Government agencies and local communities will benefit from:
- Accurate predictions of extreme events, reducing false alarms and missed warnings,
- Improved planning for flood and drought mitigation, and
- Tailored solutions for local thresholds and specific needs.
Through these advancements, SEED-FD is able to enhance how societies prepare for and mitigate the impacts of extreme hydrological events, ultimately saving lives and resources.
Conclusion and next steps
With innovative post-processing tools and AI at its core, SEED-FD is bridging critical gaps in global early warning systems. In the coming months, the project will finalize the LSTM-based model and expand its application to additional regions. By addressing the limitations of traditional methods like quantile mapping, SEED-FD is building a future where early warning systems are more precise, dynamic, and impactful than ever before.
Stay tuned for more progress updates and learn how SEED-FD’s innovations are reshaping disaster resilience on a global scale.