Evaluating Cloud Properties Simulated by Largescale Models with Satellite Data

The evaluation of the clouds in GCMs has long been hampered by the lack of global observations of the vertical structure of clouds. The situation is now radically changing with the arrival of new observations from the A-Train constellation of satellites, including the spaceborne radar (CloudSat) and lidar (CALIOP/CALIPSO) instruments. The observational definition and detection of clouds, however, depends strongly on the type of measurements and sensitivity of sensors, as well as the vertical overlap of cloud layers in the atmosphere. This definition also differs from the definition of a cloud layer in large-scale models or in high-resolution models (e.g., CRMs). Therefore, a raw and direct comparison of cloud products derived from observations with model simulations does not guarantee that apples are not compared with oranges.

To make more meaningful comparisons between models and observations, it is better to use a simulator to diagnose from the model outputs some quantities that are directly comparable with observations. Such an approach has been widely used to compare model cloud covers with ISCCP data (e.g., Klein and Jakob 1999; Webb et al. 2001; Zhang et al. 2005). New simulators, which compare the observed radar and lidar backscatter profiles with those profiles calculated from the model parameters, are now under development. First studies using a CloudSat simulator (Bodas-Salcedo et al. 2008) and an ICESAT (Wilkinson et al. 2008) or a CALIPSO (Chepfer et al. 2008) lidar simulator show already how promising the approach is to evaluate the cloudiness simulated by climate models. Biases can now be identified much more clearly and in more detail (in particular, the vertical structure can be documented) than with previous comparisons using passive measurements.

Global comparisons of histograms of CloudSat radar reflectivity (Bodas-Salcedo et al. 2008) as a function of height computed over several months from the Met Office model, which has an implicit exponential ice particle size distribution with an intercept parameter that is a function of temperature and an ice particle density which is inversely proportional to size, appear perhaps initially discouraging. The observed histograms are much smoother than those observed, but they do show that the model is underestimating mid-level clouds. However, when the comparisons are subdivided into geographical regions, they are much more revealing. Over the North Atlantic, the model performance for the ice cloud is quite good, indicating that parameterization of the intercept parameter as a function of temperature performs well. Problems are evident for low cloud: the model has two separate drizzle regimes rather than one. Comparisons over the California stratocumulus region and the tropical Pacific also reveal specific errors in the model. As noted with the ground-based measurements, the lack of mid-level clouds seems to be ubiquitous, and the occurrence of drizzle in low-level clouds seems to be overestimated in the model. Clearly, such an approach has powerful implications, although care is needed to distinguish between the relative influence of cloud and precipitation biases in errors of the simulated radar reflectivities. As with the ground-based observations, comparing the statistics of the mean values of the reflectivity histograms with the observations are just the first step. The next step is to classify the data in terms of different weather regimes; this is accomplished by separating the data according to, for example, large-scale vertical motion and surface stability. Thereafter, those processes which are being poorly represented must be identifi ed to establish if, for example, the lack of mid-level clouds in the models is important for some aspects of the simulated climate.

The computation and interpretation of a radar simulator is reasonably straightforward in that the model holds an implicit size distribution of the ice particles, and for water droplets there is a prescribed droplet size over the ocean and over land, and that in general the attenuation of the radar signal is rather small. Lidar measurements are very sensitive to the presence of cloud particles, and the horizontal and vertical resolutions of the measurements are very high (330 m and 30 m, respectively, for CALIPSO). The analysis of lidar measurements thus constitutes a powerful means of diagnosing the vertical distribution of cloud layers and their overlap. However, the attenuation of lidar signals is much larger than that of radar reflectivities, so that the signal from the satellite can be totally extinguished at low altitudes when thick upper-level clouds are present. The attenuation is related to the observed lidar backscatter through the "lidar ratio," or the ratio of backscatter to extinction, but this lidar ratio is very sensitive to the (unknown) ice particle shape and size. In addition, the penetration of the lidar beam through multiple levels of broken cloud is very sensitive to the degree of cloud overlap; this could be considered as an advantage in that the very sensitivity to the cloud overlap could be regarded as an excellent method of diagnosing if the cloud overlap implied in the model is in fact realistic. In the presence of upper clouds, as in the tropics, the simulated and observed backscatter from lower-level clouds depends on how well the thicker higher-level cirrus clouds are represented. However, a large fraction of tropical oceans are associated with large-scale subsidence in the free troposphere, so boundary-layer clouds are not overlapped by upper-level clouds. In these situations, attenuation problems are minimal and the lidar is able to provide unambiguous returns from cloud top. The lidar simulator is thus particularly useful for studying those (ubiquitous) clouds which are important for the Earth's radiation budget but are often below the sensitivity of radar, such as stratus, stratocumulus, and fair weather cumulus clouds, which we identified earlier as being of crucial importance. Lidar can observe cloud top to 30 m, and thus, for an ensemble of clouds, it should be possible to identify the cloud base of these clouds. Cloud water droplets will generally yield a radar reflectivity too low to be detected from space, so any observed radar reflectivity will indicate the presence of small drizzle droplets or precipitation. If this is combined with inferred values of liquid water path in the cloud and effective radius from passive "MODIS"-type instruments, then the properties of the fair-weather cumulus clouds can be compared in detail with their representation in models. Evaluating the ability of climate models to simulate accurately the geometrical thickness and the precipitation efficiency of shallow-level clouds, together with their variation with natural climate fluctuations, is of paramount importance if we are to have confi dence in the simulated response of these cloud properties to climate change and then in the model cloud feedbacks. At high latitudes, the persistent low-level polar clouds should also be well detected by lidar measurements.

Note that the use of simulators is also a way to fill the gap between the different cloud scales since the comparison of the cloud covers predicted at the large scale can be compared to observations derived at a much smaller scale. For example the lidar signals are, in principle, available for each lidar pulse, with a horizontal resolution of 330 m or so for the highly reflecting water clouds and a vertical resolution of 30 m; the radar reflectivity has a horizontal resolution ofjust 1.1 km and 500 m in the vertical. Unfortunately, when we are considering the representation of tropical broken cumulus clouds, high-resolution observations of the PDF of humidity via Raman lidar do not seem to be possible from space, and, at present, only values of water vapor path integrated over the vertical are available with a horizontal resolution of some 20 km.

Turning to ice clouds, the ice particle size can be derived from the ratio of the radar return (which varies as the sixth power of the particle diameter) to the lidar backscatter signal (which, when corrected for attenuation, depends on the square of the particle diameter). The first stage would be to compare the inferred ice particle size and its variation globally with location and temperature, and then compare it with the particle size, which in most models is prescribed in terms of the temperature alone. Results of this analysis should be available very soon; Delanoë and Hogan (2008) have demonstrated how the errors of the derived products from a combination of active and passive sensors can be obtained using a variational technique. Depending upon the results, one can envisage having a prescribed ice particle size in the models which varies not only with temperature but also with other environmental conditions. The next stage could be to have a double moment scheme to represent the ice particles, as is done in CRMs, provided the particle size in such a scheme could be constrained to agree with the size inferred from the active radar and lidar onboard the satellites.

Supercooled layer clouds can be identified relatively easily from space by their very high lidar backscatter and sharp backscatter gradient at cloud top. Using data from the LITE mission on the space shuttle Hogan et al. (2004) found that around 20% of all clouds between -10°C and -15°C contained supercooled layers. Such thin layer clouds have a much larger radiative impact than ice clouds of the same water content because of their smaller particle size, yet they are scarcely represented in climate models. Quantification of the radiative impact of such clouds on a global scale will soon be possible using CALIPSO data.

In addition, CALIPSO lidar data provides us with high-resolution observations of aerosol backscatter, with the "color ratio" of backscatter at the two wavelengths and depolarization ratio giving us aerosol size and shape information, respectively. The origin of these particles may be desert dust lofted by convection. Clearly, the lidar observations have the potential to quantify the global occurrence of both anthropogenic and natural aerosol but cannot by themselves distinguish the two types. It should be possible to establish just how widespread is the modification by man of the sizes and concentrations of the droplets within liquid water clouds when such clouds are embedded within haze. For ice clouds, lidar returns should reveal the frequency with which dust aerosols (natural or anthropogenic) are being lofted and transported large distances, and whether they are significantly modifying the glaciation rates and ice particle sizes of these high-level clouds. An essential first step is to quantify the magnitude of these effects on a global scale.

Thus far we have discussed the evaluation of NWP models using satellite or ground observations to see if they are producing clouds with the correct average properties and the correct PDF of these properties. This requires several months of data. On a global scale, observations over a few years should be sufficient to establish the characteristics associated with the Madden-Julian Oscillation, interannual variability, and possibly El Niño. The ground-based studies have shown that in regions where there are abundant observations, NWP mesoscale models have skill in producing the right cloud at the right time, and this skill can be evaluated on a monthly basis. In data-sparse regions, this skill is much lower so only an evaluation of the correct statistical properties of the clouds can be achieved. Evaluating the fidelity of clouds in climate models run for many years is more difficult; only the global statistics of mean cloud properties, their PDFs, and the temporal fluctuations of these metrics can be determined. However, experiments in which climate models have been run in a forecast mode indicate that some systematic biases (e.g., in the cloud and humidity fields), noticed in the climate mode, appear in a few days in the model. This suggests that the evaluation of clouds in climate models may be done in part based on short-term experiments and high-frequency observations (e.g., data from field experiments if the model is initialized with large-scale forcings from this experiment). One word of warning is in order: currently, active satellites with active radars and lidars are in sun-synchronous orbits and thus information on the diurnal cycle is limited.

0 0

Post a comment