@PhdThesis{duepublico_mods_00074986,
  author = 	{Alvarado Montero M.Sc., Rodolfo},
  title = 	{Variational data assimilation for operational flood forecasting},
  year = 	{2021},
  month = 	{Dec},
  day = 	{15},
  keywords = 	{Data assimilation; Variational methods; Hydrological forecasting},
  abstract = 	{Hydrological forecasts lie at the core of early warning systems. Their accuracy and reliability highly depend on the use of models that can represent the physical characteristics of a hydrological system. The uncertainty in the forecast can therefore be attributed not only to the inputs to the model, but also to the model itself as well as to the initial conditions of the model at the time the forecast is produced. The latter often relies on data assimilation techniques in order to provide accurate estimates of model state variables. Data assimilation is often described as a set of procedures that merge observed and simulated data to improve the true estimate of a given value. This definition acknowledges that both the observed and simulated data have an intrinsic error, which can be improved by combining the two quantities together. Most data assimilation techniques can be classified as sequential or variational types. While sequential assimilation has been thoroughly investigated in hydrological applications, its counterpart, variational assimilation, has seldom been subject of research in this field. The main reason for this is arguably the need to compute an adjoint model, which can estimate first-order sensitivities of the hydrological model. This is a major drawback for most practitioners. On the contrary, variational assimilation has been enthusiastically adopted in meteorological sciences, where a vast number of observations are required to estimate the atmospheric current state. This research focuses on the assessment of variational data assimilation in hydrological forecasting systems. The research includes the development of a general formulation of variational data assimilation in order to assimilate multiple sources of observations. A wide range of properties of the variational approach are analyzed, among them the length of the assimilation window, the type of noise variables incorporated in the objective function, as well as the spatial resolution of the noise. Furthermore, the research proposes an extension of the variational assimilation to provide a probabilistic initialization of model states. This is done by integrating multiple model parametric and model structures sets. The former rely on parametric reductions techniques in order to maximize the parametric distances for a given number of preselected sets. Finally, the impact of the variational approach in forecast performance is compared to the equivalent results obtained with Ensemble Kalman Filter, one of the most commonly applied sequential data assimilation techniques in hydrology. The results show that variational assimilation provides a flexible formulation to simultaneously assimilate multiple sources of observations. It demonstrates the trade-off between the number of noise variables, their corresponding bounds, their resolution, the length of the assimilation window, as well as the weighting factors for each of the terms of the objective function. The approach is particularly relevant to adjust delaying processes in the system, such as snow accumulation, snowmelt conditions, and slow reacting components in the hydrological system, e.g. low zone budget zones.},
  doi = 	{10.17185/duepublico/74986},
  url = 	{https://duepublico2.uni-due.de/receive/duepublico_mods_00074986},
  url = 	{https://doi.org/10.17185/duepublico/74986},
  file = 	{:https://duepublico2.uni-due.de/servlets/MCRFileNodeServlet/duepublico_derivate_00074721/Diss_AlvaradoMontero.pdf:PDF},
  language = 	{en}
}