Results
In terms of cross-validation errors, the three spatial interpolation methods performed similarly for both the total-day and consecutive-day count variables (Table: Interpolation method comparison for total-day variable and Table:Interpolation method comparison for consecutive-day variable). The GIDS approach, as suggested by the ME values as well as the actual versus the predicted means, tended to over-predict slightly more than the other two techniques. The RMSE results indicate that, for some years, the GIDS approach yielded a few more extreme errors, although GIDS had a lower RMSE than cokriging for the total-day variable in 2002 and 2003, as well as a lower MAE in 2001, 2002, and 2003. In general, error differences among the three were not substantial, with MAE consistently holding at approximately 16 percent of the total-day mean value and 25 percent of the consecutive-day mean value for all three techniques.
The GIDS-derived maps for the two count variables (Figures 1, 2) most obviously show a great deal of annual variability. For the consecutive-day variable, the Eastern United States generally tended to have higher values than the Western United States, with parts of the Appalachian Mountain region and States along the Gulf of Mexico typically exhibiting high values. However, the extent and spatial distribution of the highest-value area fluctuated substantially year-to-year. The total-day maps exhibited a similar spatial pattern, but more clearly highlighting some relatively high-value areas in the Southern and Central Rocky Mountains. Perhaps unsurprisingly, the patterns of the GIDS-derived maps were quite different than the patterns depicted in the PRISM-derived maps.
Encyclopedia ID: p3410
