Www.nr.no Evaluation of a dynamic downscaling of precipitation over the Norwegian mainland Orskaug E. a, Scheel I. b, Frigessi A. c,a, Guttorp P. d,a,

www.nr.no

Evaluation of a dynamic downscaling of precipitation over the Norwegian mainland

Orskaug E.a, Scheel I.b, Frigessi A.c,a, Guttorp P.d,a, Haugen J. E.e, Tveito O. E.e, Haug O.a

a Norwegian Computing Center, Oslo, Norway

b Department of Mathematics, University of Oslo, Oslo, Norway

c Department of Biostatistics, University of Oslo, Oslo, Norway

d University of Washington, Seattle, USA

e The Norwegian Meteorological Institute, Oslo, Norway

Motivation

► Climate research produces an increasing number of data sets combining different GCMs, CO2 emission scenarios and downscaling techniques.

► For impact studies, but also as an issue of separate interest, the quality of these data need to be verified.

Goal

► We want to compare downscaled ERA-40 reanalysis data (RCM) against observations of Norwegian precipitation.▪ How good are the RCM data?▪ Where (in the distribution) does the RCM differ from

the observations?▪ Where (geographically) does the RCM perform

best/worst?

Why is this work important?

► It assesses the quality of a dynamic downscaled data and highlights which areas these data capture reality and where there are deviations from the truth.

► Another aim is to show how standard methods of statistical testing may be used to assess dynamic downscaling.

Data

Model data► RCM model, dynamically

downscaled HIRHAM model, forced by ERA-40 reanalysis data from the ENSEMBLES project.

► Spatial resolution of 25 x 25 km2.

► Reliant on the downscaling, still supposed to possess properties similar to real weather locally over longer time periods.

Observations

► Interpolations (1 x 1 km2) from a triangulation of the official measurement stations operated by The Norwegian Meteorological Institute.

► Aggregated to 25 x 25 km2 scale by collecting 1 x 1 km2 grid cells with centre points within the RCM cell, the mean is representing the precipitation within that grid cell.

Data – The RCM

► The RCM from the ENSEMBLES project

Data – properties for both data sets

► Climate variable: precipitation

► Time period: 1961 – 2000

► Time scale: Daily, seasonal

► Resolution: 25 x 25 km2

► Number of grid cells: 777 grid cells covering Norway

Methods for comparison

► Evaluate the distributions1. Global measure:

Kolmogorov Smirnov test

2. Local measures:

Comments

► Drizzle effect avoided: conditioned on wet days; i.e. days with precipitation below a small, positive threshold (0.5 mm/day) are discarded.

► Day-to-day correlation in the RCM is partly lost due to downscaling, hence the distributions have to be compared instead of comparing day by day.

► Separate tests for each grid cell and each season.

Kolmogorov-Smirnov test

► K-S two sample test is used to check whether the empirical distributions from the RCM and the observations are equal.

► To avoid the problem of tied data, a small, random normally distributed number, N(0, σ2), is added to each data point.

σ = 1e-7

Kolmogorov-Smirnov test – Results

► The null hypothesis of equality of the distributions are rejected for almost all grid cells for all the four seasons.

► Global picture: the RCM does not have the same distribution as the observations.

► Next: want to find out where the distributions differ; local measures.

Methods for comparison

► Evaluate the distributions1. Global measure:

Kolmogorov Smirnov test

2. Local measures:

Test equality of quantiles

Construction of the 2 x 2 contigency table

0.05-quantile – Results

► Hardly any rejections of null hypothesis of equality.

► For low quantiles: the RCM reproduces the observations well both season- and nationwide.

0.95-quantile – Results

► Mainly rejections of the null hypothesis of equality.

► Overall picture: the RCM underestimates high precipitation.

Generalized Pareto Distribution (GPD)

►  

GPD – Results

► One-year return levels from GPD are more similar than expressed through the Kolmogorov-Smirnov test.

► But still: tendency that the RCM underestimates high precipitation.

Wet day frequency

► Wet day frequency = Proportion of wet days (among all days in the data)

► A wet day is defined to be above 0.5 mm/day for both data sets.

► The equality of the wet day frequency is tested by permutation testing.

Wet day frequency – Results

► Mainly rejections of the null hypothesis of equality.

► Total picture: Wet day frequency of the RCM is greater than for the observations.

Summary

► Small amounts of rainfall: the RCM shows good agreement with the observations.

► When rainfall amounts is beyond the first quartile, the agreement disappear.

► The RCM has too many and too small rain events for all seasons.

► This work is accepted for publication in Tellus A.

An improvement/correction of the RCM is needed.

What to do next?

► We want to add a statistical correction method to the output of the RCM, especially improve the right tail.

► Simple linear regression was tried out, but did not improve the results.

► We are currently working on a more complex transformation with spatial corrections.