Deep learning-based uncertainty characterization of precipitation forcing in a hydrologic data assimilation framework
Abstract
Precipitation is arguably the most erroneous variable in hydrologic modeling and prediction. While precipitation data can be obtained from reanalysis products, for hydrologic applications that often require fine spatiotemporal resolution, reanalyses show large errors in capturing the timing and location of precipitation events. This is especially critical for the task of hydrologic modeling in which precipitation plays a crucial role. In this study, we will characterize the uncertainty of multiple precipitation products, including Precipitation Regression on Independent Slopes Method (PRISM), Integrated Multi-satellitE Retrievals for GPM (IMERG), and Phase 2 of the North American Land Data Assimilation System (NLDAS-2), and then fuse all datasets using a state-of-the-art deep Bayesian belief network and Monte-Carlo sampling to fully capture the associated uncertainty in precipitation. Then, the fused precipitation product along with its associated uncertainty will be used in a hydrologic data assimilation framework to improve streamflow prediction. Instead of simply perturbing the precipitation variable, which has a significant effect on the characterization of the uncertainty of the model outputs, here we will use the actual probability distribution of this variable obtained from a DL approach by fusing various precipitation products. The results will be then compared with various perturbation methods to see how much the characterization of precipitation uncertainty helped improve streamflow prediction.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H25K1239G