A parsimonious, data-based UH rainfall-runoff model for flood forecasting

Abstract:

Converting rainfall into runoff with lumped hydrological models requires less information, as compared to distributed models. Also, these models are relatively simple, typically do not suffer from over-parameterization, and only require a single value of precipitation for the whole watershed over each time step. Short-term flood forecasting with concurrent information of recent total rainfall over the watershed allows one to avoid the uncertainty in weather forecasting. Point depth measurements from rain gauges are the most reliable, however, they are scarce. Also, the hydrological response of a watershed will vary according to its antecedent conditions. We explore and try to find solutions to these issues in a study on three sub-basins of the Obion-Forked Deer River System, in Tennessee, that have USGS gaging stations at Owl-City (1860 km²), Martin (964 km²), and Trenton (191 km²).

We use Stage IV QPE radar hourly precipitation data, which have been manually quality-controlled at NWS-River Forecast Centers using gage data, to analyze numerous events spanning 2004-2014. These data help us understand the spatial distribution of rainfall events on a 4 km grid and identify their effective coverage of the basin. To derive unit hydrographs (UHs), we utilize gaged discharge data in conjunction with average radar precipitation, for those events with a weighted-average spatial coverage in excess of 80% of the watershed area. We use ESA CCI Soil Moisture Combined Active-Passive dataset V4.4 with a grid size 25 km to find daily average soil moisture. Our analyses show that it can be used to predict infiltration capacity of a catchment (assuming constant ф-index). However, in the forecast mode, we use river baseflows at the beginning of a rainfall event as an index of pre-existing watershed condition, and then predict runoff coefficients and effective precipitation based on linear regression equations. This information is finally used to estimate shape factors of gamma probabilistic distribution functions, that are used as UHs for flood forecasts. This method performs better than using a fixed, averaged UH. Instead of representing the complex physical processes of runoff generation, we focus on the variability of the input rainfall and antecedent conditions and attempt to make the model as parsimonious as possible.