Intercomparison of PRISM and Daymet Temperature Interpolation from 1980 to 2003

• Main Author: Scully, Rebecca A.
• Format: Others
• Published: DigitalCommons@USU 2010
• Subjects: comparison; Daymet; PRISM; Atmospheric Sciences
• Online Access: https://digitalcommons.usu.edu/etd/578; https://digitalcommons.usu.edu/cgi/viewcontent.cgi?article=1574&context=etd

As ecosystem modeling becomes increasingly integrated with Geographic Information Systems (GIS) there is a rise in demand for spatially and temporally continuous meteorological data. But in order to justify management decisions or to provide robust scientific insights, the accuracy of meteorological data used as model input must be thoroughly quantified. Current methods to create spatially continuous climate data from discrete weather station data include inverse distance weighting, geostatistical techniques such as kriging and splines, local regression models such as Parameter-Elevation Regression on Independent Slope Model (PRISM) and Daymet, and regional regression models. For the conterminous United States, PRISM and Daymet are perhaps the most commonly used interpolated datasets. Both use similar inputs but apply different interpolation methods. To date, no comprehensive comparison of their respective accuracies exists. Here I show that for a wide range of conditions, PRISM is the preferred interpolation. I reached this conclusion by comparing the accuracy of predictions of annual and monthly minimum (Tmin) and maximum (Tmax) by PRISM and Daymet for the conterminous United States from 1980-2003. My goals were: (1) to determine which interpolation was more robust at predicting temperature values; and (2) to assess whether the performance of each method varies, either temporally (annual or seasonal), spatially, or by elevation. To evaluate comparative performance, I analyzed PRISM and Daymet temperature predictions of ground station temperatures by calculating the logs odds ratio (LOR), mean absolute error (MAE), and bias. In all the comparative performance analyses, PRISM was the better model. The monthly results followed the same trend as the annual average results. I found a spatial performance difference across the entirety of the conterminous United States with the largest difference on the coasts and in the mountainous western regions. Stratifying data by elevation demonstrated that as elevation increases, uncertainty from both PRISM and Daymet increased. Unless the daily resolution provided by Daymet is required, PRISM appears to be a more robust predictor of continuous temperature data over the conterminous United States from 1980-2003.