Development of a large-sample watershed-scale hydrometeorological dataset for the contiguous USA : dataset characteristics and assessment of regional variability in hydrologic model performance

Introduction Conclusions References


23
With the increasing availability of gridded meteorological datasets, streamflow records 24 and computing resources, large sample hydrology studies have become more common in the last   The next section describes the development of the basin dataset from basin selection 68 through forcing data generation. It then briefly describes the modeling system and calibration 69 routine. Next, example results using the basin dataset and modeling platform are presented.

70
Finally, concluding thoughts and next steps are discussed.      In terms of annual mean CDFs, Daymet estimated basin mean temperatures range from -2 ºC to 152 23 ºC with precipitation amounts of 0.7 to 9.4 mm day -1 (Fig. 2). Annual observed mean runoff 153 ranges from 0.01 to 9.3 mm day -1 with PET estimates ranging from 1.9 to 4.8 mm day -1 . 154 Interestingly, this implies that Daymet precipitation itself is not enough to balance the observed  Precipitation/PET). Immediately it can be seen that some basins lie above the water limit line 178 (Y=1) indicating more runoff than precipitation and many basins are near it (Y > 0.9). In these 179 cases the model calibration process would struggle to produce an unbiased calibration, or never 180 in basins above the water limit, because the basic water balance requires nearly zero 181 evapotranspiration (ET) or is not satisfied. This requires a modification to incoming 182 precipitation, which is discussed in the next section. Not coincidentally, the basins near and

229
We employed a split-sample calibration approach following Klemes (1986), assigning the  Table 1). The SCE algorithm was run using 10 different random seed starts for the 252 initial parameter sets for each basin, in part to evaluate the robustness of the optimum in each 253 case, and the optimized parameter set with the minimum RMSE from the ten different 254 optimization runs was chosen for evaluation.

255
For Snow-17, six parameters were chosen for optimization (Table 1) (Fig 5a, section 3c). When examining MNSE (Fig 5b), basins with high non-seasonal   Fig 7b). However, nearly all basins have too little modeled variance (values less than one) 370 for both the calibration and validation phases (Fig. 7c). The total volume biases are generally variance and near zero total flow bias (Fig 7). This manifests itself in the simulated hydrograph 403 as under predicted high flows, generally over predicted low flows and a more positive slope to 404 the middle portion of the FDC (Fig. 9). It is worth repeating that the goal of this initial 405 application is to provide to community with a benchmark of model performance using well   (Fig 11a-b). The arid basins are generally dry with sporadic high precipitation (and 423 flow) events, while the Appalachian basins are wetter (Fig. 1b) with extreme precipitation events 424 interspersed throughout the record. Basins with significant snowpack tend to have lower error 425 contributions from the largest error days (Fig. 11a-b). The E50 metric highlights mean peak 426 snow water equivalent (SWE) and frequent precipitation basins as well. These regions contain 427 and order of magnitude more days than the high plains and desert SW, giving insight into how 428 representative of the entire streamflow timeseries the optimal model parameter set really is.

429
Additionally, ranking the basins using their fractional error characteristics provides a 430 similar insight. As the aridity index increases, the fractional error contribution increases for 431 basins with little to no mean peak SWE. For basins with significant SWE, the fractional error 432 contribution decreases with increasing aridity (Fig. 12). Alternatively, for a given aridity index