Estimating areal rainfall time series using input data reduction, model inversion, and data assimilation
|Title||Estimating areal rainfall time series using input data reduction, model inversion, and data assimilation|
|Year of Publication||2017|
|Academic Department||Monash University|
|Number of Pages||146|
|Keywords||Emergency management, Flood, hydrology, modelling., rainfall|
Floods are devastating natural hazards that can have severe socio-economic impacts and lead to a loss of life. Consequently, the key driver for this research is to provide techniques that will lead to an increase in flood forecast skill. As fluvial floods are a direct result of rainfall, detailed knowledge of uncertainties in rainfall observations provides a fundamental foundation for improving both rainfall and flood forecast skill. As the understanding of uncertainties present in rainfall time series is developed, so too will the confidence in rainfall forecasts, and short- and long-term streamflow forecasts. For a description of rainfall uncertainty to be complete it must take into account uncertainty when rainfall was not observed, thus allowing, model structural errors to be correctly identified, analyzed, and treated. Therefore the focus of this thesis is to develop a robust methodology to estimate rainfall time series and its uncertainty such that it is consistent with both streamflow and soil moisture observations. To effectively estimate rainfall time series, a method to reduce hydrological input data dimensionality was identified. The effective reduction of hydrological input data dimensionality allows modern parameter estimation algorithms to simultaneously estimate rainfall time series and model parameters. Due to their wide-spread use as model input data reduction techniques in other fields, the discrete cosine transform (DCT) and discrete wavelet transform (DWT) were used, for comparative purposes, to reduce the dimensionality of observed rainfall time series for the 438 catchments in the Model Parameter Estimation Experiment (MOPEX) data set. Once the time series were reduced to a small number of parameters, the rainfall time series were reconstructed for comparison with the observed hyetographs. The rainfall time signals are then reconstructed and compared to the observed hyetographs using standard simulation performance summary metrics and descriptive statistics. Analysis of the results demonstrate that, when compared to the DCT, the DWT is superior at preserving both short- and long-term rainfall patterns. Second, the DWT was used to reduce the dimensionality of the input rainfall time x series for the catchment of Warwick, Queensland, Australia. The DREAM(ZS) sampling algorithm, in conjunction with a likelihood function that considers both rainfall and streamflow, was then used to estimate the input rainfall time series. Model parameters and rainfall time series were simultaneously estimated. The inclusion of rainfall in the estimation process improved the root mean square error (RMSE) of streamflow simulations by a factor of up to 1.78. This was achieved while estimating an entire rainfall time series, inclusive of days when none was observed. Last, rainfall time series for the catchment of Warwick were estimated using three different rainfall-runoff models. Using the rainfall time series and model parameter estimates, remotely sensed soil moisture observations from the Soil Moisture Ocean Salinity (SMOS) and Advanced Microwave Scanning Radiometer - Earth observing system (AMSR-E) satellites were assimilated into each of the models using an ensemble Kalman filter (EnKF). Through analysis of the innovations from the observed and simulated soil moisture it was found that the combination of model choice and remotely sensed soil moisture product had a significant impact on the quality of rainfall estimated. When compared to streamflow simulations obtained via the sole estimation of model parameters, all models that jointly estimated rainfall time series and model parameters produced superior streamflow estimates. Rainfall estimates obtained using the Sacramento Soil Moisture Accounting (SAC-SMA) model were the most realistic. When the SMOS remotely sensed soil moisture product was assimilated into the SAC-SMA, innovations that indicated errors are of a Gaussian nature were obtained. Further, streamflow simulations obtained from the SAC-SMA had the best RMSE. The research presented in this thesis developed a methodology that can be used to estimate and evaluate rainfall estimates obtained using model input data reduction, model inversion, and data assimilation techniques. These rainfall estimates can be used to condition rainfall forecasts and consequently improve flood forecast skill.