Comparison of Deterministic and Probabilistic Variational Data Assimilation Methods Using Snow and Streamflow Data Coupled in HBV Model for Upper Euphrates Basin

Abstract

The operation of upstream reservoirs in mountainous regions fed by snowmelt is highly challenging. This is partly due to scarce information given harsh topographic conditions and a lack of monitoring stations. In this sense, snow observations from remote sensing provide additional and relevant information about the current conditions of the basin. This information can be used to improve the model states of a forecast using data assimilation techniques, therefore enhancing the operation of reservoirs. Typical data assimilation techniques can effectively reduce the uncertainty of forecast initialization by merging simulations and observations. However, they do not take into account model, structural, or parametric uncertainty. The uncertainty intrinsic to the model simulations introduces complexity to the forecast and restricts the daily work of operators. The novel Multi-Parametric Variational Data Assimilation (MP-VarDA) uses different parameter sets to create a pool of models that quantify the uncertainty arising from model parametrization. This study focuses on the sensitivity of the parametric reduction techniques of MP-VarDA coupled in the HBV hydrological model to create model pools and the impact of the number of parameter sets on the performance of streamflow and Snow Cover Area (SCA) forecasts. The model pool is created using Monte Carlo simulation, combined with an Aggregated Distance (AD) Method, to create different model pool instances. The tests are conducted in the Karasu Basin, located at the uppermost part of the Euphrates River in Türkiye, where snowmelt is a significant portion of the yearly runoff. The analyses were conducted for different thresholds based on the observation exceedance probabilities. According to the results in comparison with deterministic VarDA, probabilistic MP-VarDA improves the m-CRPS gains of the streamflow forecasts from 57% to 67% and BSS forecast skill gains from 52% to 68% when streamflow and SCA are assimilated. This improvement rapidly increases for the first additional model parameter sets but reaches a maximum benefit after 5 parameter sets in the model pool. The improvement is notable for both methods in SCA forecasts, but the best m-CRPS gain is obtained for VarDA (31%), while the best forecast skill is detected in MP-VarDA (12%).