a. AIRS cloud-detection schemes

2) CO₂-slicing scheme

The CO₂-slicing method (Chahine 1974; Menzel et al. 1983), based on radiative transfer principles, is widely used to retrieve CTP and N_e. Here N_e conceptually represents the product of the geometrical cloud fraction and the gray body emissivity of the cloud. This method uses a simplistic cloud model: the cloud is considered as a single layer of opaque or semitransparent thin cloud with a homogeneous emissivity. The algorithm uses observed radiances of a set of 124 AIRS channels selected in the CO₂ absorption band (between 649.612 and 843.913 cm⁻¹), which is very sensitive to clouds. For each AIRS pixel, and each channel in the set, the following function is calculated:

View Expanded

where p is the pressure level number, k is the channel in the CO₂ band, k_ref represents the reference window channel (979.1279 cm⁻¹), is the measured radiance for channel k, is the simulated clear radiance for channel k, and represents the simulated blackbody radiance for channel k at the cloud level p. The cloud-top pressure level assigned to each channel k is the pressure level p_c,k, which minimizes the function F_k,p. Before ascertaining the CTP of a hypothetic cloud, a filter that distinguishes channels with δTBs (difference between observed brightness temperature and simulated brightness temperature) lower than the radiometric noise is applied to the algorithm. If all channels are filtered out, the pixel is flagged clear. If the pixel is cloud contaminated, the CTP is then calculated by the following expression:

View Expanded

where w_k = δF_k,p/δ lnp is the derivative of the cloud pressure function.

The effective cloud fraction is obtained for the reference window channel by the following expression:

View Expanded

If the algorithm produces a retrieved N_e smaller than 0.1, the pixel is flagged clear. Both above-calculated parameters are used by RTTOV to simulate cloud-affected radiances.

b. Results and discussion

The evaluation of both cloud-detection schemes is fundamental as unfiltered cloud-affected observations can have a detrimental impact on the quality of the analysis if they are assimilated as clear. For this study, a cloud-characterization product based on the Moderate Resolution Imaging Spectroradiometer (MODIS) data has been used as a reference. The MODIS imager is a key instrument on board the Earth Observation System satellites. We have used data product collected from the Aqua platform from the Interactions Clouds Aerosols Radiations Etc. (ICARE) center (more information available online at http://www.icare.univ-lille1.fr/archive/archive.php?dir=MODIS/MYD06_L2/), which are produced at an horizontal resolution of 5 km at nadir and cover a 5-min time interval. Because of downloading resource limitations, the comparison is limited to the Atlantic region. The validation is performed within a 10-day period: from 1 to 10 September 2006. A total of 15 706 AIRS pixels have been processed. As the evaluation of both cloud-detection schemes has already been performed (Dahoui et al. 2005) and results are briefly discussed in this paper.

Performances for both schemes are similar in terms of detection of cloudy and clear pixels. These results confirm previous studies (Dahoui et al. 2005). Figure 1 highlights the accuracy of the cloud detection for the CO₂-slicing and the ECMWF schemes for several ranges of CTP inferred from MODIS. The detection of medium clouds (between 400 and 800 hPa) delivers the best results. In contrast, the detection of low-level and high clouds (CTP > 800 hPa and CTP < 400 hPa, respectively) is not as efficient as for medium clouds for both schemes. For low-level clouds, this must be due to the similarity of the measured signal from clear-sky and low-cloud scenes in some particular situations which make the detection of these low clouds more difficult. Concerning high clouds, the problem is mostly due to the insensitivity of the sounding-based method to thin clouds (identical thresholds are applied for the detection of each type of clouds). Furthermore, clouds whose CTP varies between 250 and 300 hPa and those whose CTP varies between 900 and 1000 hPa are slightly better detected by the ECMWF scheme. Clouds whose CTP varies between 300 and 850 hPa are slightly better detected by the CO₂-slicing scheme.

The retrieved CTP from CO₂-slicing exhibits a correlation of about 0.79 with the CTP inferred from MODIS. Correlation between the N_e retrieved from the CO₂-slicing scheme and the ones inferred from MODIS is quite low for each range of N_e (about 0.51). This low correlation may first be explained by the presence of clouds distributed in the vertical (which are assumed to be single-layer clouds by the RTTOV) whose radiative effect is comparable to a cloud with a higher cloud fraction located at a single level. Moreover, the relative bad detection of low clouds that exhibit a low contrast with the surface in some particular situations may also account for this low correlation. These CTP and N_e correlations are of the same order of magnitude as those found by Pavelin et al. (2008) with the minimum residual method: between 0.67 and 0.81 for the CTP and between 0.46 and 0.78 for the N_e.

a. Methods to determine background cloud parameters

The observation operator used here for the assimilation of cloud-contaminated radiances is the 8.5 version of RTTOV, complemented by horizontal and vertical interpolations. It simulates infrared and microwave radiances observed by satellite sounders (Saunders et al. 2002). For each channel, the optical depth is obtained from linear regressions based on RTTOV input variables: temperature and humidity profiles and surface data. Although RTTOV has initially been designed to be used in clear-sky conditions, it is able to simulate cloud-contaminated radiances. The radiative transfer in cloudy conditions can be performed by two different techniques:

• The cloud is assumed to be a single-layer blackbody of negligible depth that CTP has been provided to the RTM. The cloud-affected radiance is calculated using the following expression:

  

  

  View Expanded

  

  where is the cloudy monochromatic radiance of frequency ν going away from the top of the cloud toward space with an incidence angle θ, τ_ν(z_cld, θ) represents the transmittance between the cloud top and space, T(z_cld) represents the cloud-top temperature, T(z) is the temperature of an atmospheric layer at altitude z, and B_ν represents the Planck function. Equation (4) defines the overcast radiance. In reality, most of the pixels are partially cloud contaminated. These kinds of pixels are defined by a linear blending of overcast and clear radiances using the following equation:

  

  

  View Expanded

  

  where represents the clear-sky component of the radiance and N represents the cloud cover.

• A more realistic multilayer parameterization of clouds is in the RTTOVCLD module: clouds are defined by several vertical levels and for different cloud water phases (liquid or ice) from vertical profiles provided to the RTM (temperature profiles, humidity, cloud cover, liquid cloud water, and cloud ice) by a physical cloud scheme. Clouds are assumed to be multilayer gray bodies.

b. Experiment setup

The NWP model used in this work is ARPEGE (Courtier et al. 1991), which is the Météo-France operational global model. It uses a stretched grid with a horizontal resolution of 23 km over France and 133 km over the antipodes. The assimilation scheme is a 4D-Var assimilation (Courtier et al. 1994; Rabier et al. 2000). It uses the following data sources: land surface stations, sea surface station (drifting buoys, ship reports, etc.), aircraft data, in situ sounding data (radiosonde and pilot balloon reports, etc.), wind profiler radar data, Global Positioning System (GPS) ground-based data, geostationary satellite winds (atmospheric motion vectors), and polar-orbiting satellite radiances and winds. Radiances from polar-orbiting satellites provide ARPEGE with roughly 50% of the total volume of assimilated data. AIRS data represented roughly 65% of the volume of polar-orbiting satellite data and more than 30% of the total volume of assimilated data. All these data sources are subject to various quality control checks and suspect ones are rejected. The quality control applied to AIRS is based on differences between observations and background values, with a rejection of the observation if the difference is larger than a given threshold. Moreover, a geographical thinning of 250 km is applied to all radiance data. Satellite data are bias corrected with an adaptive variational method (viz., VarBC), using geometric and flow-dependent predictors to remove systematic errors in satellite radiance data (Dee 2004; Auligné et al. 2007). In this study, three types of predictors were used: a global offset applied to each channel, a scan angle correction based on the nadir-viewing angle (to the power of 1, 2, and 3) and an air mass correction (predictors are the four atmospheric thicknesses 1000–300, 200–50, 10–1, and 50–5 hPa). We have used the operational subset of 54 channels located in the temperature longwave band chosen among the 324 channels available on the Global Transmission System. This subset contains 19 stratospheric channels with weighting functions peaking from 56 to 85 hPa, 12 channels with weighting functions peaking from 102 to 222 hPa sampling the tropopause, and 23 upper-tropospheric channels with weighting-functions peaking from 253 to 478 hPa (Fig. 2). These channels are only assimilated over seas. The assimilation period of this study covers a full month, from 1 to 30 September 2006. In this approach, once a cloud-affected pixel is detected by the CO₂-slicing scheme, cloud parameters (CTP and N_e) are retrieved from infrared radiance measurements out of the observation operator and are then directly used as inputs of RTTOV, which will then simulate the cloud-affected spectrum. Thus, N_e and CTP are determined at the beginning of the assimilation cycle with no adjustments during the minimization. Furthermore, the ECMWF scheme identifies each channel of the pixel as clear or cloudy. Finally, channels affected by clouds whose CTP ranges between 600 and 950 hPa are assimilated together with clear ones and all other cloud-contaminated channels are rejected from the assimilation scheme. Two assimilation experiments have been run to test the impact of the direct assimilation of cloud-contaminated radiances: the first one is a reference experiment that only assimilates clear AIRS radiances with the others above-described data. This reference experiment is called REF in the following study. The second one assimilates cloud-affected radiances on top of clear radiances and other above-described data sources. In the following study this experiment is called EXP. The adaptive variational bias correction from REF is applied to both clear- and cloud-affected radiances in EXP. An experiment similar to EXP (figures not shown) has been run with its own adaptive bias correction but differences with EXP remain small (e.g., no drift in number of active assimilated observations). To disentangle effects of assimilating cloud-affected radiances from those of bias correction variation, only EXP is discussed in this paper.

c. Application for mid- to low-level clouds

As previously seen, cloud-affected radiances are assimilated here for clouds whose CTP is included between 600 and 950 hPa. On one hand, the rejection of clouds situated higher than 600 hPa is motivated by the assimilation of tropospheric channels (up to 478 hPa) into our assimilation scheme and only clouds situated lower than peaks of the weighting function are selected. Indeed, the assimilation of clouds situated higher than the weighting function peak of a given channel can lead to large biases and root-mean-square (RMS) errors for both background and analysis as shown by Pavelin et al. (2008). On the other hand, it has been shown in section 2b that the detection of low-level clouds (800 < CTP < 950) was not as efficient as the detection of higher clouds and their assimilation could thus appear delicate. A sensitivity study has been performed to check whether or not the AIRS spectrum was affected by the nondetection of these low clouds. It consists in comparing EXP (which simulates cloud-contaminated radiances) and REF (which only simulates clear radiances) averaged innovations (observations minus simulated radiances using the background state of the atmosphere) with respect to the channel number at a given CTP range (800–950, 600–800, 400–600, and 200–400 hPa) for all the observations (Fig. 3). The AIRS spectrum is almost not affected by the simulation inside RTTOV of clouds whose CTP is included between 800 and 950 hPa (Fig. 3a). Difficulties of cloud-detection schemes to detect some low clouds have thus a negligible impact on the background simulation, contrary to other level clouds where the simulated AIRS spectrum is more affected by the simulation of clouds inside RTTOV (Figs. 3b–d). The assimilation of such low-level clouds is also motivated by the large number of retrieved CTP between 900 and 950 hPa (Fig. 4). However, very low clouds (CTP > 950 hPa) were rejected because the huge number of these cloud types (Fig. 4) does not seem physically plausible. The CO₂ slicing is obviously not able to accurately treat these very low clouds, most probably because of their weak impact on the signal measured by the sounder.

Our choice to assimilate cloud-affected channels whose CTP is included between 600 and 950 hPa appears thus to be a good compromise between, on one hand, the rejection of doubtful observations and those potentially spoiling the analysis, and, on the other hand, the assimilation of a nonnegligible volume of additional cloud-affected observations.

a. Number and localization of additional assimilated observations

Figure 5 shows the number of active observations for each assimilated channel for the first assimilation window (0000 UTC 1 September 2006). This assimilation window is particularly interesting because the background used is the same for EXP and REF. The number of active observations is roughly the same for EXP and REF until channel 138, which is one of the first channels whose weighting function is pointing into the tropopause. Previous channels (from 15 to 117) are stratospheric channels, thus seldom cloud contaminated, and exhibit the same behavior for both experiments. Channels 138 to 251 are upper-tropospheric channels and are thus potentially contaminated by clouds. This hypothetical contamination leads to a rejection of these contaminated channels in REF whereas some of these contaminated channels are assimilated in EXP. The lower the channel points in the atmosphere, the larger the probability of a cloud contamination and the larger is the difference between the assimilated observation numbers for the two experiments. Furthermore, the number of active clear observations is lower for EXP than for REF for each tropospheric channel. This may be due to the geographical thinning of the assimilated pixels. This thinning is based on the maximum number of nonrejected channels (monitored and assimilated channels) per profile. REF keeps the profile with the maximum number of clear channels as cloud-contaminated channels are rejected from the assimilation. It is the opposite for EXP that has a maximum number of nonrejected channels in the profile in the presence of a cloud whose CTP is between 600 and 950 hPa (in that case, all channels are nonrejected, except for the gross error).

Figures 6a,b, respectively, exhibit the number of assimilated pixels and assimilated channels (many channels are potentially assimilated in a given pixel) for both experiments with respect to their latitudinal position and for the whole study period. Most of the active pixels and channels are situated in the Southern Hemisphere, which is consistent with the assimilation of overseas observations (the oceanic surface is larger in the Southern Hemisphere than in the Northern Hemisphere). EXP assimilates between 1% and 4% more pixels than REF (Fig. 6a) with a slightly larger rate at mid- to high latitudes. The 250-km geographical thinning restricts these rates that would be higher if the geographical density of assimilated radiances were higher. The difference of assimilated channels between EXP and REF in mid- to high latitudes (Fig. 6b) is much larger than the difference of assimilated pixels (Fig. 6a): on average, 10% of additional channels against 3.5% of additional pixels. These differences are not visible for low latitudes (1.5% of additional pixels and channels on average). Thus, even if the assimilation of cloud-affected radiances yields additional active pixels to ARPEGE with quite a uniform repartition on the globe, the additional information is much larger for mid- to high latitudes than for low latitudes. This could be due to the nonassimilation of high-level clouds (CTP < 600 hPa) that are most frequent in the tropics [see the International Satellite Cloud Climatology Project (ISCCP) Web site online at http://isccp.giss.nasa.gov].

The daily evolution of the number of assimilated observations has also been studied at 1200 UTC (not shown) for assimilated channels peaking at the tropopause, upper troposphere, and midtroposphere in order to check whether the quality control was stable all along the assimilation period. For these channels, the number of active observations is constant for both experiments all along the study period, which proves that the quality control (and the cloud detection) is reliable.

b. Residual bias of observations

Figure 7a compares innovations of REF and EXP for each channel and for all observations (used and rejected observations) for the first assimilation window. Only the impact of the brightness temperature simulation by CO₂ slicing is thus seen in this figure. Innovation amplitudes fluctuate from −8 to 3 K for REF, with an absolute amplitude maximum for longwave surface peaking channels (from 759.6 to 1030.5 cm⁻¹ and from 1061.8 to 1135.6 cm⁻¹) and from 1.5 to 4 K for EXP (with an isolated peak at −5.5 K for channels 72 and 73). Innovations obtained by EXP are thus globally 4 times smaller than the ones obtained by REF for surface channels. It globally represents a cooling of the background estimation of about 6 K. Another significant cooling (from 6 to 3 K) is observed for lower-tropospheric peaking channels sensitive to water vapor (from 1218.5 to 1392.2 cm⁻¹). A smaller cooling (1–2 K) is observed for upper-tropospheric peaking channels sensitive to water vapor (from 1397.1 to 1605.1 cm⁻¹). Each shortwave channel (except shortwave stratospheric channels) presents also a significant cooling of about 4 K (from 2181.9 to 2298.7 cm⁻¹ and from 2380.4 to 2664.1 cm⁻¹). These differences show that to take into account cloud parameters into the radiative transfer calculations leads to model equivalents more consistent with observations. Figure 7b is a zoom on the active channel part of the previous figure. Innovations for EXP and REF are similar from 651.1 to 701.1 cm⁻¹ (stratospheric channels). Innovation amplitude differences are larger from 702.5 to 721.5 cm⁻¹, which are tropospheric peaking channels and are potentially more cloud contaminated. We have verified for the last day of our assimilation period (30 September 2006) at 0000 UTC that the variational bias correction was efficient throughout the assimilation period: innovation amplitudes are of the same order than those observed in the first assimilation window for both experiments.

a. Impact on analysis

EXP and REF statistics (differences in terms of biases and RMS) from various types of assimilated observations (conventional and satellite data) have been studied over the whole assimilation period. Both background and analysis statistics are in the majority of cases equivalent for both EXP and REF (not shown), but slight improvements have been noticed. With regards to conventional data, a slight reduction of, respectively, the bias and the RMS is observed for winds from profilers in the stratosphere over the Southern Hemisphere and in the upper-troposphere over the Northern Hemisphere. Regarding satellite data, a better fit to microwave sounders sensitive to atmospheric humidity has been noticed: analysis and background bias reduction for the Special Sensor Microwave Imager (SSM/I) in the tropics and over the Southern Hemisphere in the wider atmosphere and a background bias reduction of midtropospheric channels of the Advanced Microwave Sounding Unit-B (AMSU-B) in the tropics.

Figure 8a highlights differences between EXP and REF analyses for the temperature at 500 hPa for the first analysis. Differences in model fields are, as expected, located in places where cloud-affected observations are assimilated in EXP showing that these differences are mainly due to the assimilation of cloud-affected observations. These differences are globally located in the Northern Hemisphere between 40° and 90°N and in the Southern Hemisphere between 30° and 70°S. By contrast, the intertropical zone is almost not affected by the assimilation of cloud-affected observations. Furthermore, our assimilation period spreads out the austral winter, where oceans below 70°S are fully covered by sea ice, which is not the case for oceans above 70°N, mostly free from ice. Our assimilation system includes a test that rejects observations situated over sea ice, therefore explaining the negligible impact from 80° to 65°S.

A longitudinal cross section of analysis temperature differences between EXP and REF is shown (Fig. 8b) for the first analysis to characterize the impact of the assimilation of some cloud-affected observations on the analysis with respect to their diagnosed top pressure (from the CO₂ slicing). The cross section at 50°S from 40°W to 40°E is shown because this area displays many active cloud-contaminated observations with different retrieved CTP. From 30° to 20°W, medium clouds are assimilated (600 < CTP < 700 hPa): a warming spreading from the CTP down to the surface is visible. A small cooling is also visible above the CTP up to the upper troposphere (350–400 hPa). From 20°W to 0° and from 20° to 40°E, very low clouds are assimilated (900 < CTP < 950 hPa): a small cooling is visible just above the CTP spreading within the whole lower troposphere (until 650–700 hPa). The upper troposphere is also affected from 600 to 300 hPa: a warming is indeed observed. For very low clouds, these changes can also be accompanied by a slight cooling of the tropopause. Even if these changes highlight the capacity of the cloud-affected radiances to modify the analysis, the lack of verifying data prevents us from assessing which analysis better fits the real state of the atmosphere.

b. Impact on forecasts

The impact of the assimilation of cloud-affected radiances on global forecasts was determined by comparing the objective forecast scores in terms of RMS errors with respect to radiosonde observations. Table 1 shows the RMS forecast error differences for the geopotential height. These errors have been computed with respect to radiosonde observations for the validation period ranging from 0000 UTC 1 September 2006 to 1800 UTC 4 October 2006. The impact is rather neutral to slightly positive. The most significant improvement was obtained for the geopotential height according to a statistical Student’s t test with respect to radiosonde observations with a level of confidence of 95%. Over the Northern Hemisphere, positive impacts are found for a 24–96-h forecast range in the stratosphere and in the mid- to upper troposphere. Improvements at 100 and 250 hPa for a 48-h forecast range are statistically significant. Over the Southern Hemisphere, a positive impact is found for a 24–48-h forecast range in the whole atmosphere. These improvements are significant at the 24–48-h forecast range at 850 and 250 hPa and for the 48–72-h forecast range at 100 hPa. A negative impact has been observed at longer forecast ranges but it is not statistically significant. Over the tropics, positive impacts are found for the 24–72-h forecast range in the stratosphere and in the mid- to low troposphere. These improvements are significant at 850 and 100 hPa for the 24–48-h forecast range. Over Europe, the impact is mainly positive until the 72-h forecast range in the whole atmosphere and negative for the 96-h forecast range. These positive impacts are larger than for the other domains but only significant at 100 hPa at the 48-h forecast range. Scores with respect to ECMWF analyses are rather neutral to slightly positive.

Regarding the other parameters (temperature, humidity, and wind), impacts are roughly neutral to slightly positive but not significant with respect to radiosonde observations and ECMWF analyses. Similar results were found over another 3-week period (from 0000 UTC 29 May to 1800 UTC 20 June 2008) in a different context: more satellite data were assimilated (ATOVS from MetOp, IASI, and GPS radio occultation observations) and the model resolution was higher.

Figures 9a–d represent RMS errors and biases of the 48-h forecast over Europe for the geopotential, temperature, humidity, and wind, respectively. The geopotential is improved in terms of RMS in the whole atmosphere and in terms of bias in the stratosphere. These improvements are significant from 10 to 200 hPa. The temperature is slightly improved in terms of bias and RMS errors from the surface to the tropopause with significant impacts at 200 hPa. The humidity is slightly improved in terms of RMS in the upper troposphere and in the tropopause but not significantly. The wind is improved in terms of RMS errors from the surface to the tropopause and from the upper troposphere to the tropopause in terms of bias. These improvements are significant from 100 to 200 hPa.

The impact of the assimilation of cloud-affected observations on forecasts will now be further discussed by an evaluation of the predictability on a meteorological situation: the so-called Mediterranean hurricane (MEDICANE) case.

a. Synoptic description of the event by ECMWF analyses

On 26 September 2006, a strong mesoscale storm with some resemblance to a polar low hit the southeast part of Italy and caused intense precipitation. This event has been investigated from an observational and a numerical (analysis and forecasts from WRF model) point of view by Moscatello et al. (2008). In this paper, this description has been completed by analyses from the ECMWF model, which will be used as a verification to display the relative performance in terms of predictability of both experiments. The trajectory and the evolution of the mean sea level pressure (MSLP) of the low with respect to each time step are summarized in Fig. 10. ECMWF analyses show that a small-scale cyclone was generated in the southeast of the atlas Tunisian Ridge near the Algerian–Tunisian border at around 0000 UTC 25 September 2006 with a MSLP of about 1007 hPa (Fig. 11a). During the following 18 h, the low moved through an east–northeastward direction without intensifying to reach eastern Malta Island at some 800 km away from its initial position (Fig. 11b) at 1800 UTC 25 September. At 0000 UTC 26 September, the low entered the Ionian Sea near the Ionian coast of Calabria. At this stage, it began to intensify reaching a MSLP of 1005 hPa. Between 0000 and 0600 UTC, the low moved northeasterly certainly in response to a low pressure system (1001 hPa) visible in the Tyrrhenian Sea (Moscatello et al. 2008) to reach the southern Torente Gulf at 0600 UTC (Fig. 11c). All along this time, the low strongly intensified, reaching a MSLP of about 996 hPa at 0600 UTC. Between 0600 and 1200 UTC, the trajectory of the low curved northward and the depression then crossed the Apulia regions to reach the Adriatic Sea with a MSLP of about 994 hPa, the absolute minimum of the depression at 1200 UTC 26 September. For the next 6 h, the low moved along a northwestward direction and reached the Gargano promontory at 1800 UTC (Fig. 11d) with decreasing intensity (997 hPa). The low then moved inland where it finally died.

b. Predictability of the event

Four sets of numerical simulations have been performed for both experiments in order to compare the predictability, position, and intensity of the cyclone during its whole development, from 0000 UTC 25 September 2006 to 1800 UTC 26 September 2006.

1. The first set, hereinafter called FCST2300, begins at 0000 UTC 23 September 2006 and ends at 1800 UTC 26 September 2006 after a 90-h forecast.

2. The second set, FCST2312, begins at 1200 UTC 23 September and ends after a 78-h forecast.

3. The third set, FCST2400 begins at 0000 UTC 24 September and ends after a 66-h forecast.

4. The fourth set, FCST2412 begins at 1200 UTC 24 September and ends after a 54-h forecast.

Figures 13a,b highlight the evolution of forecast MSLP minimum by FCST2300 and FCST2312 and FCST2400 and FCST2412, respectively, with respect to the ECMWF analysis. The evolution of the intensity of the low is better forecast by EXP than REF throughout its whole evolution period for each forecast set. FCST2400 displays the largest averaged difference between EXP and REF (Fig. 13b): differences of MSLP between EXP forecasts and ECMWF analysis are 1.5 hPa on the mean against 3.5 hPa for REF. Furthermore, these differences between EXP and REF tend to grow with the forecast range (at least for FCST2400).

c. Accumulated precipitation

According to Moscatello et al. (2008), the storm caused intense precipitation over the southeastern part of Italy. The first MEDICANE landfall occurs between 0600 and 1200 UTC 26 September 2006 over the Salentine Peninsula. Equitable threat scores (ETS), which are defined as the intersection of the forecast and observations divided by the union of the forecast and observations at a given threshold, have thus been calculated (Fig. 14). These scores highlight forecast performance of both experiments in terms of cumulated precipitation (within a 6-h window). The closer the score is to 1, the better the forecast. These scores are calculated with respect to cumulated precipitation observations (SYNOP situated between 28° and 55°N and between 0° and 30°E) available from 0600 to 1200 UTC 26 September 2006 for four differents threshold (0.1, 1, 5, and 10 mm) and according to seven different forecast ranges (6, 12, 36, 48, 60, 72, and 84 h).

For very low precipitation (threshold = 0.1 mm), performances of both experiments are quite bad at each forecast range (not shown) but they are slightly improved in EXP. The same assessment can be made regarding a threshold of precipitations of 1 mm (not shown). However, for longtime forecast ranges (from 72 to 84 h), performances of REF are slightly better than EXP for this threshold. As can be seen in Figs. 14a,b, regarding thresholds of 5 mm (steady precipitation) and 10 mm (strong precipitation), respectively, EXP globally exhibits better scores with respect to observations than does REF. For steady precipitation (threshold = 5 mm), performances of both experiments are better than for lower precipitations (Fig. 14a) with better performances of EXP at each forecast range except for the 84-h forecast range. Regarding stronger precipitation (threshold = 10 mm), both performances are reasonable (Fig. 14b) but EXP also exhibits better performances at each forecast range. Cumulated precipitation forecasts are thus improved whatever the rainfall rate but especially for steady and strong precipitation.