Statistical downscaling of daily precipitation and the stationarity assumption

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Monday, 3 February 2014
Hall C3 (The Georgia World Congress Center )
Carlos Felipe Gaitan, University of Oklahoma - NOAA/GFDL, Princeton, NJ; and K. W. Dixon, R. A. McPherson, B. Moore III, V. Balaji, and A. Radhakrishnan

Handout (10.9 MB)

To test the statistical downscaling time-invariance assumption we used daily precipitation outputs from a high resolution (~25km grid spacing) global atmospheric model as predictands, and a coarsened version of the same high resolution outputs - interpolated to a ~100km grid - as predictors of a hybrid downscaling model based on classification and regression trees (CART) and support vector regression (SVR) with evolutionary strategies. This experimental setup, known as “Big-Brother” allows us to use the high resolution (historical and future) model outputs as pseudo-observations so we can validate the downscaled values against them.

The study region examined here focuses on 16 points across North America. We evaluated the downscaled results in terms of historical and future mean absolute error skill score (MAE SS) to assess if the skills were time-invariant. Our results show that for 9 out 16 points our hybrid CART-SVR downscaling model had positive historical and future MAE SS. We also found that the CART model under predicted the total number of rainy days. Future implementations will test other classification methods (e.g. support vector classification, drizzle threshold) and will expand the predictor set aiming to improve the overall MAE SS.