High-Resolution Precipitation Mapping in a Mountainous Watershed: Ground Truth for Evaluating Uncertainty in a National Precipitation Dataset

A 69-station, densely-spaced rain gauge network was maintained over the period 1951-1958 in the Coweeta Hydrologic Laboratory, located in the southern Appalachians in western North Carolina, USA. This unique dataset was used to develop the first digital seasonal and annual precipitation maps for the Coweeta basin, using elevation regression functions and residual interpolation. It was found that a 10-m elevation grid filtered to an approximately 7-km effective wavelength explained the most variance in precipitation (R2 = 0.82-0.95). A “dump zone” of locally high precipitation a short distance downwind from the mountain crest marking the southern border of the basin was the main feature that was not explained well by the precipitation-elevation relationship.

These data and maps provided a rare “ground-truth” for estimating uncertainty in the national-scale Parameter-elevation Relationships on Independent Slopes Model (PRISM) precipitation grids for this location and time period. Differences between PRISM and ground-truth were compared to uncertainty estimates produced by the PRISM model and cross-validation errors. Potential sources of uncertainty in the national PRISM grids were evaluated, including the effects of coarse grid resolution, limited station data, and imprecise station locations.

The PRISM national grids matched closely (within five percent) with the Coweeta dataset. The PRISM regression prediction interval, which includes the influence of stations in an area of tens of km around a given location, overestimated the local error at Coweeta (12-20 percent). Offsetting biases and generally low error rates made it difficult to isolate major sources of uncertainty in the PRISM grids. However, station density and selection, and mis-location of stations were identified as likely sources of error. The methods used in this study can be repeated in other areas where high-density data exist to gain a more comprehensive picture of the uncertainties in national-level datasets, and can be used in network optimization exercises.