Uncertainty estimation with deep learning for rainfall–runoff modeling

• DOI: 10.5194/hess-26-1673-2022
• Publication Date: 2022-03-31T09:08:53Z
• Authors: Daniel Klotz; Frederik Kratzert; Martin Gauch; Alden Keefe Sampson; Johannes Brandste...; Günter Klambauer; Sepp Hochreiter; Grey Nearing
• Keywords: Benchmarking; Dropout (neural networks); Hydrological modelling; Uncertainty Quantification

Abstract. Deep learning is becoming an increasingly important way to produce accurate hydrological predictions across a wide range of spatial and temporal scales. Uncertainty estimations are critical for actionable prediction, while standardized community benchmarks part model development research, similar tools benchmarking uncertainty estimation lacking. This contribution demonstrates that can be obtained with deep learning. We establish procedure present four baselines. Three baselines based on mixture density networks, one Monte Carlo dropout. The results indicate these approaches constitute strong baselines, especially the former ones. Additionally, we provide post hoc analysis put forward some qualitative understanding resulting models. extends notion performance shows learns nuanced behaviors account different situations.