Application
- Reliable flood forecasting is important for saving lives and minimizing damages and risks caused by floods.
- Process-based hydrological models forecast floods through explicit representations of the hydrological process as understood by the hydrological community. Recent studies have found that deep learning can improve flood forecasting in prediction accuracy, scalability, and regional generalization.
- Interpretability of deep learning flood forecasting models will lead to a better understanding of hydrological processes leading to floods and thus improve process-based flood models.
Our Innovation
- Our technology combines process-based and deep machine-learning modeling in an innovative approach.
- We develop interpretability tools aiming at gaining an understanding of flood-controlling processes from well-performing deep-learning models.
- We aim to improve both process understanding and forecasting capabilities for floods.
Advantages
- Current practice in flood forecasting models utilizes mostly existing process-based models and, recently, machine-learning models, but there is no combination of the two. Our research uniquely integrates the two modeling approaches and improves theory and practice.
- The flexibility of machine learning models makes their application much easier than process-based models. The latter typically requires substantial expert efforts in constructing and calibrating hydrological models. Our approach will allow easier implementation of reliable flood models in many locations worldwide.
Opportunity
We propose a new process learning framework that is applied to flood forecasting. This approach can be extended to other disciplines.
We seek collaborations with stakeholders interested in flood forecasting with an integrative view that combines predictability and process understanding.
