Preliminary Assimilation of Satellite Derived Land Surface Temperature from SEVIRI in the Surface Scheme of the AROME-France Model

1.1 AROME-France configuration

AROME is the limited area convective-scale model of Météo-France in operations since December 2008 ((Seity et al., 2011) and (Brousseau et al., 2016)). AROME has been developed to improve the forecast of mesoscale phenomena such as fog and thunderstorms. The AROME-France domain covers the area centered over France from 12.45° W to 16.67° E and from 37.26° N to 55.69° N as shown in Figure 1 with 53% of the domain surface covered by land.

Figure 1

AROME is a spectral model derived from the global NWP model ARPEGE (Action de Recherche Petite Echelle Grande Echelle, (Courtier et al., 1991)). The current operational version (2020) used for this study has a 1.3 km × 1.3 km grid and 90 vertical levels from 5 m high up to 10 hPa. AROME is coupled to the ARPEGE global model and produces eight forecasts everyday at 00, 03, 06, 09, 12, 15, 18 and 21 UTC with a forecast range of 48 h for 00 and 12 UTC runs, of 45 h for the 03 UTC run, of 42 h for the 06 and 18 UTC runs and of 7 h for the 09, 15 and 21 UTC runs. AROME uses the dry thermal and shallow convection mass-flux convection scheme (Pergaud et al., 2009) and the TKE (Turbulent Kinetic Energy) turbulence scheme (Cuxart et al., 2000). In terms of radiation, the RRTM scheme (Mlawer et al., 1997) is used for long wave spectrum while the Fouquart-Morcrette scheme is used for the short wave spectrum. The land surface scheme uses SURFEX-ISBA-3L (Interaction Soil-Biosphere-Atmosphere) ((Noilhan and Planton, 1989) and (Boone et al., 1999)) with a surface initialization based on the force-restore method (2 layers for soil temperature and 3 layers for soil moisture content) with a dedicated analysis scheme (Giard and Bazile, 2000) described in the next section.

1.2 The 3D-var assimilation system in AROME-France model

AROME-France model uses a 3D variationnal assimilation scheme with 1 h assimilation window. The assimilated observations vary during the day but they consist in radiosondes, wind profilers, aircraft reports, ship and buoy reports, automated land surface stations, satellite observations, Global Positioning System Zenith Tropospheric Delay, and Doppler radar wind and humidity observations from radar reflectivities. Both satellite and conventional observations are assimilated in AROME upper air assimilation. Among these, radar observations represent the major part of assimilated conventional observations (around 75%) and represent around 60% of the total assimilated observations. Satellite observations represent around 20% of the total and are composed of wind observations from atmospheric motion vectors and scatterometers, microwave and infrared radiances. Infrared observations represent around 75% of satellite observations, with half of them coming from SEVIRI radiometer onboard Meteosat Second Generation, and the other half provided by IASI (Infrared Atmospheric Sounding Interferometer) hyperspectral instrument onboard Metop platforms.

The background used here is a 1h-forecast and background error statistics determine how an observation can modify analysed fields and how the information propagates from the observation location to other model points and to unobserved model variables. The background error statistics are estimated using forecast error statistics of vorticity, divergence, temperature, surface pressure and specific humidity. The background error covariance matrix is a climatological matrix, spatially homogeneous and isotropic, and is estimated using (Berre, 2000) multivariate formulation with an ensemble of differences of 3h range forecasts ((Brousseau et al., 2016)). The background error covariances were calculated from two periods of 15 days each of ensemble data assimilation with a limited number of members (6) at high resolution, one in summer and one in winter. It includes static covariances between variables and between different points of the 3D space for each variable. Recent developments within AROME model have proved the positive impact of a B matrix derived from ensemble data assimilation of the day, including covariances between variables and levels of the atmosphere and points of the space. This new B matrix results in a 3D-EnVar which improves forecasts of extreme events such as convective precipitation and winds. Concerning surface assimilation, some developments are ongoing to improve the optimal interpolation scheme and move towards a variational framework which could also benefit from the ensemble data assimilation. A 2D-EnVar is under development for AROME model surface assimilation with a B matrix derived from ensemble data assimilation of the day, including spatial covariances and covariances between variables, which could alleviate some of the optimal interpolation drawbacks which is univariate and has static background error statistics.

The variational analysis is performed every hour at the resolution of the model. The spin-up was reduced using a background error covariance matrix calculated from forecasts coming from an ensemble data assimilation at the AROME-France resolution, in order to simulate the evolutions of errors during the assimilation cycle (Brousseau et al., 2016). The increments obtained with these background errror statistics are more balanced and avoid the part of the spin-up coming from imbalanced analysis increments. Residual spin-up comes from imbalance between the model fields, and between the upper-air analysed fields and the surface fields that come from surface assimilation. It is noteworthy that the long-range (48 h) forecasts at 00 and 12 UTC also benefit from the analysis at 01 and 13 UTC, respectively, using an incremental analysis update (IAU) in order to improve their quality. It has been proven that the quality of forecasts updated using IAU is equivalent to the quality of forecasts initialized one hour later.

1.3 LST retrieval in AROME-France model

Satellite infrared and microwave instruments do not observe the Earth directly but measure the radiation reaching the top of the atmosphere. These radiations are the sum of the radiation emitted by the Earth surface, by the atmosphere and the radiation received and reemitted by the surface and the atmosphere. The approach of LST retrieval is based on the use of window channels for which the atmosphere is transparent in a way that the radiances observed by the satellite instruments result from the Earth surface emission only and are not impacted by the atmosphere in the absence of clouds. In order to calculate the surface parameters knowing the atmospheric state, the LST or surface emissivity retrieval needs to invert the radiative transfer equation ((Guedj et al., 2011) and (Karbou et al., 2006)):

where ∊_v is the surface emissivity, R_v(θ) is the observed radiance, θ_v is the incidence angle, Γ_v is the atmospheric transmittance, T_s is the surface temperature, $L_{\nu }^{\uparrow }$ and $L_{\nu }^{\downarrow }$ are respectively the up-welling and the down-welling radiances for the channel ν. In case of surface temperature retrieval for infrared channels, the emissivity is taken from an emissivity atlas (Seemann, 2007) and the LST is given by the inverse form of the radiative transfer equation:

where L is the Planck function. Given the higher temporal and spatial variation of the surface temperature compared to 2 m temperature and the inaccurate representation of surface conditions in the model, previous studies (Karbou et al., 2006) and (Guedj et al., 2011) have shown the interest of using the retrieved surface parameters for the satellite radiances data assimilation. At Météo-France, the retrieval of surface temperature is performed for infrared instruments assimilation (SEVIRI and IASI) in clear sky. On the contrary, the retrieved surface emissivity is used to assimilate the microwave instruments knowing the larger variation of surface emissivity in the microwave bandwidth. The objective of the study is to assimilate the retrieved land surface temperature from SEVIRI in the land surface model of AROME model. (Sassi et al., 2019) have shown that the representation of the diurnal cycle with satellite derived land surface temperature is more accurate than the representation of the model, especially in spring and summer. It has also been shown in a previous study (Sassi et al., 2019) that the use of the land surface temperature retrieved from an instrument was useful for the simulation of the brightness temperature of other instruments and that it improves the simulations using the model surface temperature. These results were encouraging to develop the assimilation of satellite retrieved land surface temperatures in the surface scheme.

2.1 SURFEX

SURFEX (Externalized surface) is a surface modelling platform developed at Météo-France (Masson et al., 2013) that can be used in offline mode or coupled to atmospheric models such as AROME. In coupled mode, the atmospheric model and SURFEX exchange atmospheric fluxes (radiation, precipitation) and near surface meteorological parameters (temperature, humidity, wind) from the atmosphere, and surface turbulent fluxes (heat, moisture, momentum) and surface temperature from the surface at each time step, as described in Figure 2. These surface quantities are computed over four different tiles representing the nature of the soil: nature (soil/vegetation), town (urban areas), lake and sea. Open water areas are assumed to be saturated surfaces with a temperature imposed from an external analysis. On the contrary, for the other tiles (nature and town), explicit surface energy and water balances are solved at each model time step. The Town Energy Balance (TEB) allows the estimation of various surface temperatures (roads, building walls and roofs) over urban areas. Details can be found in (Masson, 2000) but they are not useful for the current study which is focused on improving the soil temperatures of the nature tile. Here, the ISBA (Interactions Soil Biosphere Atmosphere) parametrization is used for the nature tile ((Noilhan and Planton, 1989) and (Noilhan and Mahfouf, 1996)) in its three-layer version (Boone et al., 1999). It predicts the evolution of two soil temperatures (a superfical one TG1 associated with the diurnal forcing τ₁ (1 day) and a deeper one TG2 associated with longer time scales τ₂ = 2πτ₁ (about 6 days)) using the force-restore method first proposed by (Bhumralkar, 1975). It is important to point out that over snow-free surfaces, the surface temperature TG1 is representative of a mean value accounting for both bare soil and vegetation contributions in the nature tile (one single energy balance is solved). Such a mean value makes the comparison with satellite estimates rather straightforward, since the land surface temperature seen by a radiometer is a mixture of the longwave emission from various surface types present in the pixel. Regarding moisture reservoirs, in summer conditions (i.e. no soil water freezing), four prognostic variables are evolved with time : an interception reservoir Wr for the vegetation (rain, dew), a superficial soil reservoir Wg1 (that controls bare soil evaporation), a root-zone reservoir Wg2 (that controls plant transpiration), and a drainage reservoir Wg3 (that interacts with ground hydrology). For NWP, only the soil moisture reservoirs Wg1 and Wg2 are of importance in terms of initial contents (Mahfouf et al., 2009).

Figure 2

2.2 The surface analysis system

A dedicated surface analysis scheme allows the initialisation of the soil prognostic variables of ISBA (Tg1, Tg2, Wg1, Wg2) every 3 hours even though the atmospheric analysis is performed every hour (this can be understood because deep soil variables evolve slowly in time with respect to atmospheric processes). This scheme that is described in (Giard and Bazile, 2000) is based on atmospheric increments of screen-level temperature (T2m) and relative humidity (RH2m), measured at two meters by synoptic (SYNOP) weather stations, in order to correct the four soil variables (Tg1, Tg2, Wg1 and Wg2). Therefore a first step is a bidimensional analysis of T2m and RH2m using surface weather stations (SYNOP reports and the French high density surface station network RADOME) and a short-range AROME forecast as background field. This is done by a univariate optimum interpolation scheme with mostly homogeneous and isotropic structure functions (some dependency with orography are accounted for as explained in (Soci et al., 2016)). The structure functions for T2m and Hu2m are as follows:

where dist is the distance between the observation and the analysis point and D is a characteristic distance, D is set to 100 km for T2m and Hu2m in AROME model. g(Δ_alti,Δ_LSM) is the function accounting for differences between the land-sea mask of the observation and of the analysis point, and between the altitude of the observation and the analysis point. The T2m and Hu2m data benefit from an advanced quality control on observations. Observations with missing or wrong orography are rejected. Observations for which the difference between observation and the background is larger than $5\sqrt{{\sigma }_o^2+{\sigma }_b^2}$ for T2m for SYNOP and buoy observations, or larger than $2.5\sqrt{{\sigma }_o^2+{\sigma }_b^2}$ for Hu2m are rejected before the assimilation. Observations for which the innovation (obs-guess) is larger than $3.5\sqrt{{\sigma }_o^2+{\sigma }_b^2}$ for T2m or larger than $1.75\sqrt{{\sigma }_o^2+{\sigma }_b^2}$ for Hu2m are checked after analysis and they are kept if the analysis departure (obs-analysis) if lower than $3.5\sqrt{{\sigma }_o^2+{\sigma }_b^2}$ for T2m or $1.75\sqrt{{\sigma }_o^2+{\sigma }_b^2}$ for Hu2m (Stjepan, 2016).

In a second step, the T2m and RH2m analysis increments are used to correct the soil variables with linear relations. For soil moisture, optimal interpolation coefficients depending upon surface properties and atmospheric conditions as proposed by (Giard and Bazile, 2000) are used operationally. For soil temperature, these are empirical corrections initially proposed by (Coiffier, 1986) for a previous land surface scheme and adapted by (Giard and Bazile, 2000) to the ISBA scheme (Mahfouf et al., 2009), that is:

The atmospheric correction is imposed as such to the surface temperature Tg1 and is reduced for the deeper temperature Tg2 (to account implicitly for the heat wave damping). A similar scheme is also used at ECMWF for soil temperature (Douville et al., 2000) despite the use of a Simplified Extended Kalman Filter for soil moisture analysis (De Rosnay et al., 2013). The only difference with the Météo-France set-up is the use of an empirical function that prevents from correcting soil temperatures during daytime (since errors are more likely to arise from soil moisture from the turbulent fluxes). This operational methodology has been modified in order to include SEVIRI LST. First the bidimensional optimal interpolation has been adapted from T2m and RH2m observations to allow the bidimensional optimal interpolation of surface temperature with SEVIRI LST observations and the ISBA surface temperature Tg1 as a background field (1h forecast). The various parameters defining background and observation errors statistics (variances and correlation length) have been setup from the results of the diagnostic studies presented in Section 4. Then, the analysis increments ΔLST are used to modify the soil temperature increments as:

This first choice of equal weights between the two increments will have to be revised in the future when more optimal coefficients will be used for the soil temperature analysis. It allows to account for the LST information in the soil layers without discarding the information from screen-level variables. As explained later the use of SEVIRI LST in the soil analysis will be restricted to night time periods, in agreement with results from (Mahfouf, 2007) and the ECMWF set-up. When SEVIRI LSTs are not used, the soil temperature corrections are provided by equation 4. Figure 3 summarizes the implementation of SEVIRI LST assimilation in the surface analysis system together with the assimilation of screen-level data from surface observations.

Figure 3

3.1 Diagnostics of observation errors

Using the Desroziers approach described in (Desroziers et al., 2005), the observation error covariances are diagnosed based on the analysis and the background departure statistics. The consistency check verifies the equation:

where R is the observation error covariance matrix, $d_b^o$ is the innovation vector, $d_a^o$ is the vector of differences between the observation and the analysis in the observation space and E is the mathematical expectation. The initial a priori values for the observation and background error standard deviation are 3 K and 1.5 K, respectively. As the diagnosed values are rather close to the arbitrary a priori values, only one iteration has been carried out (Bathmann, 2018). Following conclusions of (Sassi et al., 2019) that have shown that nighttime observations are closer to in-situ surface temperature observations than the model surface temperature, only nighttime observations have been considered here. Therefore only observations at 21, 00 and 03 UTC analysis times are first kept. However large observation standard deviations have been observed at 21 UTC, greater than 5 K over a large part of the domain, that is why the 21 UTC analysis time has been excluded form further experiments. The standard deviations over July and August 2019 of the observation errors for 00 and 03 UTC show spatial variability over the AROME-France domain as described in Figure 4 with smaller mean standard deviations over the southern part of the domain especially over the Iberian peninsula and higher values to the North especially over the United Kingdom.

Figure 4

The examination of the amount of observations used to calculate the standard deviation shows a consistent spatial variability of the number of available observations with the values of the diagnosed standard deviation. The higher standard deviations were observed in areas with a fewer number of clear observations, which might be due to a higher cloudiness. In fact, more observations are available over the southern part of the domain (around 100 clear observations per pixel) than over the norther part (around 50 clear observations per pixel), and larger standard deviations in average are observed over the northern part of the domain. Standard deviations of observation errors vary from 1.5 K to 4.5 K at some places.

Over the major part of the domain the standard deviation is between 1.5 and 2.5 K. Slightly larger observation errors up to 3 K are observed over some parts of the Iberian peninsula, which can be related to semi-arid regions. Larger values over 3 K and up to 4.5 K are observed over the northern part of the domain, which can be related to regions with a smaller amount of pixels entering the analysis. These regions are also characterized by wet soils and possible undetected or misclassified clouds impacting the diagnostics, which should be done only for clear sky.

Hereafter we have applied an orography filter that rejects observations over high or complex orography. In fact, pixels with altitude higher than 1000 m and those for which the standard deviation of altitude points with surrounding pixels (in a 10 km radius) is above 100 m have been discarded in the computation of the error statistics. This double filter enabled rejecting the observations above orographic areas and also the areas with high orographic gradients while keeping observations on mid-altitude plateaux. The rejected observations are located mainly over the Alps, the Pyrenees, the Massif Central, the Iberic Monts, the Guadarrama mounts and the Cantabriques mounts (as shown in Figure 4).

3.2 Diagnostics of background errors

In addition to observation error covariances, we have diagnosed the background error covariances using a similar approach based on the Desroziers method (Desroziers et al., 2005).

where B is the model error covariance matrix, H is the tangent linear version of the observation operator and H^T its adjoint version, $d_b^a$ is the vector of differences between guess and analysis, and $d_b^o$ is the innovation vector.

Figure 5 shows the mean background error standard deviations for 00 h and 03 h analysis times of July and August 2019. Background errors appear to be more homogeneous compared to the observation ones as displayed in figure 4. We also notice smaller standard deviation in average for the background errors compared to the observation errors with maximum values around 1 K in most cases. However, these values exceed 1.5 K over few locations such as near mountainous areas for example over the Iberian peninsula. This fact might be explained by a misrepresentation of sub-grid heterogeneities. According to these diagnostics, we considered as tunings for the following part an observation error of 3 K and a background error of 1.8 K for the SEVIRI LST assimilation. Note that the observation and background errors for the 2 m temperature are respectively of 1.4 and 1.6 K as in the operational version of AROME. Moreover, diagnostics of spatial error correlations have been carried out on the background surface temperature (not shown). They show that from 10 km the correlation between two points is reduced by 50% and from 30 km background error correlation is close to zero and the points can be considered as uncorrelated. Therefore we chose a correlation length of 30 km for LST error. In comparison the correlation length of 2 m temperature error is 100 km in the operational version of AROME and is kept to the same value in the following experiments.

Figure 5

4.1 A case study: La Loire basin sub-domain

For a first experimental set-up, we chose a summer period covering July and August 2019 to assimilate clear sky SEVIRI LST to optimize the number of clear sky data. We rely on the NWC SAF (Le Gleau, 2019) cloud classification provided with the SEVIRI data to select pixels classified as “clear over land”. A spatial thinning is also applied with one pixel retained out of five. This corresponds to the operational thinning ratio used in the atmospheric analysis. Finally only nighttime pixels are assimilated since the observation error standard deviations are rather low by nighttime, as demonstrated in the previous section, and as previous studies showed that LSTs retrieved from different infrared sensors agree better during nighttime (Sassi et al., 2019). In order to understand the impact of the SEVIRI LST assimilation, we considered a sub-domain within the AROME-France domain and located in the North-Western part of France, near the Loire river estuary (3 W – 0.5 E, 46,5N 48,5 N).

Figure 6 shows the surface temperature from the 1h forecast in addition to SEVIRI retrieved LST (squares) and 2 m temperature (circles) for July, 4th 2019 at 03 UTC on the Loire basin sub-domain. The south-western part of the domain (surface temperature varying between 291 and 292 K) corresponds to the Atlantic Ocean. Figure 6 shows a high spatial variation of surface temperature in the background from the North to the South, with up to 5 degrees difference. Moreover, we notice a global qualitative agreement between the model surface temperature, the SEVIRI LSTs and the 2 m temperatures. However, few SEVIRI LSTs show higher differences (colder by more than 4 K) with model surface temperature and other neighboring SEVIRI LST and 2 m temperature observations. To help understanding these differences, we show in Figure 7 the surface temperature differences between analysis and guess while assimilating SEVIRI LST only (no 2 m temperature observations have been assimilated at this stage).

Figure 6

Figure 7

Two large negative areas of LST analysis increments can be noticed at 47.75 N, 0.5 W and 47.4 N, 2 W. After having performed 2D OI SEVIRI using LST and 2m temperature analysis, surface temperature analysis increments can be derived as shown in Figure 8. The use of 2m temperature analysis, warmer than SEVIRI LST by more than 2 K over the two negative areas has reduced the amplitude of these increments.

Figure 8

From this example, the SEVIRI LSTs show high variability at small spatial scales. One possible reason might be the presence of clouds misclassified as clear sky pixels that could induce the observed cold bias. To further investigate this hypothesis, we evaluated the first guess departures distribution for the SEVIRI LST for July and August 2019 at 03 UTC as shown in Figure 9.

Figure 9

We notice a cold residue that might be due to a mismatch in cloud type detection. In order to reduce such impact on the assimilation, we considered, as an additional criteria, a background departure threshold of –4.5 K below which the LST is rejected from the analysis.

After evaluating the impact of SEVIRI LST on surface temperature analysis, and in order to assess its impact on atmospheric parameters, we have performed a two month experiment (EXP-ARO) in which we consider SEVIRI LST assimilation at 00 and 03 UTC on top of 2 m temperature and relative humidity already assimilated every three hours in the operational version of AROME model surface analysis. We took then into account the findings of the previous work to improve the experimental set-up. While the observation and background errors of 2 m temperature remain equivalent to the operational version, we maintained the observation and background errors of LST, previously defined for the case studies, at respectively 3 and 1.8 K and also the correlation length of LST errors at 30 km. Table 1 summarizes the main settings of the assimilation experiment (EXP-ARO) and also the control experiment (REF-ARO) that reproduces the operational AROME configuration. We present in the following sections the impact of SEVIRI LST assimilation, first on the assimilation of different parameters and then on AROME forecasts.

4.2 Impact on assimilation

The first step to evaluate the SEVIRI LST assimilation in the AROME model was to evaluate its impact on surface and atmospheric analyses in terms of improvement of model guess compared to observations. For temperature, we evaluated the statistics of innovations (observation – guess) at 2 m compared to surface stations observations (synoptic stations). Figure 10 shows the mean differences between the 2 m temperature observations and the model short range (1 h) forecasts as a function of the analysis time for EXP-ARO and REF-ARO experiments.

Figure 10

Figure 10 shows, during night-time (up to 6 UTC), a slight decrease in the 2 m temperature mean innovation difference for EXP-ARO compared to REF-ARO, which represents an improvement of the background quality as it is closer to independent observations. This is consistent with the fact that SEVIRI LST have been used only at 00 and 03 UTC. Moreover, we notice a slight decrease in the mean differences on the last assimilation times of the day (19 UTC – 00 UTC). This might be due to a better short range forecast closer to the observation, due to better surface temperature in the forecast. This improvement indicates that the assimilation of SEVIRI LST has still an impact 15 to 21 hours later the assimilation time (at 00 and 03 UTC). However, a slight increase in the mean differences for EXP-ARO with respect to REF-ARO are noticed around noon (10 UTC – 14 UTC) despite the fact that no SEVIRI LST are assimilated during the day.

In addition to 2 m temperature innovations, we compared the mean differences between 2 m relative humidity observation and model background for EXP-ARO and REF-ARO as described in Figure 11.

Figure 11

Consistently with what has been noticed for the 2 m temperature, the mean differences slightly decrease with EXP-ARO compared to REF-ARO not only during the first assimilation times but also around noon and at the end of the day. However, the impact is very small after 6 h.

Furthermore, in order to evaluate the impact of SEVIRI LST assimilation on the assimilation of satellite observations, we considered microwave instruments in the atmospheric component of the model, since the model surface temperature is used to retrieve surface emissivity in a window channel for each microwave instrument in order to assimilate the other channels. Consequently, a modification of the model surface temperature is expected to induce an impact on the model equivalent to observations and then on the assimilation of microwave channels. For infrared sensors, as mentioned before, a LST is retrieved separately for each sensor to assimilate its sounding channels. We present in Figure 12 the impact of the SEVIRI LST assimilation on MHS microwave sensor which has shown the most significant impact, especially for the channel 5 (183 GHz+/– 7) sensitive to water vapour absorption at 700 hPa in terms of RMSE (Root Mean Square Error) differences in %.

Figure 12

We notice an improvement in MHS channel 5 assimilation with EXP-ARO over most of analyses times (corresponding to negative values of relative RMSE differences), and not only around 0 and 3 UTC at which SEVIRI LST are assimilated. This could result from a more realistic surface temperature obtained in EXP-ARO compared to REF-ARO which propagates in time. It might be interesting to evaluate the impact of the assimilation of LST at all analysis times, also during daytime with adequate observation error statistics, in order to investigate if a larger improvement is obtained in the microwave radiances assimilation.

4.3 Impact on forecasts

A first evaluation of the SEVIRI LST assimilation on the forecast quality has been considered according to surface parameters using surface stations (synoptic + RADOME) observations represented in Figure 13. The number of observation remains the same all along the day.

Figure 13

Various meteorological parameters have been evaluated, and we focus here on 2 m temperature and relative humidity, that show small but significant impacts especially by nighttime. Table 2 gives the differences of RMSE for the 2 m temperature and relative humidity parameters according to forecast range. Positive values indicate an improvement of EXP-ARO with respect to REF-ARO.

Table 2 shows small but significant impacts at different forecast ranges. First, for 2 m temperature, a very small increase in RMSE at the analysis time 0 UTC with EXP-ARO is noticed. This fact can be explained by the assimilation of SEVIRI LST that increases the distance between the analysis of EXP-ARO and the 2 m temperature observations, compared to REF-ARO that assimilates only 2 m temperature for the soil temperature analysis. On the other hand, there is a slight and significant improvement in 2 m temperature forecast at nighttime forecast ranges (24 h and 48 h). This behaviour can be explained by a better analysis of surface temperature thanks to the SEVIRI LST assimilation leading to more realistic surface temperature forecasts.

For 2 m relative humidity, significant positive impacts take place during nighttime forecast ranges (18 h to 30 h and 48 h). These improvements of the nighttime forecast range have also been found for the 2 m temperature parameter. This shows that SEVIRI LST assimilation at 00 UTC and 03 UTC assimilation times has given a consistent impact on the nighttime forecasts for both temperature and relative humidity parameters at 2 m.

To further assess the impact of SEVIRI LST assimilation, temperature and relative humidity forecasts in the atmosphere are evaluated at various heights using radiosonde observations at 1000, 850, 700, 500 and 400 hPa (Table 3).

In terms of temperature, Table 3 shows small but significant impact of SEVIRI LST assimilation on the forecast up to 700 hPa level with a maximum amplitude of 0.03 K at 1000 hPa level. The significant impacts on forecast quality with EXP-ARO remain for the first 24 h range. After 24 h, the impact turns to be neutral to slightly negative.

In terms of relative humidity, Table 3 shows an improvement in forecast at 0 UTC that is maximum between 850 and 500 hPa, reaching 4 % of RMSE difference at 700 hPa, which is consistent with the Figure 12 showing smaller RMSE for water absorption channel 5 (sensitive to 700 hPa). For forecast ranges from 12 h to 48 h, significant impacts are mainly observed at 850 and 700 hPa with positive impact at 850 hPa at 12, 36 and 48 h, and a slight degradation at 24 h and at 1000 hPa at 12, 24 and 36 h with a maximum amplitude around 0.7% on both sides. These results are consistent with the temperature forecast evaluation. The consideration of these two parameters for verification is based on the operational verification system available at Météo-France which considers relative humidity close to the surface which is indeed correlated to temperature.

The above results have shown small but significant impacts mainly on surface and low levels temperature and relative humidity forecasts. We remind that during this experiment, as it represents the first LST assimilation work with the AROME model, we have chosen to assimilate the nighttime SEVIRI LST only and also to consider equal assimilation weights for SEVIRI LST and 2 m temperature for soil analysis. The different results will be summarized and these hypothesis will be further discussed in the next section.