Sensitivity of the Weather Research and Forecasting model 
(WRF) to downscaling extreme events over Northern Tunisia

Abstract. Rainfall is one of the most important variables for water and flood management. We investigate the capacity of the Weather Research and Forecasting model (WRF) to dynamically downscale the ECMWF Re-Analysis data for Northern Tunisia. This study aims to examine the sensitivity of WRF rainfall estimates to different Planetary Boundary Layer (PBL) and Cumulus Physics (Cu) schemes. The verification scheme consists of three statistical criteria (Root Mean Square Error (RMSE), Pearson correlation, and the ratio bias coefficient). Moreover, the FSS coefficient (fraction skill score) and the quality coefficient SAL (structure amplitude latitude) are calculated. The database is composed of four heavy events covering an average of 318 rainfall stations. We mean by heavy event, each event occurred a rainfall of more than 50 mm per observed day at least in one rainfall station. The sensitivity study showed that there is not a best common combination scheme (PBL and Cu) for all the events. The average of the best 10 combinations for each event is adopted to get the ensemble map. We conclude that some schemes are sensitive and others less sensitive. The best three performing schemes for PBL and Cu parametrizations are selected for future rainfall estimation by WRF over Northern Tunisia.


tested at 35 km horizontal resolution to quantify the seasonal biases of rainfall. It was found that rain rates were 48 predominantly sensitive to Cu schemes and much less to PBL and MP schemes. They found that WRF simulates 49 accurately seasonal gradients of rainfall also the seasonal large-scale rainfall patterns. However, they noticed 50 strong seasonal biases fluctuation from an experiment to another. We conclude from this study of (Crétat et al., 51 2011) that without testing numerous physical parameterizations one couldn't find satisfactory rainfall estimations.

52
Another sensitivity study was achieved by Evans et al. (2011), over the south of Australia, to evaluate the ability 53 of a 36 member multi-physics WRF ensembles to reproduce four East Coast Low events. Two PBL schemes, two 54 Cu schemes, three microphysics (Mp) schemes, and three radiation (Ra) scheme combinations of shortwave and 55 longwave schemes respectively were used to create these 36 members. A weak sensitivity appears for weak 56 weather systems in comparison with extreme events. In agreement with previous WRF parameterizations studies 57 (Jankov et al., 2005; Flaounas et al, 2011), not a single preferred member is the best for all cases and all metrics.

58
To study WRF sensitivity over Tunisia, this paper contains four other sections organized as follows: Section 2 59 describes the in situ data and used WRF parametrizations, Section 3 provides the sensitivity study methodology, 60 Section 4 represents the results, and the last section summarizes the conclusions and perspectives. The spatial interpolation of the in situ precipitation data was achieved using an inverse distance weighted (moving

142
For the rest of the work we will use the four selected days out of the eleven undetected events by MSGMPE.

143
Furthermore, based on quantile quantile comparison of the three different parameters (PBL, Cu, Mp) schemes, 144 we will choose which parameters will be used for the sensitivity study.

153
We assume the three most commonly adopted parameters (PBL, Cumulus (Cu) and microphysics (Mp)) to analyze 154 the sensitivity of WRF over the study area. Figure  For the PBL schemes simulation, the Cu scheme was fixed to 2 and Mp scheme to 6 ( Fig.5a).

159
We notice that for the PBL parameter (Fig.5a), the rainfall estimation differs from one scheme to another. It is 160 concluded that there is some WRF sensitivity for this parameter over the study area. To illustrate the sensitivity of 161 the Cu schemes the PBL scheme was fixed to 9 and Mp scheme was fixed to 6 (

169
The choice of these events is based on the incapability of MSGMPE to detect them. Also, we chose them because 170 of the difference in the type of rain (scattered or very localized in space, in topographic area) and for the location 171 difference of the extreme values in the ground.

172
A threshold of 0.1 mm is used in SAL and FSS verification to distinguish between rainy and no rainy pixels. In 173 case of undetected events, they will be deleted in the SAL diagram. The number of these non-represented cases in 174 SAL will indicate the poor forecasts. This will appear foremost for the high thresholds (30 and 50 mm/day).
The variable d is the largest distance among two points in the specified domain.

207
While X(Rmod) and X(Robs) is the mass center of the observed and modeled precipitation fields respectively. 224 The values of S are within [-2, +2]. When S is more than 0 that means the predicted rainfall objects are too outsized 225 and/or too smooth (Fig.6), while when it is less than 0 that means that the predicted objects of rainfall are too small The FSS coefficient in Fig.8 (b) helped us to identify the best 10 combinations (  After achieving the ranging of the schemes based on the sum metrics methodology, we select the best 20 schemes 287 to evaluate them using the FSS and the SAL verification method (Fig.9).
288 FSS helps us to select the best 10 combinations (Fig.9b) that are mentioned in Table 3 The various FSS thresholds clarify the skills of combinations (Fig.10b). After calculating the sum of metrics, we 296 selected the 10 best combinations (Table 4).

297
[  To find the best 10 combinations we represented the 20 best combinations selected previously by the metrics sum. 306 Fig.9b helped us to identify only 9 best combinations. We select the 10 th combination based on the metrics sum 307 (Cu5Pb8) which was not so representative of FSS (Table 5).