Atmospheric Research 82 (2006) 83 – 101 www.elsevier.com/locate/atmos
Model predicted low-level cloud parameters Part II: Comparison with satellite remote sensing observations during the BALTEX Bridge Campaigns Marc Schröder a,⁎, Nicole P.M. van Lipzig b , Felix Ament c , Jean-Pierre Chaboureau d , Susanne Crewell b , Jürgen Fischer a , Volker Matthias e , Erik van Meijgaard f , Andi Walther a , Ulrika Willén g a
Institut für Weltraumwissenschaften, Freie Universität Berlin, Carl-Heinrich-Becker-Weg 6-10, 12165 Berlin, Germany b Meteorologisches Institut, Universität München, München, Germany c Meteorologisches Institut, Universität Bonn, Bonn, Germany d Laboratoire d'Aerologie, OMP, Toulouse, France e Institut für Küstenforschung, GKSS, Geesthacht, Germany f Royal Netherlands Meteorological Institute, KNMI, DeBilt, The Netherlands g Rossby Center, SMHI, Norrköping, Sweden Received 27 January 2005; accepted 12 December 2005
Abstract A pressing task in numerical weather prediction and climate modelling is the evaluation of modelled cloud fields. Recent progress in spatial and temporal resolution of satellite remote sensing increases the potential of such evaluation efforts. This paper presents new methodologies to compare satellite remote sensing observations of clouds and output of atmospheric models and demonstrates their usefulness for evaluation. The comparison is carried out for two MODerate resolution Imaging Spectrometer (MODIS) scenes from the BALTEX Bridge Campaigns. Both scenes are characterised by low-level clouds with a substantial amount of liquid water. Cloud cover and cloud optical thickness of five different models, LM, Méso-NH, MM5 (non-hydrostatic models), RACMO2, and RCA (regional climate models) as well as corresponding retrievals from high resolution remote sensing observations of MODIS onboard the Terra satellite form the basis of a statistical analysis to compare the data sets. With the newly introduced patchiness parameters it is possible to separate differences between the two scenes on the one hand and between the models and the satellite on the other hand. We further introduce a new approach to spatially aggregate cloud optical thickness. Generally the models overestimate cloud optical thickness which can in part be ascribed to the lack of subgrid-scale variability. However, LM underestimates the frequency of occurrence of cloud optical thickness at values around 25. Furthermore, we compare the standard operational output of the non-hydrostatic models to simulations of the same models including parameterised shallow convection. However, clear improvements in the representation of low-level clouds are not found for these models. A change of the coefficients for autoconversion in RCA shows that LWP and precipitation strongly depend on this parameter. Refined vertical resolution, implemented in RACMO2, leads to a better agreement between model and satellite but still leaves room for further
⁎ Corresponding author. Tel.: +49 30 83852751; fax: +49 30 83856664. E-mail address:
[email protected] (M. Schröder). 0169-8095/$ - see front matter © 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.atmosres.2005.12.005
84
M. Schröder et al. / Atmospheric Research 82 (2006) 83–101
improvements. In general, this study reveals deficiencies of the models in representing low-level clouds, in particular for a stratiform cloud. © 2006 Elsevier B.V. All rights reserved. Keywords: Patchiness; Non-hydrostatic models; Climate models; Satellite remote sensing
1. Introduction Numerical weather prediction and climate models are essential tools for understanding the hydrological, radiative and energy budgets of the Earth. Hence, the development and examination of evaluation strategies is of great importance in order to assess the reliability of the models and to identify limitations. In the past, clouds were not the main focus in model evaluation. With the development of new measuring techniques available for studies on clouds and precipitation, this is changing (e.g. Yu et al., 1996; Barros and Bindlish, 1999; Klein and Jacob, 1999; Hollars et al., 2004; Hennemuth et al., 2003). Most frequently, averages, correlations and histograms of parameters of interest as well as their differences are utilised for evaluation purposes (e.g. Hollars et al., 2004). Yu et al. (1996) concentrated on the subgrid model variability to identify best cloud overlap within a model column. Klein and Jacob (1999) focused on the positioning of midlatitude baroclinic systems while Barros and Bindlish (1999) approached the evaluation with the help of texture analysis. Ryan et al. (2000) presented an evaluation of models with different resolutions (ranging from 5 to 300 km), e.g. concluding that the accuracy of regional climate models depends on large scale forcing. The evaluation techniques can partly be utilised to classify cloud systems. For example, the texture parameters introduced by Barros and Bindlish (1999) provide a measure for changes in climate monitoring and forecasting activities. Walther and Bennartz (2006) used texture and shape analysis to discriminate frontal from non-frontal contributions in precipitation fields. Besides evaluation of atmospheric models with ground-based remote sensing observations (Hennemuth et al., 2003; Hollars et al., 2004), simulations of such models are compared to satellite remote sensing observations (Yu et al., 1996; Barros and Bindlish, 1999; Klein and Jacob, 1999). Several parameters describing cloud properties or atmospheric conditions are available from satellite remote sensing observations. The comparison of such parameters to corresponding model parameters is in many cases a challenging task.
Firstly, appropriate parameters need to be defined for comparison. Apart from predefined priorities, like e.g. radiation, this decision should be oriented at the expected reliabilities of the satellite and model product. Secondly, aggregation is needed since the spatial resolutions of satellite and model are generally not identical. For parameters like liquid water path averaging of subgrid properties is appropriate. For cloud optical thickness the situation is more complex. If radiative aspects are in the focus of the analysis, Rossow and Schiffer (1991) propose to assign weights to cloud optical thickness within the averaging process according to its effect on the planetary albedo (adopted by Lau and Crane, 1995). Yu et al. (1996) and Klein and Jacob (1999) created pseudo-satellite observations within a General Circulation Model (LMD GCM) and the ECMWF model environment by radiative transfer models. They compared cloud optical thickness and cloud top height/pressure in terms of fractions belonging to a certain classification of cloud type. Thirdly, retrieval or parameterisation deficiencies might introduce biases which can sometimes be avoided by simple adjustments. For example, an overestimation of the fraction of highlevel clouds is found in a comparison between atmospheric models and radar observations by Hogan et al. (2001) and Willén et al. (2005). They could increase the level of agreement by introducing a threshold of minimum detectable ice water content. This study is the second part in a series of two papers in this issue. In the following the first part is referred to as van Lipzig et al. (2006). Both papers evaluate the representation of low-level clouds in five different regional atmospheric models. During the World Meteorological Organization (WMO) cloud modelling workshop two days from the Baltic Sea Experiment (BALTEX) Bridge Campaigns (BBC) have been selected as case study examples for low-level clouds. Five different models are utilised to evaluate the representation of such clouds in the two BBC cases: LM, Méso-NH, MM5 (non-hydrostatic models with a spatial resolution of around 3 km), RACMO2, and RCA (regional climate models with a spatial resolution of 19 km). A grid spacing of 3 km is expected to be insufficient to resolve shallow convection. Therefore,
M. Schröder et al. / Atmospheric Research 82 (2006) 83–101
we investigate possible differences in the representation of such clouds between model operations with and without parameterised shallow convection in the nonhydrostatic models. In van Lipzig et al. (2006) an overview of the synoptic situation during the two BBC cases is given and model related parameterisations are introduced. They give an overview of the various purposes for which the models have been designed at the different weather services or research centers and argue that the evaluation of the models in the configuration of their operational application is of relevance. In this paper, a few aspects are repeated or extended, if relevant for the discussion. This part analyses the spatial characteristics of the same regional atmospheric models and compares the results to observations of MODIS onboard the Terra satellite. The major goal of the paper is to examine the representation of low-level clouds in atmospheric climate models. For this purpose, 2D spatial fields of cloud cover, cloud optical thickness, effective cloud droplet radius, cloud top pressure, and precipitation are utilised. Section 2 introduces the MODIS observations and retrieval schemes, the atmospheric models, and the BALTEX Bridge Campaigns. This is followed by a presentation of the methodology of the comparison of cloud cover and cloud optical thickness. The results are presented in Section 4. In Section 5 the effect of including parameterised shallow convection, possible improvements achieved by refining the vertical resolution, and the comparison of the full domain of RACMO2 and RCA to MODIS observations are discussed. Finally, conclusions are presented. 2. Data sources 2.1. Remote sensing observations Remote sensing observations utilised in this study were carried out with the MODerate resolution Imaging Spectrometer (MODIS) onboard the U.S. Terra satellite. Terra is a polar-orbiting satellite that achieves global coverage every 2–3 days. It flies over central Europe 1– 2 times a day with overpass times around 10:30 UTC. MODIS takes 3D measurements of radiances: two spatial dimensions (along track through the propagation of the satellite and across track with 1354 spatial pixels, which corresponds to a swath of ∼2330 km) and one spectral dimension (36 channels, spanning the visible, near infrared and thermal wavelength range). Several operational algorithms for the retrieval of various atmospheric and cloud products exist which are
85
accessible via internet (Distributed Active Archive Center, DAAC: http://daac.gsfc.nasa.gov, funded by NASA). Among the products is the MOD05 product (King et al., 1997) which provides cloud optical thickness (τ), effective particle radius (reff), liquid water path (LWP), cloud top pressure, and thermodynamic phase. Additionally, MOD05 data includes a cloud mask that has been developed by Ackerman et al. (1997). Clouds are generally characterised by higher reflectances and lower brightness temperature differences than the underlying surface of the Earth. Therefore, the algorithm relies on a set of radiance differences and ratios as well as absolute values. Uncertainties in discriminating overcast from clear sky may arise from thin cirrus and low stratus at night. Due to insufficient contrast small cumulus and cloud edges may also be problematic (Ackerman et al., 1997). The cloud mask shows agreement with satellite and ground-based lidar observations, in particular during day and in snow-free regions, but also identifies the above criterions as uncertainty sources (King et al., 2003; Mahesh et al., 2004). Cloud phase information is determined after Menzel et al. (2002) and utilises differences in the absorption characteristics of liquid and ice water. The retrieval of cloud top pressure is also described in Menzel et al. (2002) and is based on observations at four channels in the CO2 absorption band and a window channel in the infrared. The so-called CO2 slicing method is applied and utilises the correlation between the absorption intensity, i.e. the ratio of a CO2-absorption affected channel to a window channel, and cloud top pressure. The absorption of CO2 varies strongly over the four different channels and allows correction of semitransparent clouds. When the difference between clear sky and observed radiance is within the instrument noise, the 11 μm channel is used only to retrieve cloud top pressure, assuming black body radiation. The retrieval of τ and reff relies on observations at four visible and near-infrared channels. At non-absorbing channels the reflected intensity is a function of cloud optical thickness. The algorithm separates between τ of liquid water and ice clouds, utilising the cloud phase information. The relation between radiance and τ is non-linear, and for increasing τ the radiance asymptotically approaches independency of τ. Therefore, the range of τ is limited to τ ≤ 100, and a flag is assigned, if τ > 100. The channels in the near infrared are affected by absorption of liquid water and allow the retrieval of reff. The inverse problem is solved by look-up tables. The tables have been composed by the output of radiative transfer simulations which have been performed with a
86
M. Schröder et al. / Atmospheric Research 82 (2006) 83–101
wide variety of τ, reff, ice crystal information, and viewing geometry as input. The algorithm is valid for single layer and plane-parallel geometry. While deviations from these assumptions have minor impacts on τ in the majority of cases, it may strongly affect products relying on absorption, i.e. reff. The satellite measures reff of the cloud top and no information is available from layers below. The parameters are flagged, if saturation occurs or if suspiciously high reflectivities of underlying targets are observed. Cloud top pressure and cloud phase have a spatial resolution of 5 km while cloud cover, τ, LWP, and reff have a spatial resolution of 1 km. 2.2. Regional atmospheric models Output from five models is compared to MODIS observations in this study: the Lokal-Model (LM) from the Deutscher Wetterdienst, the MESOscale NonHydrostatic model (Méso-NH) developed by the Centre National de la Recherche Scientifique and Météo France, the Mesoscale Model MM5 from the Pennsylvania State University and the National Center of Atmospheric Research operated by GKSS, the Regional Atmospheric Climate Model version 2 (RACMO2) developed by Royal Netherlands Meteorological Institute (KNMI), and the Rossby Center Atmospheric Model (RCA). LM, Méso-NH, and MM5 all are nonhydrostatic models while RACMO2 and RCA are hydrostatic (limited-area) models used for the purpose of regional climate prediction. The models are introduced in detail in van Lipzig et al. (2006), including a discussion of their scientific purposes. Here, we will
describe some details of the cloud schemes and the estimation of τ and reff relevant to this paper. In LM, Méso-NH, and MM5, an all-or-nothing scheme is used to determine cloud cover, assuming that clouds are resolved by the model grid. Additionally, LM utilises a diagnostic subgrid cloud scheme which contributes only to radiation and not to the hydrological cycle. In the climate models a subgrid-scale cloud scheme is used allowing for fractional layer cloud amount, prognostic in RACMO2 and diagnostic in RCA. In all models horizontal fluxes are neglected, so that turbulence is treated as a one-dimensional process. In MM5 τ is calculated after Stephens (1978) while in Méso-NH the determination of τ follows Fouquart (1987) and Bonnel et al. (1983). In LM, RACMO2, and RCA the cloud optical thickness τj for each layer j is calculated from the liquid water content accounting for fractional cloud cover. Model values of cloud optical thickness larger than the maximum cloud optical thickness from MODIS are set to 100 and form a class of grid cells with τ > 100. RCA and LM consider a variable reff: In RCA it is function of temperature and in LM it is discretised by a set of eight cloud types, thereby affecting τ (Ritter and Geleyn, 1992). In RACMO2, reff is assigned two distinct values, 10 μm over land and 13 μm over ocean/sea. In the remaining models, a variable reff is not explicitly considered. In addition to the standard (control; CTL) simulations, the non-hydrostatic models performed simulations with an implemented shallow convection scheme (SHC), see van Lipzig et al. (2006). The domain size of each model (together with resolution) is given in Fig. 1. The models were initiated at 12 UTC and integrated
Fig. 1. Illustration of the different domain sizes of the non-hydrostatic and climate models. The MM5 domain lies completely within the domain of LM. The resolutions of the models are 2.8 (LM), 2.5 (Méso-NH), 3.0 (MM5), and 19 km (RACMO2, RCA).
M. Schröder et al. / Atmospheric Research 82 (2006) 83–101
over a period of 36 h. 2D spatial fields are recorded every full hour, beginning 12 h after initialisation. The study carries out one of the first tests with the new version of the LM. (The operational version of LM uses Δx = 7 km.) 2.3. The BALTEX Bridge Campaigns The BALTEX Bridge Campaigns took place around the Cabauw Experimental Site for Atmospheric Research (CESAR), the Netherlands, in August and
87
September 2001 (BBC) and in May 2003 (BBC2). BBC was a joint venture between the European CLIWA-Net project (Crewell et al., 2002) and the German 4DClouds project. BBC2 was a cooperation of several institutions across Europe with major contributions from the Royal Netherlands Meteorological Institute (KNMI). During both experiments coordinated observations of clouds by various ground-based and airborne instruments were carried out. An overview of BBC including introductions to utilised instruments and first results is given by Crewell et al. (2004).
Fig. 2. Cloud cover (a, b) and cloud optical thickness (c–f) retrieved from MODIS data measured at 10:45 UTC on D1 (left panels) and at 10:05 UTC on D2 (right panels). (a–d) Are aggregated to 3 km and (e, f) to 19 km resolution, utilising Eq. (3). Cabauw, 4.93°E and 51.96°N, is marked in all panels. The black areas in the center of some regions of large cloud optical thickness are due to saturation of the sensor (e.g., (c) in the center of the top part of the image).
88
M. Schröder et al. / Atmospheric Research 82 (2006) 83–101
The comparison of MODIS and model observations is carried out for two scenes which had been taken at 10:45 UTC on 23 September 2001 (D1) and at 10:05 UTC on 21 May 2003 (D2). Consequently, each model provides 2D cloud cover at 11:00 (RCA at 10:45 UTC) and 10:00 UTC for D1 and D2, respectively. Although the time difference of 15 min can cause discrepancies for individual clouds, we do not expect that the statistics of the cloud fields have changed significantly. Details on the synoptic situation of these days are provided in van Lipzig et al. (this issue). In Fig. 2 the cloud cover (a, b) and the cloud optical thickness (c–f) observed by MODIS on D1 (left panels) and D2 (right panels) are shown. Cloud optical thickness in Fig. 2c–e and f have been aggregated onto a grid with 3 km and 19 km resolution, respectively. The D1 cloud cover is dominated by a large, relatively homogeneous liquid water cloud deck over the Netherlands (with Cabauw close to the edge of the cloud deck) and two relatively large clear sky areas to the (south-)west of Cabauw and at the south-east corner of the image. The stratiform cloud moves slowly to the east and decreases in size until mid-afternoon. In the afternoon broken clouds spread over a relatively large area around Cabauw. The clear sky area to the (south-)west of Cabauw is located further to the southwest and increases in size. In contrast, the D2 cloud mask and the cloud optical thickness images are characterised by many small clear sky areas and by a large number of small cloud cells with large cloud optical thicknesses, in particular over land. The brokenness increases until noon. Both images reflect the convective nature of this day. The focus of our study is on low-level clouds. To demonstrate that the D1 and D2 cases are well suited for this purpose we analysed the percentage of high- (above 6.1 km), medium- (1.98–6.1 km), and low-level (below 1.98 km) clouds on D1 and D2. Cloud top pressure is converted to cloud top height with radiosound data from Cabauw. The respective percentages (normalisation to total number of pixels) observed by MODIS are 10%, 29%, and 41% on D1 and 1%, 59%, and 31% on D2. The percentage of low-level clouds is reasonably large, and the increased percentage of medium-level clouds on D2 can be an indication of the convective nature of this day. The sum of the percentage does not give 100% because a cloud top pressure value is not assigned to clear sky pixels. Though the evaluation approach presented in the following was developed for low-level clouds it is not restricted to such cases and can be applied to any cloud system.
3. Approach All five models have a lower resolution than MODIS (except cloud top pressure and phase). In this work aggregation is carried out spatially, and not temporally. A newly developed parameter and a new aggregation approach are introduced to extend standard evaluation strategies. The first step of the comparison includes an appropriate aggregation of the satellite data to each model grid. If the model utilises a rotated grid, the geographic parameters of MODIS are rotated using the same parameters for rotation as the models do. The transformation follows the equations provided by Pearson (1990) and is applied for aggregation onto the LM, RACMO2, and RCA grids. For the remaining models, the geographic information of MODIS is aggregated onto the model grid by triangulation. Since the models and MODIS cover different areas, the minimum area of overlap among all six data sources is considered to allow a better intercomparison of the performances of the models. After aggregation of the satellite data onto the model grids we speak of grid cells for both, models and satellite. 3.1. Spatial averages related to cloud cover The cloud cover of MODIS is determined by arithmetic spatial aggregation, so that a fractional cloud cover is introduced for MODIS. In general, fractional cloud cover is utilised to calculate average characteristics, i.e. total cloud cover is determined from fractional cloud cover. As soon as an interpretation of the grid cell being either overcast or clear sky is required, a threshold of 0.5 is applied to fractional cloud cover: All grid cells with fractional cloud cover larger than 0.5 are considered as overcast and the remaining grid cells as clear sky. The chosen threshold is reasonable in the sense that it divides the range of values in equal parts. Note however, that a change of the threshold to 0.3 or 0.7 does not lead to qualitative changes of the results, apart from a decreased difference of p1 from MODIS and LM data on D1 which is slightly smaller than the uncertainty measure for p1, defined and estimated below. In the following the term cloud mask is used, if a threshold of b < 0.5 has been applied to cloud cover. In case of RACMO2 and RCA an application of the threshold to cloud cover is not reasonable and therefore parameters relying on the application of the threshold are not determined for both models. The quality of the spatial matching between the MODIS and the model cloud masks is assessed: A
M. Schröder et al. / Atmospheric Research 82 (2006) 83–101
89
Fig. 3. Example for two cloud masks: white and black grid cells indicate overcast and clear sky grid cells, respectively. The larger patchiness is found in the left panel (p1 = 11 / 150), dominated by clear sky areas (p2 = − 1 / 150), while the cloud mask in the right panel is dominated by clouds (p1 = 6 / 150 and p2 = 4 / 150).
contingency table is constructed from which a large number of skill scores can be extracted, as discussed, e.g. in Woodcock (1976). The majority of scores depends on the distribution of the trial conditions. Here, we consider the Kuipers skill score V as defined in Wilks (1995). The Kuipers skill score is also referred to as Hanssen and Kuipers discriminant and summarises the contingency table by a single value. V is independent of the trial conditions, i.e. of cloud cover, gives equal weight to both possible events (either overcast or clear sky), and was recommended by Woodcock (1976) for the evaluation of forecast accuracy. It varies between − 1 (perfect disagreement) and 1 (perfect agreement). A value of 0 would indicate a randomly chosen forecast. In biology the patchiness parameter describes the structural diversity of a habitat. Here, it is introduced in a similar manner: The degree of patchiness increases with increasing number of cloud and clear sky areas (Ncld and Nfree, respectively). A visual impression is given in Fig. 3. The cloud mask in the left panel has a larger degree of patchiness, caused mainly by a dominance of clear sky areas, than the cloud mask in the right panel which is dominated by overcast areas. The necessity to include both, overcast and clear sky areas, becomes obvious. In order to identify a single connected cloud area an eight-connected neighbour algorithm is applied: The algorithm searches for overcast grid cells in x- and y- as well as in diagonal direction of the data field to identify connected grid cells. Each connected cloud area is labelled with a unique area index. In case of clear sky areas a fourconnected neighbour algorithm is applied in order to avoid grid cells being connected across the links of overcast grid cells. The patchiness is then characterised by two parameters: The first parameter gives the overall patchiness of the cloud mask and is defined as p1 ¼ ðNcld þ Nfree Þ=n;
ð1Þ
with n being the total number of grid cells. The degree of patchiness is large, if p1 is large, and maximum patchiness is achieved, when the cloud mask appears as a chess board with each grid cell surrounded by four clear sky and four overcast grid cells. Then, all overcast grid cells compose a single cloud, and all clear sky grid cells count as individual areas, which leads to the maximum p1 = 1 / 2 + 1 / n. The second parameter assigns a qualitative attribute to p1, and is determined as the normalised difference between Ncld and Nfree: p2 ¼ ðNcld Nfree Þ=n:
ð2Þ
If p2 is positive (negative), overcast (clear sky) areas contribute most to the patchy appearance of the cloud mask. Since Nfree relies on a four-connected neighbour algorithm, p2 is biased to smaller values. Note, that p1 and p2 do not change, if identical areas are added, i.e. the domain size is duplicated. However, the aggregation to grids with different resolution in combination with the application of the threshold might lead to discrepancies. Furthermore, different geographic coordinate systems in combination with the necessity to analyse a rectangular data field can also result in differences. In order to assess the robustness of patchiness we consult the patchiness for the satellite among the non-hydrostatic models (Tables 1 and 2). It can be concluded that differences in patchiness of the order of 0.001 are not significant. Note, that p1 was multiplied by 100 in Tables 1 and 2. The usefulness of the introduction of the patchiness parameters is addressed with a sensitivity study. In an exemplary study, carried out with Méso-NH, the efolding time for numerical diffusion has been set to 1500 times and 100 times the time step. We found p1 = 0.0041 and p1 = 0.0042 on D1 and D2, respectively, if the typical, small time for numerical diffusion is utilised. In case of the large time for numerical diffusion, we get p1 = 0.0124 and p1 = 0.0128 on D1 and D2, respectively.
90
M. Schröder et al. / Atmospheric Research 82 (2006) 83–101
Table 1 Comparison of total cloud cover (b), Kuipers skill score (V) and patchiness (p1, p2) extracted from the D1 data Parameter
LM
Méso-NH
MM5
b V p1 ⁎ 100 p2 ⁎ 100
0.84/0.61 (0.79) 0.16/0.54 0.37/0.28 (0.50) − 0.17/−0.07 (−0.03)
0.83/0.86 (0.78) 0.16/0.15 0.41/0.39 (0.41) − 0.11/− 0.24 (−0.11)
0.92/0.80 (0.79) 0.20/0.12 0.56/0.44 (0.42) − 0.36/− 0.19 (−0.16)
The total cloud cover from RACMO2 and RCA is 0.62 (MODIS: 0.81) and 0.76 (MODIS: 0.81), respectively. The first values refer to the CTL- and the second to the SHC-runs. The values in brackets refer to MODIS. p1 and p2 were multiplied by 100 to provide larger values. The total number of grid cells is between 11,000 and 19,000.
While the patchiness shows a large difference between small and large times for numerical diffusion, averages and histograms of cloud optical thickness exhibit no significant differences between both integrations. For an intercomparison of models with different resolutions the interpretation of the results might be problematic. But in this paper the focus is on a comparison of model output and MODIS observations, so that the patchiness parameters are a useful extension of standard evaluation approaches and may contribute to an evaluation of existing and future turbulence schemes. 3.2. Spatial aggregation of cloud optical thickness Cloud optical thickness can be seen as a link between cloud properties and radiative transfer. It is defined as the integral over the droplet size distribution weighted by a product of extinction efficiency and particle cross section (e.g., Liou, 1992). This expression can be either applied to a single grid cell or can be integrated in the vertical. However, no definition exists to directly infer a spatial average of τ from spatially varying fields of cloud properties. Since τ is a radiative property a spatial average should be defined via radiative transfer. Therefore, the most appropriate approach to assess the effect of a spatially variable τ on radiation would be via observations and averages of radiances. Since the implementation of a radiation module that is capable of simulating radiances corresponding to the observation geometry of MODIS is very ambitious and beyond the scope of this analysis, we propose to estimate the average (or effective) cloud optical thickness (τ) from the transmission (T) according to T = exp(− τ/μ) (μ being the cosine of the sun zenith angle). The directly transmitted radiance can be estimated, if the incident radiance is known and multiplied with T. The transmission T provides the
link to radiative transfer, and we aggregate τ by (with i being a spatial grid cell index and N being the number of grid cells): ! X s ¼ l ln 1=N expðsi =li Þ : ð3Þ i
In this way, the radiative effect of a spatially variable cloud optical thickness is approximated to first order. Eq. (3) is used to aggregate MODIS cloud optical thickness onto the model grid and yields a horizontally averaged effective cloud optical thickness on the basis of independent spatial grid cells. Furthermore, it offers a general approach to allow a comparison of high resolution cloud optical thickness retrieved from satellite observations to corresponding output from atmospheric models with lower resolutions. LM, RACMO2, and RCA have fractional cloud cover for each of their layers. Therefore, we have to assume a cloud overlap function in order to derive an effective cloud optical thickness valid for the entire vertical column. The fractional cloud cover indirectly requires a spatial average of overlapping clear sky and overcast parts of the layers. On the one hand we want to stay as close as possible to the cloud optical thickness that enters the radiation code of the model. On the other hand we want to be as close as possible to the cloud optical thickness as retrieved by the satellite. For this purpose we integrate as follows: ! Y scloud ¼ l ln ð1 bj þ bj expðsj =lÞÞ ; ð4Þ j
with j being a layer index. bj has been normalised to the columnar cloud cover which has been determined applying maximum-random overlap, and therefore τcloud is the cloud optical thickness and not the optical thickness of the grid box. 1 − bj incorporates the impact of the clear sky part of the layer on transmission (with the clear sky transmission being one) while the second term within the product gives the transmission of the Table 2 As Table 1 but on D2 Parameter
LM
Méso-NH
MM5
b V p1 ⁎ 100 p2 ⁎ 100
0.92/0.87 (0.92) 0.19/0.16 0.44/0.49 (0.66) − 0.33/− 0.32 (−0.48)
0.80/0.79 (0.92) 0.14/0.09 0.57/0.56 (0.62) − 0.53/− 0.41 (−0.54)
0.97/0.86 (0.92) 0.22/0.25 0.38/0.19 (0.71) − 0.30/− 0.12 (−0.62)
The total cloud cover from RACMO2 and RCA is 0.71 (MODIS: 0.92) and 0.76 (MODIS: 0.92), respectively.
M. Schröder et al. / Atmospheric Research 82 (2006) 83–101
overcast part of the layer. Therefore, the summation in Eq. (4) gives a spatial average transmission of a single layer. To get the transmission of the whole column, the transmissions of the layers are multiplied with each other consecutively. Finally, the transmission is inverted to get the average cloud optical thickness of the cloudy column. τcloud is subsequently converted into an effective cloud optical thickness of the grid box by applying Eq. (4) with j = 1 and replacing bj with the columnar cloud cover. To comply with this, MODIS cloud optical thickness is aggregated to the model grid scale by applying Eq. (3) to all pixels in a model grid cell with τi = 0 for clear sky pixels. After aggregation onto the model grid, the domain average cloud optical thickness (τ) is determined over all satellite and model grid cells with bi > 0, utilising Eq. (3). Scene and domain averages of satellite and model are determined independently of each other, i.e. no overlap of clouds is required between satellite and model. In this way, misinterpretations due to missing cloud areas in the model output are avoided. 4. Results 4.1. Cloud top pressure, cloud optical thickness, and cloud cover It has been shown by Hogan et al. (2001) and Willén et al. (2005) that the ECMWF model, LM, RACMO2, and RCA overestimate the occurrence of high-level clouds, if compared to cloud radar observations. Similar findings have been presented by Van Meijgaard et al. (2001) for a comparison to satellite observations. Wyser and Jones (2005) argued that this might be caused by clouds with small cloud optical thicknesses, invisible to satellite remote sensing. We exemplary compare the percentages of high-, medium-, and low-level clouds retrieved from LM cloud top height (20%, 40%, and 34% on D1 and 32%, 41%, and 18% on D2) to the satellite values (10%, 29%, and 41% on D1 and 1%, 59%, and 31% on D2). LM overestimates the occurrence of high-level clouds in both cases what might be caused by a disability of the satellite to detect high thin clouds. Dessler and Yang (2003) analysed MODIS’ sensitivity to tropical cirrus clouds with low values of cloud optical thickness. They found that 30% of cirrus clouds with τ < 0.05 are undetected. With an expression given by Heymsfield et al. (2003) τ can be converted to ice water path (IWP). For τ = 0.05 one gets IWP = 0.7 g m− 2. If we exclude all grid cells with IWP < 0.7 g m− 2, the fractions of high-level clouds reduce to 8% and 10% in LM on D1 and D2, respectively. The percentages of
91
the other classes are almost unaffected. The application of the IWP threshold improves the agreement between satellite and model but on D2 an overestimation of highlevel clouds remains. The thin ice clouds are located in the north and north-west as well as in the west and south-west of the domain on D1 and D2 (mainly over sea), respectively. Note, that the spatial resolution of MODIS cloud top pressure product is 5 km. The qualitative results are independent of the aggregation process (either arithmetically or maximum of the cloud top height values of MODIS within each model grid cell). Images of cloud optical thickness of the five models are presented in Figs. 4 (D1) and 5 (D2), and corresponding figures of cloud cover can be found in van Lipzig et al. (2006). The regional climate models show smaller cloud optical thicknesses than the nonhydrostatic models which is a consequence of the larger number of grid cells with fractional cloud cover in combination with the utilisation of Eq. (4). On D1 the stratiform cloud over land is present in RCA and with lower homogeneity in MM5. None of the models was capable to simulate the clear sky area to the (south-)west of Cabauw, but LM and RACMO2 show small cloud optical thicknesses in that region (arithmetic averages of 2.4 and 3.8, respectively, for longitudes between 3° and 4.5° as well as latitudes between 51.4° and 52.1°). On D2 the MODIS cloud optical thickness images show cloud structure with a larger variety of cell sizes than the models, when the cell size is defined by a small cloud optical thickness threshold. In particular, small cell sizes are found with higher frequencies in MODIS than in the models. This also holds for LM and Méso-NH which exhibit the largest degree of structure among the models. However, LM and Méso-NH likewise exhibit convective structure on D1. Besides small areas at the west and north of the verification domain the satellite observes low values of cloud optical thickness over sea on D2 (arithmetic average of ∼ 2, cloud top height of ∼ 2 km). This is best resembled by RACMO2 and largely overestimated by Méso-NH and RCA while LM and MM5 show large parts over the sea as clear sky. The full scene of MODIS shows that the area with low values of cloud optical thickness extends over a larger area than is depicted in Fig. 2b, and in the lateral forcing data of LM and MM5 the clear sky area over sea covers approximately the same area as MODIS. An analysis of relative humidity and temperature in the clear sky areas reveals that RH values in LM and MM5 are only a few percent below saturation. A small increase in humidity or a small decrease in temperature would lead to cloud formation.
92
M. Schröder et al. / Atmospheric Research 82 (2006) 83–101
Fig. 4. Cloud optical thickness of LM (a), Méso-NH (b), MM5 (c), RACMO2 (d), and RCA (e) on D1. Cabauw, 4.93°E and 51.96°N, is marked with a black arrow. RACMO2 and RCA show a section of their full domain and are limited to lower maximum values of cloud optical thickness. Note that cloud optical thickness is plotted on a non-linear scale.
The number of grid cells with fractional cloud cover different from zero and one (precisely: 0.1 < b < 0.9) is around 25%, if LM is considered, and between 12% and 21% in case of cloud cover from MODIS (Δx = 2.8 km). The regional climate models yield values around 77% while the satellite observes only 34% (Δx = 19 km). Obviously, the probability of a grid cell being either clear sky or overcast decreases with increasing grid cell size. If histograms of fractional cloud cover (not shown) of MODIS are compared to the model histograms (LM, RACMO2, and RCA), these models underestimate the frequency of occurrence of clear sky grid cells and their
frequency distributions show larger skewness than the satellite observations. This was also found by Willén et al. (2005). We examined model predicted time series of cloud optical thickness and cloud amount in order to see whether the comparison is sensitive to a shift in time of the model fields relative to satellite observations. With the exception of a minor improvement for MM5 on D2, the sensitivity is found negligible. The analyses presented so far shows that the clouds of interest, i.e. the stratiform cloud deck on D1 and the cumulus on D2, are located exclusively over land while
M. Schröder et al. / Atmospheric Research 82 (2006) 83–101
93
Fig. 5. As Fig. 4 but on D2.
the clouds over sea include the majority of ice clouds and are partly optically thin. We therefore restrict our analyses to clouds over land surfaces in the following. 4.2. Comparison of spatial average characteristics of cloud cover Tables 1 and 2 provide total cloud cover, Kuipers skill score, and patchiness for the models and MODIS on D1 and D2, respectively. In this section we focus on the first values given for the models (CTL-runs). The total cloud cover is generally underestimated by the regional climate models and overestimated by the nonhydrostatic models, except for an underestimation in case of Méso-NH on D2. The overestimation of the non-
hydrostatic models on D1 is related to the clear sky area (south-)west of Cabauw which is not present in the models. The Kuipers skill score is relatively small in all cases, with largest values found for MM5. The small skills can be expected from the patchy appearance of cloud cover and cloud optical thickness on D1 and D2. The spatial variability on small scales reduces the probability for coincidences: e.g. p1 of Méso-NH on D2 is characterised by largest patchiness, in close agreement to the satellite, but has a small Kuipers skill score. The patchiness p1 is smaller for all non-hydrostatic models than for MODIS on D1 and D2 (except for an overestimation in case of MM5 on D1), and satellite and model cloud masks are dominated by clear sky areas, in evidence by a negative p2. Largest differences in p1 and
94
M. Schröder et al. / Atmospheric Research 82 (2006) 83–101
p2 are found for MM5 on D2 what is caused by a homogeneous cloud deck that covers large parts of the domain. The results related to patchiness are consistent with a histogram analysis of clear sky areas (in number of grid cells, A). With the exception of Méso-NH and MM5 on D1, the models underestimate the frequency of occurrence of small clear sky areas with A < 5–10. This finding is consistent with conclusions given in Bryan et al. (2003) that due to numerical diffusion structures at scales smaller than about six times the grid spacing do not represent physical processes. The dominant cloud structure on D1 is a relatively homogeneous cloud deck covering large parts of the land surface while D2 is characterised by convection which is reflected in a large amount of small clear sky areas. We expect p1 to indicate this difference between D1 and D2, i.e. D2 should be characterised by a larger p1 than D1. MODIS, LM, and Méso-NH observe larger values of p1 on D2 than on D1 but not MM5. This
points to the conclusion that LM and Méso-NH better represent the convective cloud structure on D2 than MM5, in agreement with the finding of van Lipzig et al. (2006). 4.3. Comparison of histograms and spatial averages of cloud optical thickness Histograms of cloud optical thickness are plotted in Figs. 6 and 7 for D1 and D2, respectively. The fraction of τ > 100 is at most 2% (Méso-NH on D2) but generally much smaller. In the majority of cases the models overestimate the frequency of occurrence of high values of τ. Remarkable is the maximum in frequency of occurrence at τ = 35 in case of MM5 on D2 which is in contrast to the almost exponential decrease of frequency of occurrence with increasing τ in all other cases (MODIS and non-hydrostatic models). The maximum in frequency of occurrence for MM5 has its origin in the
Fig. 6. Histograms of cloud optical thickness of LM (a), Méso-NH (b), MM5 (c), RACMO2 (d), and RCA (e) in comparison to MODIS (D1). In (a–c) the frequency of occurrence is shown for the CTL- and SHC-runs. Note the logarithmic scale of the frequency of occurrence.
M. Schröder et al. / Atmospheric Research 82 (2006) 83–101
95
Fig. 7. As Fig. 6 but on D2.
homogeneous and relatively large cloud deck, with values of τ around 35. The cloud deck is also present at Cabauw and remains there for another 2 h. This is consistent with the overestimation of the lifetime of clouds by MM5, found in van Lipzig et al. (2006). Because MODIS observations and the full domain simulations of MM5 show a good agreement in spatial structure, large scale advection cannot explain the above result. MM5 overestimate the humidity in the planetary boundary layer, and we suppose that the humidity of the land surface is overestimated in MM5. However, we cannot verify this assumption since the appropriate observations are lacking. LM underestimates the frequency of occurrence of τ between 15 and 45 on D1 and D2. To analyse this underestimation of τ in more detail we compared the
cloud top height of LM and MODIS. We focused on cloud top heights of liquid water clouds with LWP larger than a certain minimum. (It turned out that the results do not depend strongly on the chosen threshold.) On D1 we found an average difference between LM and MODIS of almost 600 m, on D2 an average difference of >300 m. Note that for cloud top heights below 3000 m the retrieval algorithm for MODIS is likely to use single channel information only, with uncertainties that can be larger than the above differences. The histograms of cloud optical thickness are consistent with an underestimation of the inversion strength found for LM on both D1 and D2 and with a correlation between the spatial variability of updrafts and large LWP (van Lipzig et al., 2006). It can be deduced from Figs. 6 and 7 that all models overestimate arithmetic averages of cloud optical
96
M. Schröder et al. / Atmospheric Research 82 (2006) 83–101
thickness, besides underestimations in case of RACMO2 on D1 and LM on D1 and D2. But as stated in Section 2.3 we interpret cloud optical thickness as a parameter relevant for radiative transfer and therefore determine spatial averages of cloud optical thickness with Eq. (3) (τ). In Table 3 τ is provided for MODIS and models on D1 and D2. Here, we consider the values related to the CTL-runs. We find that under- and overestimations of τ are not necessarily accompanied by under- and overestimations of arithmetic averages of cloud optical thickness. The higher (lower) frequencies of occurrence of low values of cloud optical thickness explain the differences in case of LM (Méso-NH). All models, except Méso-NH on D2, overestimate τ. The analyses of Yu et al. (1996) and Klein and Jacob (1999) also show that for middle and low clouds optically thick clouds are overestimated in the simulations of the GCM. Here, the overestimations are in part due to the approach applied for aggregation. If we hypothetically assume overcast layers in the models and overcast pixels in the satellite, then the subgrid-scale variability found in MODIS observations with Δx = 1 km always leads to larger transmissions than model simulations of the same cloud with lower spatial resolution: T(〈τ〉) ≤ 〈T(τ)〉 and therefore τmodel ≥ τMODIS. This subgrid-scale variability bias has been analysed for albedo by Cahalan et al. (1994), and they propose to implement an inhomogeneity factor of 0.7 to account for subgridscale variability. This factor has been utilised to determine cloud optical thickness in RACMO2. The factor was determined for marine stratocumulus and might depend on the type of the cloud. We also find large fractions of cloud cover with 0.1< b< 0.9, so that the bias might not be present: The fractional cloud cover can be interpreted as spatial cloud variability. In case of LM, MM5, Méso-NH and in particular in case of RCA the bias may explain the results while in case of RACMO2 the value of the inhomogeneity factor might not be appropriate. Cloud cover of MODIS is aggregated onto the model grid which introduces fractional cloud cover. This indicates cloud structures that cannot be resolved by Méso-NH and MM5 since these models utilise an all-or-
nothing scheme to determine cloud cover. If grid cells with bi = 1 in MODIS data are considered only, the difference in τ between MODIS on the hand and MésoNH and MM5 on the other hand is reduced, but an overestimation of τ by the models remains. The models need to parameterise reff to determine cloud optical thickness from LWP, and for the cloud fields of interest all models but RCA define either a constant reff or do not explicitly parameterise τ with reff. To illustrate the variability of reff even under relatively homogeneous conditions, we analysed the spatial variability of reff in MODIS observations of the stratiform cloud deck on D1. The minimum, maximum, and average values are 2, 30, and 11 μm, respectively, with a standard deviation of almost 4 μm (On D2 larger variations of reff are found.). Though we cannot conclude that a constant reff explains any biases, it demonstrates the need to introduce a variable reff or for a refinement of existing discretisation schemes. We utilise rain radar observations taken at 11:00 UTC on D1 and at 10:00 UTC on D2 (images not shown) to investigate, if underestimations (overestimations) of τ coincide with overestimations (underestimations) of precipitation. We found an overestimation of precipitation for LM on D1 and an underestimation of precipitation for MM5 on D2 which may partly explain the differences in τ. In all other cases, differences in precipitation cannot consult to explain differences in τ. Like in van Lipzig et al. (2006) we studied the effect on the representation of low-level clouds in RCA of altering the effective radius threshold that controls the onset of autoconversion. We decreased this parameter from 11 μm to 6 μm. It leaves the precipitation in RCA on D1 almost unchanged, in agreement with the observations, but leads to an increased precipitation on D2. The latter effect brings the model output closer to the radar observations. Furthermore, it leads to a decrease of spatial averages of LWP of up to a factor of three so that the averages from RCA are similar to the averages from RACMO2. A comparison of cloud optical thickness between MODIS and models would have been appropriately complimented by a corresponding analysis of LWP. However, the comparison of LWP from a high resolution
Table 3 Comparison of average cloud optical thickness (τ) on D1 and D2 Case
LM
Méso-NH
MM5
RACMO2
RCA
D1 D2
2.1/1.7 (1.4) 3.0/2.7 (2.1)
2.2/2.4 (1.5) 2.0/2.0 (2.6)
1.8/1.5 (1.3) 3.8/2.7 (1.8)
1.9 (1.0) 2.4 (1.9)
3.4 (1.0) 2.3 (1.9)
The first values refer to the CTL- and the second to the SHC-runs. The values in brackets refer to MODIS.
M. Schröder et al. / Atmospheric Research 82 (2006) 83–101
time series observation at one location and a medium resolution observation in the horizontal at one time is a challenging task (Jolivet and Feijt, 2005) and beyond the scope of this paper. Furthermore, the retrieval of LWP from MODIS observations is less direct than the retrieval of cloud optical thickness and depends on the retrieval of reff. reff is related to cloud top layers and is not necessarily representative for the whole column. In particular, it can be larger than a microphysical average of the column, leading to overestimations of LWP (also stated in Jolivet and Feijt, 2005). 5. Discussion 5.1. Sensitivity study on shallow convection Satellite observations offer an excellent opportunity to investigate the different performances of the nonhydrostatic models with regard to an operation with their standard cloud schemes and with a parameterisation of shallow convection. Because shallow convection is not fully resolved by a grid spacing of 3 km, we carried out this sensitivity study in order to see how the presentation of low-level clouds is affected in the non-hydrostatic models. The parameters related to cloud cover and cloud optical thickness of the SHC-runs are given in Tables 1– 3. Furthermore, the histogram plots in Figs. 6 and 7 also contain data from SHC-runs. Méso-NH revealed no significant sensitivity to an activated SHC (though τ increases on D1 and V decreases on D2) what was also found in van Lipzig et al. (2006). b and τ decrease for LM and MM5 in the SHC-run. The SHC scheme enhances the likelihood of mixing of moist and heat, which can result in a more homogeneous boundary layer and reduces the strength and frequency of updrafts. From that we can expect a reduction of b and τ, whenever the SHC scheme is activated (see also LWP analysis in van Lipzig et al., 2006). In case of LM V increases significantly on D1. The increase is a consequence of the clear sky area at the (south-)west of Cabauw, which is present in the SHC-run and thereby increases the spatial agreement to MODIS. The increased mixing seems to be strong enough to dissolve the thin clouds which had been present in the CTL-run. It was found in van Lipzig et al. (2006) that LM simulates strongest north-easterly winds. These winds seem to foster the mixing and dissolvement of the clouds to the (south-)west of Cabauw. The strong decrease of τ for MM5 on D2 is evident in the shift of the maximum in frequency of occurrence to a smaller cloud optical thickness. In all other models
97
frequencies of occurrence related to the SHC-run are either similar or smaller than the frequencies related to CTL-runs at all values of cloud optical thicknesses. An improvement is found at high values of cloud optical thickness for all models. The schemes supply additional mixing and thereby may reduce the probability of overshooting clouds. The decrease of the frequency of occurrence of low values of cloud optical thickness of LM reflects the dissolved optically thin clouds which were present in the CTL-run of LM on D1. Due to an increased tendency of clouds to dissolve when SHC is activated one might expect an increase in patchiness. However, apart from p1 of LM on D2, no increase of patchiness with the implementation of SHC can be found. The SHC can introduce a tendency of clear sky areas to become connected and small overcast areas to be dissolved and thereby may reduce the patchiness. The former argument might be favoured because p2 decreases, with the exception of p2 for Méso-NH. Additionally, SHC is only activated when a convective depth threshold of 250 hPa is not exceeded. This might be the case for a few grid cells only and, in consequence, might have just little effect on patchiness. In the two considered cases, an activated SHC does not result in a clear improvement of the model results in terms of b, τ, and histograms of τi. This is consistent with the findings of Zhu et al. (2005). However, b and τ can be strongly affected by an implementation of SHC and in some instances the scheme leads to results which are closer to the observations. 5.2. Increased vertical resolution A sensitivity study on increased vertical resolution was carried out with RACMO2. We analysed, if the representation of low-level clouds in RACMO2 with 60 and 80 vertical levels (RACMO2_L60 and RACMO2_L80, respectively) improves. The vertical grid spacing of RACMO2_L80 is 60–170 m in a region of 0.5–3 km. The sensitivity study relies on an analysis of b and τ which are given for RACMO2 (40 levels), RACMO2_L60, RACMO2_L80, and MODIS in Table 4 for D1 and D2. RACMO2_L60 and RACMO2_L80 exhibit a clear improvement relative to RACMO2, except for τ on D2. The simulations with 60 and 80 levels still show differences with MODIS but these are smaller than the differences found for RACMO2 in all cases. If images of cloud cover and cloud optical thickness are analysed (not shown), RACMO2_L80 is closest to the satellite on D1: A large stratiform cloud deck covers the Netherlands. A clear sky area is also
98
M. Schröder et al. / Atmospheric Research 82 (2006) 83–101
Table 4 Comparison of total cloud cover (b) and average cloud optical thickness (τ) of RACMO2, RACMO2_L60, and RACMO2_L80 to corresponding results of MODIS Parameter RACMO2 RACMO2_L60 RACMO2_L80 MODIS b (D1) τ (D1) b (D2) τ (D2)
0.62 2.9 0.71 3.6
0.78 4.8 0.80 4.4
0.82 5.8 0.79 4.8
0.81 5.0 0.92 3.0
The upper and lower halves of the table refer to observations on D1 and D2, respectively.
present but further to the north than in the MODIS images. In conclusion, the addition of vertical levels to RACMO2 increases the quality of the representation of low-level clouds, with larger improvements on D1 than on D2. 5.3. Full domain analysis of RACMO2 and RCA The area analysed above might not be representative for the full domain of RACMO2 and RCA and therefore, we have compared the models in their full domain size to MODIS with respect to cloud cover and histograms of cloud optical thickness and LWP. MODIS
observations do not cover the full domains completely, so that only overlapping areas are considered for comparison. The spatial agreement in cloud cover images is relatively large, major features are well captured by both models but appear smoothed relative to MODIS observations (not shown). The models still underestimate b, though the correspondence to MODIS is better. In Fig. 8a,b histograms of cloud optical thickness of RACMO2, RCA, RCA AUT (change of the threshold for autoconversion), and MODIS on D1 and D2 are shown. Both models overestimate the frequency of occurrence of high values of cloud optical thickness, in particular RCA. Overestimations are also found for RACMO2 at low values of cloud optical thickness. Both models overestimate arithmetic averages of cloud optical thickness which are also found for averages determined using Eq. (3) in case of RCA. However, the relative difference between τ of MODIS and RCA is smaller for the full domain than for the verification domain. In case of RACMO2 the comparison of τ exhibits excellent agreement to MODIS on D1 and D2, again showing that the large frequencies of occurrence at low values of cloud optical thickness dominate the spatial average. Histograms of LWP are plotted in Fig. 8c,d. Maximum LWP values and corresponding frequencies
Fig. 8. Frequency of occurrence of cloud optical thickness (a, b) and LWP (c, d) from RACMO2 and RCA on D1 (left panels) and D2 (right panels). The label RCA AUT marks the histograms with a change of the threshold for autoconversion.
M. Schröder et al. / Atmospheric Research 82 (2006) 83–101
of occurrence are larger for RCA than for RACMO2. The histogram of LWP of RCA is in much closer agreement to the corresponding histogram of RACMO2, if the threshold for autoconversion is changed in RCA. Here, the major differences are larger frequencies of occurrence around 0.5 kg m− 2. 6. Conclusions This paper has provided methods by which satellite data can be used for the evaluation of the representation of clouds in atmospheric models. It is the second part of a series of two papers (first part by van Lipzig et al., 2006), and utilises model output from LM, Méso-NH, MM5 (non-hydrostatic models, Δx = 3 km), RACMO2, and RCA (regional climate models, Δx = 19 km). It shows results of first applications to the BBC cases-two of the cases of the WMO cloud modelling workshop 2004. Satellite images are of particular use since they cover a large area at high spatial resolution, and we utilise MODIS observations during the BBC cases at 10:45 UTC on 23 September 2001 (D1) and at 10:05 UTC on 21 May 2003 (D2). For an atmospheric model it is not feasible (nor relevant) to forecast an individual cumulus cloud at the exact location and time. However, the models need to be able to describe the statistical properties of the clouds. The comparison was carried out for cloud cover, cloud optical thickness, effective radius, and cloud top pressure simulated by the atmospheric models and retrieved from MODIS observations. The newly introduced patchiness parameters are determined from cloud masks and turned out to be relevant to distinguish different types of cloud cover: The patchiness enables the differentiation between the more stratiform cloud cover on D1 and the convective cell structures observed on D2. It further reveals differences in spatial organisation of clouds among the atmospheric models relative to MODIS and may be useful to test future turbulence and cloud schemes. To allow an appropriate comparison of cloud optical thickness we developed an aggregation approach which accounts for radiative aspects and for the high spatial resolution of the MODIS observations. Furthermore, it can be applied to a wide variety of atmospheric models. For overcast layers a bias between model and satellite observations can be expected due to the subgrid-scale variability of cloud optical thickness in the MODIS observation. This may partly explain the frequently found overestimations. However, LM underestimates the frequency of occurrence of cloud optical thickness at values around 25 and overestimates the frequency of
99
occurrence at values larger than 45. This points to an overestimation of vertical velocities what inhibits the development of stratiform clouds. The short lifetime of clouds discussed in van Lipzig et al. (2006) are accompanied by an underestimation of the frequency of occurrence of small cloud areas with relatively large cloud optical thicknesses. In van Lipzig et al. (2006) the vertical velocities in LM, that are highly variable in time, are correlated with the liquid water content. It is unlikely that the large scale forcing imports the observed differences between models and MODIS. Rather, the parameterisations of turbulence and cloud processes are more likely candidates to be inaccurate to simulate lowlevel clouds. An exemplary analysis of the amount of high-level clouds exhibits overestimations of the fraction of high-level clouds in LM. An application of a threshold of IWP ≤ 0.7 g m− 2 to LM cloud cover reduces the fraction of high-level clouds largely, with closer agreement to the MODIS observation. The threshold was chosen to compensate for the disability of MODIS to detect thin cirrus clouds. The use of a (nearly) constant value of reff in most models stands in contrast to the strong variability of reff in MODIS observations and may partly explain the differences between MODIS and model cloud optical thicknesses, i.e. a variable reff has a direct effect on cloud optical thickness, and in general this effect is not canceled by the aggregation approach. Differences in precipitation between model simulations and rain radar observations are generally not consistent with differences in τ, except for LM on D1 and MM5 on D2. However, precipitation and LWP from RCA were strongly affected by changes in the threshold for autoconversion. The results related to the operational versions of the models are confronted with results of the same models with parameterised shallow convection included (nonhydrostatic models only). Due to an increase of the intensity of mixing total cloud cover and the frequency of occurrence of high values of cloud optical thickness are found reduced. In general, the shallow convection does not lead to a better agreement between models and satellite. It exhibits improvements in some parameters and for some models but not for the full set of parameters and all models. Refined vertical resolution (60 and 80 levels) was implemented in RACMO2 to identify possible improvements relative to an operation with 40 levels. It clearly leads to closer agreement with MODIS observations, with minor differences between the two higher resolution simulations. Satellite observations are well suited to complement the evaluation of atmospheric models with ground-
100
M. Schröder et al. / Atmospheric Research 82 (2006) 83–101
based observations. The presented findings are consistent and partly confirm the results given in van Lipzig et al. (2006). Uncertainties in simulating the strength of inversions in conjunction with too strong updrafts partly explain the presented results. The comparison reveals that atmospheric models, operating at horizontal scales as fine as 3 km, still have weaknesses in the representation of low-level clouds, in particular for the stratiform cloud. Systematic deficiencies in modelled cloud fields cannot be identified on the basis of two scenes only. In the near future the comparison of non-hydrostatic and climate models with satellite remote sensing should be extended to a significantly larger number of cases, i.e. a long-term study is planned in order to assess the above mentioned weaknesses. In particular, utilising data from the Meteosat Second Generation satellite, which takes full disk observations every 15 min, will permit the analysis of the diurnal cycle of cloud systems. The parameters and methods presented above allow a fast processing of data sets for evaluation purposes and therefore are applicable to significantly larger data amounts of long-term evaluation studies. Acknowledgements The authors like to thank Peter Albert from Deutscher Wetterdienst and Maximilian Reuter from the Freie Universität Berlin for providing their aggregation routines. Henk Klein Baltink from the Royal Netherlands Meteorological Institute (KNMI) is kindly acknowledged for providing rain radar data. The use of MM5, which is a Pennsylvania State University/ National Center for Atmospheric Research (NCAR) model, is gratefully acknowledged. N.P.M. van Lipzig and M. Schröder were funded by the German project QUEST under grant FI 435/12-1 and CR 111/5, respectively. References Ackerman, S., Strabala, K., Menzel, P., Frey, R., Moeller, C., Gumley, L., Baum, B., Schaaf, C., Riggs, G., 1997. Discriminating ClearSky from Cloud with MODIS: Algorithm Theoretical Basis Document (MOD35). Algorithm Theoretical Basis Document ATBD-MOD-06. NASA Goddard Space Flight Center. Barros, A.P., Bindlish, R., 1999. Using image analysis techniques for intercomparison of spatial variables: an application to satellite observations and model simulations of cloud fields. Syst. Res. Info. Sci. 8, 273–293. Bonnel, B., Fouquart, Y., Vanhoutte, J.-C., Farvalo, C., Rosset, R., 1983. Radiative properties in some African and mid-latitude stratocumulus clouds. Beitr. Phys. Atmos. 56, 409–428.
Bryan, G.H., Wyngaard, J.C., Fritsch, J.M., 2003. Resolution requirements for the simulation of deep moist convection. Mon. Weather Rev. 131, 2394–2416. Cahalan, R.F., Ridgway, W., Wiscombe, W.J., Bell, T.L., 1994. The albedo of fractal stratocumulus clouds. J. Atmos. Sci. 51, 2434–2455. Crewell, S., Drusch, M., van Meijgaard, E., van Lammeren, A., 2002. Cloud observations and modelling within the European BALTEX Cloud Liquid Water Network. Boreal Environ. Res. 7, 235–245. Crewell, S., Bloemink, H., Feijt, A., Garcia, S., Jolivet, D., Krasnov, O.A., van Lammeren, A., Löhnert, U., van Meijgaard, E., Meywerk, J., Quante, M., Pfeilsticker, K., Schmidt, S., Scholl, T., Schröder, M., Simmer, C., Trautmann, T., Venema, V., Wendisch, M., Willén, U., 2004. The BALTEX Bridge campaign: an integrated approach for a better understanding of clouds. Bull. Am. Meteorol. Soc. 85, 1565–1584. Dessler, A.E., Yang, P., 2003. The distribution of tropical thin cirrus clouds inferred from Terra MODIS data. J. Climate 16, 1241–1247. Fouquart, Y., 1987. Radiative transfer in climate modeling. In: Schelsinger, M.E. (Ed.), NATO Advanced Study Institute on Physically-Based Modeling and Simulation of Climate and Climatic Changes. Erice, Sicily, 11–23 May 1986, pp. 223–283. Hennemuth, B., Rutgersson, A., Bumke, K., Clemens, M., Omstedt, A., Jacob, D., Smedman, A.-S., 2003. Net precipitation over the Baltic Sea for one year using models and data-based methods. Tellus 55A, 352–367. Heymsfield, A.J., Matrosov, S., Baum, B., 2003. Ice water path-optical depth relationships for cirrus and deep stratiform ice cloud layers. J. Appl. Meteorol. 42, 1369–1390. Hogan, R.J., Jakob, C., Illingworth, A.J., 2001. Comparison of ECMWF winter-season cloud fraction with radar-derived values. J. Appl. Meteorol. 40, 513–525. Hollars, S., Fu, Q., Comstock, J., Ackerman, T., 2004. Comparison of cloud-top height retrievals from ground-based 35 GHz MMCR and GMS-5 satellite observations at ARM TWP Manus site. Atmos. Res. 72, 169–186. Jolivet, D., Feijt, A.J., 2005. Quantification of the accuracy of liquid water path fields derived from NOAA 16 advanced very high resolution radiometer over three ground stations using microwave radiometers. J. Geophys. Res. 110 (D11204). doi:10.1029/ 2004JD005205. King, M.D., Tsay, S.C., Platnick, S.E., Wang, M., Liou, K.-N., 1997. Cloud Retrieval Algorithms for MODIS: Optical Thickness, Effective Particle Radius, and Thermodynamic Phase. Algorithm Theoretical Basis Document ATBD-MOD-05. NASA Goddard Space Flight Center. King, M.D., Menzel, W.P., Kaufman, Y.J., Tanre, D., Gao, B.C., Platnick, S., Ackerman, S.A., Remer, L.A., Pincus, R., Hubanks, P.A., 2003. Cloud and aerosol properties, precipitable water, and profiles of temperature and humidity from MODIS. IEEE Trans. Geosci. Remote Sens. 41, 442–458. Klein, S.A., Jacob, C., 1999. Validation and sensitivities of frontal clouds simulated by the ECMWF Model. Mon. Weather Rev. 127, 2514–2531. Lau, N.-C., Crane, M.W., 1995. A satellite view of the synoptic-scale organization of cloud properties in midlatitude and tropical circulation systems. Mon. Weather Rev. 123, 1984–2006. Liou, K.N., 1992. Radiation and Cloud Processes in the Atmosphere. Oxford University Press, New York, NJ, USA. Mahesh, A., Gray, M.A., Palm, S.P., Hart, W.D., Spinhirne, J.D., 2004. Passive and active detection of clouds: comparison between
M. Schröder et al. / Atmospheric Research 82 (2006) 83–101 MODIS and GLAS observations. Geophys. Res. Lett. 31 (L04108). doi:10.1029/2003GL018859. Menzel, W.P., Bryan, B.A., Strabala, K.I., Frey, R.A., 2002. Cloud Top Properties and Cloud Phase—Algorithm Theoretical Basis Document. ATBD-MOD-04. NASA Goddard Space Flight Center. Pearson II, F., 1990. Map Projections: Theory and Applications. CRC Press, Boca Raton, FL, USA. Ritter, B., Geleyn, J.-F., 1992. A comprehensive radiation scheme for numerical weather prediction models with potential applications in climate simulations. Mon. Weather Rev. 120, 303–325. Rossow, W.B., Schiffer, R.A., 1991. ISCCP cloud data products. Bull. Am. Meteorol. Soc. 72, 2–20. Ryan, B.F., Katzfey, J.J., Abbs, D.J., Jakob, C., Lohmann, U., Rockel, B., Rotstayn, L.D., Stewart, R.E., Szeto, K.K., Tselioudis, G., Yau, M.K., 2000. Simulations of a cold front by cloud-resolving, limited-area, and large-scale models, and a model evaluation using in-situ and satellite observations. Mon. Weather Rev. 128, 3218–3235. Stephens, G.L., 1978. Radiation profiles in extended water clouds: II. Parameterization schemes. J. Atmos. Sci. 35, 2123–2132. van Lipzig, N.P.M., Schröder, M., Crewell, S., Ament, F., Chaboureau, J.-P., Löhnert, U., Matthias, V., van Meijgaard, E., Quante, M., Willén, U., Yen, W., 2006. Model predicted low-level cloud parameters. Part I: Comparison with observations from the BALTEX Bridge Campaigns. Atmos. Res. 82, 55–82. Van Meijgaard, E., Andrae, U., Rockel, B., 2001. Comparison of model predicted cloud parameters and surface radiative fluxes with
101
observations on the 100 km scale. Meteorol. Atmos. Phys. 77, 109–130. Walther, A., Bennartz, R., 2006. Radar-based precipitation type analysis in the Baltic area. Tellus-A 58 (3), 331–343. Wilks, D.S., 1995. Statistical methods in the atmospheric sciences: an introduction. International Geophysics Series, vol. 59. Academic Press, San Diego, CA, USA. Willén, U., Crewell, S., Baltink, H.K., Sievers, O., 2005. Assessing model predicted vertical cloud structure and cloud overlap with radar and lidar ceilometer observations for the Baltex Bridge Campaign of CLIWA-NET. Atmos. Res. 75, 227–255 (Special Issue: CLIWA-NET: Observation and Modelling of Liquid Water Clouds). Woodcock, F., 1976. The evaluation of yes/no forecast for scientific and administrative purposes. Mon. Weather Rev. 104, 1209–1214. Wyser, K., Jones, C.G., 2005. Modelled and observed clouds during Surface Heat Budget of the Arctic Ocean (SHEBA). J. Geophys. Res. 110 (D09207). doi:10.1029/2004JD004751. Yu, W., Doutriaux, M., Seze, G., Le Treut, H., Desbois, M., 1996. A methodology study of the validation of clouds in GCMs using ISCCP satellite observations. Clim. Dyn. 12, 389–401. Zhu, P., Bretherton, C.S., Köhler, M., Cheng, A., Chlond, A., Geng, Q., Austin, P., Golaz, J.C., Lenderink, G., Lock, A., Stevens, B., 2005. Intercomparison and interpretation of single column model simulations of a nocturnal stratocumulus topped marine boundary layer. Mon. Weather Rev. 133 (9), 2741–2758.
