Forecasting streamflow in real-time has applications in water management, and emergency actions. Radar data and a fully distributed physics-based hydrologic model is used to predict how high river levels with rise and when the flow will arrive downstream in reservoirs. The study presented is part of a larger project to test distributed hydrologic modeling from an end-user perspective. The goal of this project is to identify the utility and skill of distributed modeling to provide better forecasts at a timescale of less than six hours.  

WATERSHED DEMONSTRATION  

This demonstration project includes the Sandy Creek watershed in central Texas, Clear Creek located in the coastal plains of East Texas, and a basin to be selected in Puerto Rico. Participants in this project include the National Weather Service, Southern Region, local organizations, and a private sector partner, Vieux & Associates, Inc. The pilot study for the 896 km² (346 mi²) Sandy Creek watershed is briefly described below. The model resolution selected for this watershed is 150-m. This resolution affords a very detailed representation of the watershed with nearly 40,000 grid cells. The outlet of the basin is located at Kingsland, TX, and is upstream from Lake LBJ. The following documents the data used, and uncalibrated and calibrated results for a single storm event referred to herein as the July event (07/04/02). More events are expected to be tested and with other sources of rainfall data input, e.g., the Multi-sensor Precipitation Estimate (MPE) produced by NWS. Further calibration of the model will likely be warranted as more events are tested.  

MODEL DEVELOPMENT STATUS  

The following datasets are used to develop the Sandy Creek model.

 

• DEM:  30-m DEM provided by LCRA resampled to 150-m for final Vflo™ domain resolution.  Slope, flow direction, overland versus channels cells, and background image derived from DEM.
• Landuse: National Land Cover Database land cover is resampled from 30-m to 150-m and processed for overland hydraulic roughness values.
• Soils: STATSGO 1-km soils data resampled to 150-m for generation of Green-Ampt parameters obtained by regression relationships that relate soil properties to wetting front suction, hydraulic conductivity, and porosity.
• Channels: Channel widths were estimated from 20 randomly selected sample points using Digital Orthophoto Quarter Quadrangles.  The estimated channel widths for the 20 sample points were then used to develop regression equations according to regional groups within the basin. DEM channel slopes often contain artifacts resulting from sampling, which are not representative of the channel and terrain. Channel bed slopes were obtained at 20 randomly sampled locations by locating upstream and downstream cells and computing a local slope. A regression relationship between slope and flow accumulation was used to assign channel slope throughout the basin except where rating curves were established.
• Regionalized Channel Hydraulics:   Randomly selected points are grouped that have similar correlations between channel width and flow accumulation. Using several regional regression relationships, trapezoidal channel shapes are assigned to stream channels except for locations with rating curves.
• Initial Degree of Saturation: From USGS recent-daily streamflow data for Kingsland prior to July 2002, the Sandy Creek basin was relatively dry, with discharge values near the lowest during a two-year period. A basin-wide initial saturation value of 30% was chosen to represent the dry conditions. Operational implementation will rely on soil moisture modeling using evapotranspiration and distributed infiltration input.
• Radar Rainfall: Level II data calibrated to LCRA rain gauge data is used as model input. The rainfall maps sampled to 15 minute intervals at 1x1 km resolution were produced by RainVieux. Level II reflectivity and Level III DPA were adjusted using data from 22 LCRA rain gauges. After rain gauge adjustment the radar rainfall agreed with the gauges within 8.2% to 26.6%.

Figure 1 shows a map of Sandy Creek. Topography is shown as a shaded relief map with an overlay of landuse/cover. Figure 2 shows the hydrograph results from a priori parameter values derived from the geospatial data. After calibration, the simulated and observed hydrographs agree closely as seen in Figure 3.  

Figure 1 shows a map of Sandy Creek. The watershed outlet at Kingsland is just upstream of Lake LBJ, and is indicated by the large filled triangle in the upper right of the image.

Figure 2 Uncalibrated simulated (black) and observed (red) discharge at Kingsland.

Figure 3 Calibrated results showing simulated (black) and observed (red) discharge at Kingsland.  

A fully distributed physics-based hydrologic model has been used to provide short-term streamflow prediction using calibrated radar rainfall inputs. Close agreement in the uncalibrated results gives confidence in the model and its application to this watershed. Calibration by adjusting parameter values resulted in improved agreement between simulated and observed discharge at Kingsland. As more events are simulated, model calibration will be refined. Testing is underway with other sources of radar rainfall input derived from NWS products including MPE.