On noice in data assimilation schemes for improved flood forecasting using distributed hydrological models

We investigate the effects of noise specification on the quality of hydrological forecasts via an advanced data assimilation (DA) procedure using a distributed hydrological model driven by numerical weather predictions. The sequential DA procedure is based on (1) a multivariate rainfall ensemble generator, which provides spatial and temporal correlation error structures of input forcing, and (2) lagged particle filtering to update past and current state variables simultaneously in a lag-time window to consider the response times of internal hydrologic processes. The procedure is evaluated for streamflow forecasting of three flood events in two fast-responding catchments in Japan (Maruyama and Katsura). The rainfall ensembles are derived from ground-based rain gauge observations for the analysis step and numerical weather predictions for the forecast step. The ensemble simulation performs multi-site updating using information from the streamflow gauging network and considers the artificial effects of reservoir release. Sensitivity analysis is performed to assess the impacts of noise specification in DA, comparing a different setup of random state noise and input forcing with/without multivariate conditional simulation (MCS) of rainfall ensembles. The results show that lagged particle filtering (LPF) forced with MCS provides good performance with small and consistent random state noise, whereas LPF forced with Thiessen rainfall interpolation requires larger random state noise to yield performance comparable to that of LPF + MCS for short lead times.