Exploring the benefits of satellite remote sensing for flood prediction across scales

Space-borne remote sensing datasets have the potential to allow us to progress towards global scale flood prediction systems. However, these datasets are limited in terms of space-time resolution and accuracy, and the best use of such data requires understanding how uncertainties propagate through hydrological models. An unbiased investigation of different datasets for hydrological modeling requires a parsimonious calibration-free model, since calibration masks uncertainties in the data and model structure. This study, which addresses these issues, consists of two parts: 1) the development and validation of a multi-scale distributed hydrological model whose parameters can be directly linked to physical properties of the watershed, thereby avoiding the need of calibration, and 2) application of the model to demonstrate how data uncertainties propagate through the model and affect flood simulation across scales.

I based the model development on an interactive approach for model building. I systematically added processes and evaluated their effects on flood prediction across multiple scales. To avoid the need for parameter calibration, the level of complexity in representing physical processes was limited by data availability. I applied the model to simulate flows for the Cedar River, Iowa River and Turkey River basins, located in Iowa. I chose this region because it is rich in high quality hydrological information that can be used to validate the model. Moreover, the area is frequently flooded and was the center of an extreme flood event during the summer of 2008. I demonstrated the model's skills by simulating medium to high-flow conditions; however the model's performance is relatively poor for dry (low flow) conditions. Poor model performance during low flows is attributed to highly nonlinear dynamics of soil and evapotranspiration not incorporated in the model. I applied the hydrological model to investigate the predictability skills of satellite-based datasets and to investigate the model's sensibility to certain hydro-meteorological variables such as initial soil moisture and bias in evapotranspiration. River network structure and rainfall are the main components shaping floods, and both variables are monitored from space. I evaluated different DEM sources and resolution DEMs as well as the effect of pruning small order channels to systematically decreasing drainage density. Results showed that pruning the network has a greater effect on simulated peak flow than the DEM resolution or source, which reveals the importance of correctly representing the river network. Errors on flood prediction depend on basin scale and rainfall intensity and decrease as the basin scale and rainfall intensity increases. In the case of precipitation, I showed that simulated peak flow uncertainties caused by random errors, correlated or not in space, and by coarse space-time data resolution are scale-dependent and that errors in hydrographs decrease as basin scale increases. This feature is significant because it reveals that there is a scale for which less accurate information can still be used to predict floods. However, the analyses of the real datasets reveal the existence of other types of error, such as major overall bias in total volumes and the failure to detect significant rainfall events that are critical for flood prediction.