Regionalization of model parameters by incorporating rainfall uncertainty: "A large sample study on 326 rainfall dominated catchments in the USA"
Yimer, E.A.; Nossent, J. (2020). Regionalization of model parameters by incorporating rainfall uncertainty: "A large sample study on 326 rainfall dominated catchments in the USA". MSc Thesis. Vrije Universiteit Brussel/KU Leuven: Brussel. X, 55 pp.
|
| Beschikbaar in | Auteurs |
|
Documenttype: Doctoraat/Thesis/Eindwerk
|
| Author keywords |
Ungauged basins; Linear multiple regression; Rainfall uncertainty |
| Auteurs | | Top |
- Yimer, E.A.
- Nossent, J., revisor, meer
|
|
|
| Abstract |
Prediction in ungauged catchments is a topic that was given attention for several decades and still, researchers are devoted to improve the different methods of regionalization. The demerits of each regionalization technique make the process of regionalization difficult and unpredictable. Furthermore, uncertainty related to input, parameter, model structure, and streamflow data are additional challenges in the regionalization scheme. This thesis focuses on improving the regionalization method of multiple regression by considering input uncertainty. The availability of a large sample database called CAMELS, in the USA allows us to perform multiple trials and have concrete evidence on the effect of incorporating input uncertainty on regionalization. This thesis is a large sample study where we have analyzed 326 rainfall dominated catchments. Hydrological modeling is conducted using the HYMOD model. To get optimized parameter sets, we have used the DREAM(ZS) algorithm.
We applied a linear multiple regression model between independent variables (catchment attributes) and dependent variables (model parameters) with and without incorporating uncertainty. This allowed us to see the difference before and after accounting for input uncertainty. In a first step, we performed the regression with three different approaches to obtain a well-performing regression model. Approach 1 incorporates catchments which managed to score a Nash Sutcliffe Efficiency (NSE) greater or equal to 0.5 for regionalization (160 catchments). Approach 2 uses catchments that showed an NSE value greater or equal to 0.65 (104 catchments) and finally approach 3 contains catchments with NSE above 0.7 (67 catchments). We used 75 independent common catchments for all the three approaches for cross-validation.
It was observed that approach 2 and 3 performed almost the same, while approach 1 performs better than the other approaches. This is important, because we noted that adding catchments which are classified as “Satisfactory (0.65In the next phase, we incorporated input uncertainty in the model parameter optimization strategy. We selected independent rainfall events and then, we applied rainfall multipliers (additional parameters) to correct those rainfall events. Due to this, the posterior distribution of the optimized model parameters changes. The median rainfall multiplier values were around 1 for most catchments. Then using the newly generated parameter sets, the linear multiple regression has been conducted and the result during cross-validation shows a clear improvement of the NSE values. Among the 75 validation catchments, 40 and 37 of them managed to show improvements during the calibration and validation periods, respectively. This proves the vital use of accounting input uncertainty in linear multiple regression. |
|
