Low-cloud characteristics over the tropical western Pacific from ARM observations and CAM5 simulations
Abstract
This study evaluates the ability of the Community Atmospheric Model version 5 (CAM5) to reproduce low clouds observed by the Atmospheric Radiation Measurement (ARM) cloud radar at Manus Island of the tropical western Pacific during the Years of Tropical Convection. Here low clouds are defined as clouds with their tops below the freezing level and bases within the boundary layer. Low-cloud statistics in CAM5 simulations and ARM observations are compared in terms of their general occurrence, mean vertical profiles, fraction of precipitating versus nonprecipitating events, diurnal cycle, and monthly time series. Other types of clouds are included to put the comparison in a broader context. The comparison shows that the model overproduces total clouds and their precipitation fraction but underestimates low clouds in general. The model, however, produces excessive low clouds in a thin layer between 954 and 930 hPa, which coincides with excessive humidity near the top of the mixed layer. This suggests that the erroneously excessive low clouds stem from parameterization of both cloud and turbulence mixing. The model also fails to produce the observed diurnal cycle in low clouds, not exclusively due to the model coarse grid spacing that does not resolve Manus Island. This study demonstrates the utility of ARM long-term cloud observations in the tropical western Pacific in verifying low clouds simulated by global climate models, illustrates issues of using ARM observations in model validation, and provides an example of severe model biases in producing observed low clouds in the tropical western Pacific.
Key Points
- CAM5 underestimates total low clouds, produces excessive low clouds near mixed-layer top
- The macrophysics scheme is responsible for the excessive low clouds near the mixed-layer top
- The excessive humidity at the mixed-layer top is suggested due to vertical transport errors
1 Introduction
Global models show a large range of climate sensitivities, which arises mainly from differences in their cloud radiative feedbacks to a changing climate [Andrews et al., 2012; Bony et al., 2006]. Low-level cloud has been identified as a key uncertainty in model cloud feedback [Bony and Dufresne, 2005; Medeiros and Stevens, 2009; Wyant et al., 2006]. The previous generation of climate models tend to underestimate low-cloud cover [Boyle et al., 2005; McFarlane et al., 2007] and overestimate its radiative properties, such as the optical depth [Webb et al., 2001; Zhang et al., 2005; Karlsson et al., 2007]. This is particularly so in the tropical shallow cumulus regime [Nam et al., 2012]. Low-cloud climate feedback includes interaction between cloud radiative cooling, boundary layer relative humidity, and cloud cover [Brient and Bony, 2012]. In the tropical deep convective regime, such as the Indian and western Pacific warm pool, shallow cumuli may play a different important role in the weather-climate system. They have been proposed to be instrumental to large-scale disturbances such as the Madden-Julian Oscillations [Madden and Julian, 1971, 1972] through their moistening and heating effects on the lower troposphere and thereby setting the stage for transition from shallow convection to initiation and growth of deep convection [Del Genio et al., 2012; Zhang and Song, 2009].
Low clouds associated with shallow convection in weather and climate models must be represented through parameterization, because their spatial scales (~a few km) are much smaller than the model grid spacing. Designing and improving shallow cumulus parameterization have been a long-term endeavor. Biases in underestimating low-cloud cover has persisted to most recent models, namely, those that participated in the Coupled Model Intercomparison Project 5 (CMIP5). In these models, stratocumulus clouds are produced in place of shallow cumulus clouds; low clouds tend to concentrate in the lowest 1 km instead of spread throughout the boundary layer, and vertical structures of low clouds depend only weakly on large-scale environment conditions in contrast to observations [Nam et al., 2012].
There are two motivations for this current study. One is a recent new parameterization scheme for shallow convection [Park and Bretherton, 2009]. This scheme has been implemented in the Community Atmospheric Model version 5 (CAM5) with discernable improvement in its climate simulations [Neale et al., 2010]. It is desirable to assess against observations how shallow convective clouds in the tropical deep convection regime are produced in CAM5 with this new parameterization scheme in place. The second motivation of this study is the availability of long-term observations of shallow convective clouds in the tropical western Pacific from the U.S. Department of Energy ARM (Atmospheric Radiation Measurement) program tropical western Pacific sites at Manus, Darwin, and Nauru [Long et al., 2013]. These observations allow robust assessment of model performance in producing shallow cumuli.
In this study, we chose observations from the Manus site to evaluate low clouds in CAM5 simulations as a pilot study to establish a procedure for observations-model comparison. The model simulations we diagnosed in this study cover the Year of Tropical Convection (YOTC) period (May 2008 to April 2010) [Waliser et al., 2012]. Observations at Manus show abundant deep [Mather, 2005] and shallow [McFarlane et al., 2007] convective clouds. Previous evaluations of cloud fractions produced by global models against limited ARM observations from the ARM tropical western Pacific sites have suggested that models tend to underestimate shallow (or low) clouds with unrealistic vertical distributions [Comstock and Jakob, 2004; Boyle et al., 2005]. No model can produce the observed seasonal cycle of shallow cloud fractions [Qian et al., 2012].
Comparing point observations to model simulations poses many challenges. One important aspect of the comparison is to adopt a consistent methodology for identifying different cloud types in both models and observations. Most of the previous studies on model comparison and climate feedbacks [e.g., Jakob and Tselioudis, 2003; Stephens, 2005] follow International Satellite Cloud Climatology Program classification to define low clouds based only on the cloud top pressure (>680 hPa). This definition does not distinguish low-layered cloud from low cumulus clouds. The latter is the main target of this current study. Another important aspect in comparing point observations to a general circulation model grid is to choose an optimal time averaging that is representative of the model grid. Some of the previous studies [e.g., Morcrette, 2002; Song et al., 2014] used 1 h averages, and others [e.g., Ovchinnikov et al., 2006; Yang et al., 2006; McFarlane et al., 2007] used 3 h for large-scale model evaluation. Selecting a particular averaging period needs justification and requires sensitivity studies to understand the effect of averaging on variables being compared. Otherwise, the conclusions drawn from the comparisons may not be robust.
The present study is different from the previous ones in several aspects. We defined low clouds as clouds with their tops below the freezing level and bases within the boundary layer. We have tested effects of different averaging time on cloud fraction calculations. To ensure a fair cloud classification for model and observations, a particular cloud type was defined based on the continuity in cloud fraction profiles rather than defining them only based on the averaged cloud top; the latter may misclassify some deep clouds as low clouds. Low clouds were further separated into precipitating and nonprecipitating clouds, and rain rate distributions of precipitating clouds were documented. We scrutinized the vertical structures of low clouds using different methods (daily and monthly time series, the diurnal cycle, mean fraction, and probability density function (PDF) profiles, etc.) in the observation-model comparison to ensure the consistency and robustness of the results. None of these has been tried previously in model validation studies using data from the ARM Manus site.
Section 2 describes the data and methodology used. Results are presented in section 3, where we discuss general cloud characteristics but our focus is low clouds. Section 4 provides a summary and discussion.
2 Data and Methodology
2.1 Observations
Table 1 lists all data used in this study. Observations from the ARM Manus site (2.06°S, 147.425°E) are described in detail by Long et al. [2013]. The cloud locations and radar moments are from an integrated data product Active Remote Sensing Clouds (ARSCL), which combines data from four modes of the millimeter-wavelength (Ka-band at frequency of 35 GHz) cloud radar (MMCR), a laser ceilometer, and a micropulse lidar (MPL). The ARSCL product at a temporal resolution of 10 s and a vertical resolution of 45 m provides best-estimated profiles of hydrometeor reflectivity and cloud boundaries (bases and tops) up to 10 vertically stacked layers [Clothiaux et al., 2000; Kollias et al., 2007]. Reflectivity profiles from the MMCR-precipitation mode were used to identify the rainy profiles at 10 s resolution. Unless otherwise stated, MMCR reflectivity values used in this study were based on the ARSCL product.
| Datastream and Identifier | Measurement Used for the Study | Frequency | Site | Remarks |
|---|---|---|---|---|
| ARSCL, twparsclcloth C1.c1 | Cloud boundaries, reflectivity | 10 s | Manus | ARM value-added product |
| MMCR, twpmmcrmode4pr2200606161clothC1.c1. | Precip-mode reflectivity | ~8 s | Manus | Identifying rain profiles |
| ARMBEATM | T, RH, U, V | hourly | Manus | ARM value-added product |
| Twpsmet60sC1.b1 | Rain rate | 1 min | Manus | Tipping bucket rain gauge |
- a The observations from the Manus site are the main data set used to document the cloud statistics for the YOTC period (May 2008 to April 2010).
Vertical profiles of air temperature and humidity are from ARM Best Estimate Data Products (ARMBE) [Xie et al., 2010; McCoy and Xie, 2012]. Although the ARMBE algorithm updates data every hour, the data are available only when there are radiosonde observations (2100 and 0900 local time). Rain rates are from a tipping bucket rain gauge located within 100 m of the MMCR, ceilometer, and MPL. The accuracy of the rain rate measurement is 0.1 mm h−1, and its uncertainty ±0.1 mm h−1 [Habib et al., 2001]. Rain rates were interpolated to the ARSCL resolution (10 s).
2.2 Cloud-Type Identification
Low (Deep) cumulus clouds were identified when the cloud top pressure was greater (less) than 550 hPa (representing the average freezing level calculated from the soundings) and the corresponding cloud base pressure greater than 850 hPa (representing the typical boundary layer top height at Manus). A cloud base was detected from the laser ceilometer without rain, and its pressure from the nearest sounding.
Radar gates (vertical resolution, which is 45 m) were defined to contain cloud if reflectivity is greater than −50 dBZ. The reflectivity profiles from the MMCR-precipitation mode (8 s resolution) were interpolated to the ARSCL resolution (10 s) and used to identify rainy profiles. A reflectivity profile (from the MMCR-precipitation mode) was flagged as rainy if any one of the gates below the cloud base contains reflectivity values exceeding 0 dBZ. These rain flags were assigned to the corresponding ARSCL profiles. An event (3 h window) with at least one rain profile was defined as a rainy event. For a rainy event, profiles of cloud fractions were calculated only if rainy profiles are less than 25% of the total profiles and using only radar echoes of nonrain profiles because of practical limitations for separating cloud and rain layers in a column. Also, inclusion of rainy profiles would obscure the signal from low cloud, which is the focus of the present study. This methodology was adopted in previous studies for defining cloud fractions of precipitating events [McFarlane et al., 2007]. For precipitating events, a cloud fraction was defined as the number of nonprecipitating cloudy profiles to the total number of nonrainy profiles. Mean cloud base for each event is calculated by averaging the lowest 25% of the cloud base values from the ceilometer after sorting all the ceilometer detections in the ascending order. An event was flagged as clear sky if the radar was operating and there was no significant (greater than noise level) echo detected in any of the profiles within a 3 h period.
Cloud tops reported in the ARSCL product were detected from the gradient of MMCR reflectivity. The operating frequency of the MMCR (35 GHz, 8.66 mm) is more affected by attenuation from water vapor, cloud, and hydrometeors than those of the centimeter-wavelength precipitation radars (e.g., S band and C band). Because of this, cloud top height reported in the ARSCL product can be underestimated, which may result in false detection of deep clouds as low clouds. A previous study [Song et al., 2014] used an arbitrary rain rate threshold of 40 mm/d to eliminate attenuated cloud profiles over the Southern Great Plains. Whether this threshold may apply to a tropical region such as the western Pacific with heavy rain is unknown. We have examined this issue using the data from the Dynamics of Madden-Julian Oscillation field campaign [Yoneyama et al., 2013]. A correction procedure was developed to correct cloud top height derived from a Ka-band radar to account for the rain attenuation. The methodology and correction proposed is detailed in the supporting information. The main results from this exercise are the following: (1) attenuation effects are very significant for rain rates > 1 mm/h at 1 min resolution, and (2) attenuation effects on cloud top height are important for averaging time less than an hour. We have also tested (not shown) the feasibility of using Doppler velocities (instead of rain rates) as a measure of the attenuation. We found that corrections of cloud top height using either rain rates or Doppler velocities provide similar results.
The developed correction was applied to the ARSCL data at the Manus site at 10 s resolution. The correction was applied only to profiles with rain rates greater than 1 mm/h and cloud top height greater than the 550 hPa level. This prevents most true low clouds from being misclassified as deep. We found that the effect of attenuation on 10 s cloud profiles of ARSCL is minimal with less than 2% of deep clouds being misclassified as low clouds. For averaging windows greater than 2 h, there is no misclassification reported. Based on this, one can summarize that for the tropical sites, the effect of attenuation on the cloud top height can be neglected for time averages greater than 2 h. However, for studies that demand cloud top estimates with high rain rates at high resolution (less than an hour), correction of rain-attenuated cloud profiles is necessary. In this study, we used corrected cloud profiles because we tested the dependence of our results on time averaging that ranges from 1 to 12 h.
Observed profiles of cloud fractions were calculated by averaging the ARSCL clutter-free MMCR reflectivity profiles in a 3 h time period (hereafter referred to as an event). The justification for selecting the 3 h window is provided in section 3.2. For a nonrainy event, a particular cloud type was identified using uncorrected cloud fraction profiles, while for a precipitating event it was defined using corrected cloud top height. Continuous, nonzero cloud fraction profiles were treated as a cloud segment. A low (deep) cloud was defined when the pressure at the base of a cloud fraction profile in a particular cloud segment is greater than 850 hPa and the pressure at the corresponding top of the cloud fraction profile in that segment is greater (less) than 550 hPa. All other clouds with their bases above the mixed-layer top (850 hPa) are referred to as “other cloud types.” This definition was applied to both model simulations and observations.
2.3 CAM5 Simulations
The model output is from CAM5. The model horizontal grid spacing is 0.98° in longitude and 1.258° in latitude. There are 30 vertical levels (four in the boundary layer). One advancement from CAM4 to CAM5 relevant to this study is a new shallow cumulus scheme with a plume dilution and more refined estimates of entrainment-detrainment rates, cloud top height, and penetrative entrainment [Park and Bretherton, 2009]. Under the Department of Energy Cloud-Associated Parameterizations Testbed (CAPT) protocol (http://www-pcmdi.llnl.gov/projects/capt/index.php), 6 day hindcasts were performed using CAM5 with initialization every day at 0000 UTC using the European Centre for Medium-Range Weather Forecasts analysis for the YOTC period. This approach of climate model simulations in the numerical weather prediction mode with short-term runs initialized using global reanalysis products has proven efficient to reveal model systematic biases and errors [Yang et al., 2006]. The lower boundary condition of the CAM5 simulations comes from National Oceanic and Atmospheric Administration optimum interpolation weekly sea surface temperature. The model output is available for every hour. The details of the CAPT model output are described in Ma et al. [2013]. We have used day 2 (25–48 h) output in this study. Figure 1 shows the fractional land at CAM5 grids in the western Pacific region. Data from two model grid boxes (centered at 1.4136°S, 147.5°E and 2.3506°S, 147.50°E, respectively) closest to the ARM Manus site location were extracted and averaged to represent model output at Manus, which will be referred to as the Manus grid. Its center is labeled as a red cross in Figure 1. Notice that the land fraction at the model Manus grid is zero. The red square represents a land area chosen to test the diurnal cycle in the model (see section 3.3).
2.4 Model Cloud Types
There are two types of clouds in CAM5 simulations: Stratus and cumulus. Their fractions are grouped into four classes: Shallow cumulus (ash,cu), deep cumulus (adeep,cu), liquid stratus (al,st), and ice stratus (ai,st). The shallow cumulus parameterization scheme of Park and Bretherton [2009] produces both shallow and deep cumulus fractions. A deep cumulus scheme of Zhang and Mcfarlane [1995] produces deep cumulus fractions. The fraction of shallow and deep clouds combined is referred to as convective cloud fraction (aconv,cu = ash,cu + adeep,cu). Stratus clouds (at both high and low levels) are produced by a macrophysics scheme [Park et al., 2014]. The total cloud fraction consists of convective and stratus clouds at each grid point (or box), vertical level, and time step.
At a grid point, the lowest and highest levels of continuous nonzero values of the cloud fraction represent the base and top of a cloud layer, respectively. As in the observations, low clouds are identified if the cloud top pressure is greater than 550 hPa and the corresponding cloud base pressure greater than 850 hPa. Low cloud thus defined can be produced by either the shallow or deep cumulus schemes or macrophysics scheme. Identified low clouds are further classified into precipitating and nonprecipitating ones based on whether simultaneous precipitation is produced by the shallow convection, deep convection, or macrophysics schemes.
2.5 Model-Observation Comparison
Comparisons between model output and observations must face a number of issues related to the mismatch between model grid sizes and point measurements, and between model time steps and instrument sampling frequencies. Most previous studies made direct comparisons between point measurements and model grid box output under the assumptions that temporal sampling at a point yields the equivalent of a two-dimensional slice through a three-dimensional model grid box, and the fraction of the grid filled by clouds in the two-dimensional grid slice represents the amount of cloud in the three-dimensional volume [Mace et al., 1998; Hogan et al., 2001; Comstock and Jakob, 2004; Boyle et al., 2005; Bouniol et al., 2010]. Attempts have been made to quantitatively relate point measurements to areal averages represented by model grid boxes as a function of time averaging [Barnett et al., 1998; Long, 2002; Yang et al., 2006]. As the model horizontal resolution becomes coarser, longer time averaging is needed from observations in order to match the scales. Some studies considered 1 h [e.g., Morcrette, 2002; Song et al., 2014; Zhu, 2004; Wu et al., 2014] and others used 3 h [Ovchinnikov et al., 2006; Yang et al., 2006; McFarlane et al., 2007] as the averaging time in their model evaluation studies. Jakob et al. [2004] suggested model cloud can be interpreted as probabilistic at an observation point, and comparisons between model output and point measurements be done statistically without any time averaging of observations to fit the model resolutions. Based on a simple time-space transformation, a 3 h averaging period would represent a spatial scale equivalent to model resolution of 100 km assuming wind speeds of ~ 10 ms−1 in the cloud layer. In this study, we used a 3 h averaging window (centered on model time steps) for the observations in order to match the model grid resolution (~100 km) and time step (1 h) in statistical comparison between them. A justification of this choice is given in section 3.3.
There are a total of 16536 usable hours of data from the CAM5 simulations, and 15180 from the observations. We made an assumption that the small discrepancies in the sample sizes would not percolate into their statistical comparisons. All samples were used in our analysis.
3 Results
3.1 All-Sky Cloud Fraction
Detection of cloud presence and fraction depends on the instrument used [Wu et al., 2014]. We first compared time series of all-sky cloud fractions from the ARSCL product based on data from two instruments: MMCR clutter-free reflectivities and MPL cloud masks (clear, cloudy, and clutter flags). Monthly mean time series of cloud fractions from the MMCR, MPL, ceilometer, and CAM5 are compared in Figure 2. The MPL data were missing between October 2008 and June 2009 due to a component failure at the Manus site. Though the MPL component was replaced around June 2009, there were issues noticed with the MPL data, which significantly affected the overall MPL sensitivity to detect clouds at altitudes higher than around 5 km (C. Flynn, Pacific Northwest National Laboratory, personal communication, 2015). For this reason, we have plotted MPL data only up to September 2008 during the YOTC. The total cloud fraction from the MPL is higher than the MMCR. This is due to its higher sensitivity to detect thin low clouds and cirrus clouds better than the MMCR. The ceilometer is primarily meant to detect liquid clouds, which resulted in missing high cloud bases (greater than 6 km). The trend in the ceilometer total cloud fraction matches well with the MMCR cloud fraction (up to July 2009), which suggests that the liquid clouds are mainly responsible for this evolution. The lack of matching in the trends between them (after July 2009) could be due the high frequency of high cloud occurrences, which ceilometer cannot detect. CAM5 total cloud fraction exceeds the MMCR for some months (preferably May–June), which can be mainly attributed to the missing thin cirrus and wimpy shallow clouds from the MMCR. The reason for the CAM5 total cloud fraction to be lower than the MMCR is possibly due to the underproduction of low clouds, which will be examined in the later analysis. The discrepancies in the total cloud fraction between different instruments and the model could be understood better by looking at the vertical distribution of the clouds (profiles of their cloud fractions), which will be presented in the next section.
3.2 General Cloud Characteristics
Profiles of mean cloud fractions are calculated from the MMCR and MPL observations using a 3 h window, and from the hourly CAM5 cloud fraction profiles. The missing days (October 2008 and June 2009) were excluded from calculations of the mean cloud fraction profiles. Though the MPL sensitivity is affected even for the later period (after June 2009), which would not have significantly affected the cloud detection at the lower levels (below 5 km) (C. Flynn, personal communication, 2015). For this reason, we have used MPL data for the later period (after June 2009) only to make qualitative comparison with the MMCR data with a focus on low clouds. The cloud fraction reported from the MPL should be treated as a lower bound.
A comparison between mean cloud fraction profiles from MMCR, MPL, and ceilometer (Figure 3a) reveals large differences between them. The profile of the mean cloud fraction from the MMCR shows less variability (smoother) than that of the MPL. The MPL detects more low clouds and high clouds, including cirrus, than the MMCR but misses the dominant peak near 9 km seen by the MMCR. The zoomed-in comparison below 3 km (inlet) shows three cloud layers (centered around 500 m, 800 m, and 2 km) distinctly seen from the MPL. The MMCR detects the upper two (with some offset in height) but misses the lowest one (at 500 m) because of its poor sensitivity. The ceilometer captured two peaks in the cloud fraction: increased cloud fraction near the cloud base around 500 m and the other at 1800 m. The lowest peak (around 500 m) from the ceilometer matches closely with the MMCR, but less than the MPL by more than a factor of 2. It is partly because ceilometer is unable detect thin clouds with visibility less than hundred meters. A previous study by McFarlane et al. [2007] at the Manus site also reported the peak in the MMCR cloud fraction at around 800 m, but not the peak around 2 km. This is may be due to the short analysis period (6 months: February 2000 to July 2000) in their study. The lowest peak in the cloud fraction (around 500 m) from the MPL is nearly at the mean cloud base where the cloud fraction is highest, which is confirmed by the peak in the ceilometer cloud fraction.
The MPL shows a peak around 5.5 km, which corresponds to the melting layer. A previous study by McFarlane et al. [2007], hereafter referred to as MF07, also documented this peak from the MMCR. The higher peak (10 km) from the MMCR is slightly lower than that in MF07 (around 12 km). This is due to the short analysis period (6 months) in MF07 and the seasonality in the high clouds observed at Manus [Mather, 2005]. One should note that, although the MPL performs better than the MMCR in detecting high-level thin cirrus clouds, it might underestimate thin cirrus when there are low-level clouds obscuring its signal. Unlike MPL, the ceilometer optics is not affected by the Sun contamination. However, its limitation for not detecting thin clouds (due to vertical visibility threshold of ~ 100 m) and no cloud information above the cloud base adds no extra information over MPL for studying the vertical distribution of low clouds.
For better model-observation comparisons, we averaged cloud fractions from the observations within each model vertical layer in pressure coordinates (Figure 3b). The model overproduces high clouds (with peak around 200 hPa), which is consistent with MF07, where MMCR observations were compared to CAM3 simulations. A striking difference is that the model also over produces the low clouds (with peak around 965 hPa) by more than a factor of 2, which is not the case in MF07. There are systematic discrepancies in cloud fractions detected by the MMCR and MPL (except around 500 m). Unfortunately, due to a large gap (9 months) in the MPL data and also MPL sensitivity issues thereafter, it is impossible to use them in place of the MMCR data. But their discrepancies are much smaller than those between the observations and simulations. Hereafter, we use only the MMCR observations from the ARSCL product to validate the model simulations. The conclusion drawn between ARSCL and CAM5 simulations, particularly the cloud fraction (magnitude), should be treated as qualitative.
Daily profiles of cloud fractions from the MMCR observations (averages of 3-hourly cloud fractions over 24 h) and CAM5 simulations (averages of hourly cloud fractions over 24 h) are shown for the YOTC period (Figures 4a and 4b). The MMCR observations are available for most of the time except for three episodes (1 November 2008 to 5 November 2008, 10 December 2008 to 4 February 2009, and 24 May 2009 to 30 May 2009) of missing data as marked by grey areas in Figure 4a. The two time series offer a glimpse of total cloud behaviors at or near Manus from the observations and simulations. The Manus site is situated in the western Pacific warm pool and experiences frequent deep convection throughout the year. This is evident from the deep-cloud vertical extents (from near the surface to 16 km and above) during all seasons in both observations and simulations. This feature of frequent deep convection at the Manus site is also reported in MF07. Most observed cloud fractions are capped by the tropopause (minimum tropospheric temperature detected by the sounding data, marked by the red line in Figure 4a). The most obvious discrepancies between the two time series are abundant high clouds in the simulation that are missing in the observations and the dearth of middle-level clouds in the simulations in comparison to the observations. In the lower troposphere, there is a particular layer of high cloud fractions near 944 hPa in the simulation that is absent from the MMCR observations, which will be one of the main subjects of discussion later in this study.
Figure 5a shows mean profiles of MMCR cloud fractions for all cloud events (black) and only low cloud events (red). The total cloud fraction reaches a maximum at 250 hPa. The low-level peak (around 800 m) is clearly visible only in the height coordinates as in Figure 3a. The PDF of cloud base profiles shows three peaks, while that of cloud top shows two (Figure 5b). The sharp increase in the observed cloud tops around 550 hPa might be due to the presence of the stable layer near the melting layer as observed by Riihimaki et al. [2012] over the Darwin site. They suggested that the thin midlevel altocumulus are important component contributing to the peak at 550 hPa, which seems to explain the difference between total and low cloud fraction below freezing level observed in this study.
The profile of the total cloud fraction of CAM5 (Figure 5c, black line) shows a dominant peak in the upper troposphere and a secondary peak in the lower troposphere. CAM5 produced a dominant PDF peak of the cloud top in the upper troposphere and a dominant peak of the cloud base near the surface (Figure 5d); both are absent from the MMCR observations (Figure 5b). The large discrepancies in the high-level cloud top occurrence (Figure 5d) may come from errors in both the observations and model. The MMCR cannot well detect thin cirrus clouds, which significantly contributes to this peak in the simulations. Meanwhile, CAM5 may overproduce deep convective and cirrus clouds. The fraction profile of low-cloud events (Figure 5a, red line) decreases almost monotonically with height in observations. In contrast, it peaks around 930 hPa in CAM5 (Figure 5c, red line). Possible causes for this spurious peak of cloud fraction in the simulations are discussed in section 3.3.
Table 2 summarizes percentage occurrence of different cloud events from the MMCR observations and CAM5 simulations. It is clear that CAM5 produced an insufficient low cloud fraction (~13% of total cloud events) compared to that in the MMCR observations (~48%). Another major discrepancy is that most low clouds (>90%) precipitate in CAM5, whereas only around 36% precipitate in the observations. The fact that about one quarter of the total cloud fraction is low cloud (Figure 5a) yet their event number contributes to 48% of the total confirms that low clouds are small in horizontal scale and short in duration, hence their nominal contribution to the total cloud fraction. Deep clouds are much more frequent in the simulations (~86%) than in the observations (25%). In both simulation and observations, most deep clouds precipitate. The occurrence of other types of cloud (all cloud events except low and deep clouds) is less frequent in the simulation (<2%) than in the observations (27%). They also rain too often in the simulation (~60%) in comparison to the observations (~10%). Previous studies [Zhang et al., 2010; Badas-Salcedo et al., 2008] also observed that models (CAM 3.1 and UK Met Office forecast model) produce more precipitating low clouds than observed.
| Low Cloud | Deep Clouds | Other Clouds (With Bases Above 1.5 km) | |
|---|---|---|---|
| CAM5 (16536 h) | 12.54% (2071) 92.85% (1923) | 85.90%(14181) 99% (14048) | 1.71% (283) 60.77% (172) |
| MMCR (15180 h) | 47.73% (6633) 35.79% (2374) | 25.23% (3507) 97.38% (3415) | 26.69% (3756) 11.53% (433) |
- a Percentages of precipitating clouds for each cloud type are given in italic.
To illustrate the contribution of precipitation from low clouds to total rainfall in the simulations, we examined different physics that produce precipitation. In the simulations, precipitating events were identified by total rain at the surface, which is the sum of rain produced by the shallow, deep, and macrophysics schemes. Table 3 shows a decomposition of these precipitating low- and deep-cloud events into different schemes. For precipitating low-cloud events (2071), the macrophysics scheme almost always produces precipitation (in 1923 events); in about half (1097) of the 2071 events, the shallow scheme produces precipitation without rain from the deep scheme. In 826 of 14181 deep precipitating events, the shallow scheme produces rain while the deep scheme does not. This illustrates that most precipitation in the model is not produced by low clouds, even though simulated low clouds rain too often.
| Shallow Convection Scheme (SS) | Deep Convection Scheme (DS) | Macrophysics (LS) | SS U DS | SS ∩ DS | SS - DS | SS - DS | |
|---|---|---|---|---|---|---|---|
| Low clouds (2,071 events) | 1,784 | 826 | 1,923 | 1,923 | 687 | 1,097 | 139 |
| Deep clouds (14,181 events) | 7,225 | 13,222 | 14,037 | 14,048 | 6,399 | 826 | 6,823 |
- a The union operator “U” means “or” and the intersection “∩” means “both.” The operator “ - ” between A and B (e.g., A - B) means in A but not in B.
3.3 Low Clouds: Mean Profiles and Structures
Vertical PDFs of low cloud fractions in Figure 6 illustrate the main discrepancies between the observations and simulations. There is a peak of low clouds in a layer between 955 and 932 hPa in the simulations, which is absent from the observations. The peak is near the mean cloud base, and its shallow layer renders characteristics of stratiform clouds. CMIP5 model comparisons also reported that low clouds are concentrated in the lowest 1 km instead of spread throughout the boundary layer [Nam et al., 2012]. A few of the CMIP5 models produce a single stratocumulus-like layer near the cloud base instead of shallow cumulus clouds [Nuijens et al., 2015]. Immediately, above this peak (880–733 hPa), low clouds are underestimated in the simulations in comparison to the observations. Further up (above the 733 hPa level), low clouds are overestimated in the simulations. To get more insight to the lower end of cloud fraction values (<0.025), we compared the frequency occurrence of cloud fraction between the MMCR and CAM5 using all low cloud fraction values from the 955 to 932 hPa layer (Figure 6c). It shows that CAM5 produces low cloud fractions more frequently than the MMCR. This may partially be attributed to the difficulty of the MMCR to detect very low fraction of low clouds.
The stark contrast in the low-cloud profiles between the simulations and observations, particularly the low-level (955–932 hPa) peak only in the simulations, must be explained. One possibility is that the observations may have missed the low-level peak because of the specific time averaging. Another is underestimate of low clouds by the MMCR in comparison to MPL at lower levels (Figure 3a inset). To evaluate these possibilities, we examined the dependence of MMCR profiles of different cloud types on averaging time (T). While numbers of clear-sky and deep-cloud events decrease with increasing T gradually, numbers of low and other types of clouds decrease exponentially as T increases from 1 to 6 h (Figure 7a). Mean low cloud fractions detected by the MMCR decrease quickly with increasing T from 1 to 2 h at all levels but gradually with larger T (Figure 7b). For averaging time greater than 3 h, the mean low cloud fraction changes only by a factor less than 0.3. There is a tendency that the profiles converge as T increases further. Compared to the MPL, the MMCR underestimates the total low cloud fraction by about a factor of two. These uncertainties in the observations, however, cannot explain the excessive low cloud fraction in the layer of 955–932 hPa produced by the model (Figure 3b). The probability distribution of low cloud fractions in this layer as a function of T is shown in Figure 7c. While chances of high fractions (>0.1) to be detected generally decrease with increasing T, the difference is less than 1%. All these rule out a particular choice of averaging time T as a reason for the model-observation discrepancy in the layer of 955–932 hPa where a large peak of low cloud fraction exists only in simulations. Hereafter, we consider 3 h as an optimal choice for averaging observables to compare with the model statistics.
The shape of the low-cloud profile is dictated by the mixing of cloudy and environmental air and also due to the variability in buoyancy profiles or surface forcing. The decrease in cloud cover with height is indicative of decrease in mass fluxes. This is consistent with the observed profiles of low cloud fraction at Manus with a maximum at the cloud base and a decrease with height (Figure 8a). In addition to the spurious low-level peak, simulated low clouds are underestimated in a layer of 847–798 hPa and overestimated above that (Figure 8b).
We next address the question as which part of the model is responsible for the large cloudiness (lower peak) near the cloud base. We compared the CAM5 profiles of low clouds produced by different parameterization schemes. There is no low-level peak in the mass flux of low cloud produced by the deep cumulus scheme (Figure 8e, red line). There is a weak low-level peak in the mass flux of low cloud produced by the shallow cumulus scheme (Figure 8e, black line). But the shallow scheme is not mainly responsible for the erroneous peak in the total low cloud fraction. The erroneous low-level peak in the simulated total cloud fraction (Figure 8b) is absent in the profile of the total convective cloud fraction produced by the deep and shallow schemes in combination (Figure 8d). The only remaining suspect is the macrophysics cloud scheme. The fact that the macrophysics cloud scheme relies a lot on relative humidity and the coincidence between the erroneous low-cloud peak and the relative humidity peak (Figure 8c) supports this suspicion. One possible reason for the macrophysics cloud scheme to overproduce low clouds is excessive moisture that reaches saturation near the mixed-layer top, which is too high in the simulations in comparison to the sounding observations (Figure 8c). This might be related to insufficient vertical mixing by the turbulence scheme and shallow cumulus transport.
Vertical structures of low clouds were further examined in terms of the PDFs of cloud bases, tops, and depths. For fair comparisons, the MMCR observations of a finer vertical resolution (45 m) were averaged over the model vertical layers with layer centers as averaging bin centers. In the CAM5 simulations, low-cloud bases are biased toward lower levels (below the 955 hPa level) in comparison to those detected by the ceilometer (Figure 9a), tops of low clouds have an erroneous peak in the 798–733 layer (Figure 9b), and low-cloud depths show an unrealistic peak at 1.5–2 km (Figure 9c).
3.4 Low Clouds: Monthly and Diurnal Statistics
We now reexamine the time series of cloud fractions shown in Figures 4a and 4b but for monthly means of low clouds: monthly averages of 3-hourly cloud fractions from the observations and hourly cloud fractions from the simulations. Model errors in low clouds are again evident. The distinct but spurious peak in the CAM5 low cloud fraction in a thin layer between 955 and 932 hPa, seen in Figures 4-8, is obvious in the monthly mean (Figure 10). This large cloudiness peak coincides with the layer of maximum relative humidity in the simulations (Figure 10d), which is about 10% higher than in the observations (Figure 10c). Boundary layer heights are included in the output of the CAM5 simulations. In the observations, the boundary layer heights are estimated by degrading the original soundings to the resolution of the model in order to make the comparison independent of the vertical resolution. The observed values are estimated based on two approaches: one is the algorithm of Heffter (1980) and another is an algorithm based on the Richardson number (with critical thresholds of 0.5 and 0.25). We have applied the same methodology and thresholds as defined in the ARM Value added product (VAP) for planetary boundary layer heights [Sivaraman and Rihimaki, 2001]. Boundary layer heights are lower for most of the months in the simulations than in the observations (Figures 10c and 10d). Also, the variability in the observed boundary layer heights is larger compared to the simulations. Assuming that the mean of the observational estimates from different approaches are representative of the observations, we can say that the model boundary layer heights are underestimated at least by a model grid resolution. The excess humidity located near the mixed-layer top and shallow boundary layer in the simulations again suggests errors in the vertical mixing and transport as the main reason for the excessive humidity there, which in turn leads to excessive low clouds produced by the macrophysics cloud scheme. This resonates with the recent studies [Nuijens et al., 2015] that show several climate models systematically overestimate amount of cloud near the mean cloud base, often along with a too shallow boundary layer.
The diurnal cycle is an important component of the cloud variability at major islands of the Maritime Continent. For example, strong diurnal signals were observed up to 600 km off the coast of New Guinea [Liberti et al., 2000]. CAM5 does not resolve Manus Island (Figure 1). So one would not expect a strong diurnal cycle at Manus in the simulations. This is indeed the case. However, it is important to document the observed diurnal cycle in low clouds and examine whether or not the missing diurnal cycle in the model is entirely due to its inability of resolving the land fraction. Observed low clouds exhibit a diurnal cycle in their fraction at Manus (Figure 11a) with a peak during local afternoon (2–3 P.M.) and a minimum near midnight. There is no sign of any diurnal cycle in simulated low clouds (Figure 11c). The high cloudiness peak between 955 and 932 hPa persists at all hours. The sounding observations show discernable differences in relative humidity in the boundary layer between later morning and midnight (Figure 11b). There is no diurnal fluctuation in simulated humidity (Figure 11d). The observed two peaks of low clouds near the surface might be related to convergences due to land and sea breezes.
To determine the degree to which CAM5 misses the diurnal cycle at Manus because of the unresolved island, we selected an area northwest of New Guinea (red square in Figure 1) where the model does resolve the local island. We found no diurnal cycle in low cloud at this location either. This indicates that even if the model resolved Manus Island, it may still miss the observed diurnal cycle.
3.5 Low Cloud: Precipitation Characteristics
Previous studies [Johnson et al., 1999; Mapes, 2000; Zhang and Hagos, 2009] have proposed that shallow convection might play a unique role in the tropics because of its effects of moistening and/or heating. Both precipitating and nonprecipitating low clouds may moisten the environment through evaporation of condensates and detrainment of uncondensed vapor. Only precipitating low clouds release net latent heat through formation of rain. Shallow precipitation contributes up to 20% of the total precipitation over the tropical oceans [Short and Nakamura, 2000; Schumacher and Houze, 2003]. For a better understanding of the role of low clouds in the tropical cloud life cycle and circulation, it is desirable to know their occurrence of precipitation and rain rate distribution in comparison to other types of clouds.
Out of the total 6633 observed low-cloud events, about 36% (2374) precipitated. Nonprecipitating low clouds from the CAM5 simulations were not included in the following diagnostics, because most simulated low clouds precipitate as shown in section 3.2. In spite of this, documenting their properties from model and observations provides some insight into the representativeness of point observations of rain rates versus spatial domain. Observed vertical distributions of precipitating and nonprecipitating low clouds, distributions of rain rates for low and deep clouds were documented. The contribution of rainfall from the shallow and deep schemes in the low- and deep-cloud events was examined. Figures 12a–12c show vertical distributions of cloud fractions for observed precipitating, nonprecipitating, and all low clouds. At a given vertical level, the precipitating low cloud fraction undergoes a larger variability than nonprecipitating clouds. This is perhaps related to the fact that nonprecipitating clouds have a limited cloud depth, whereas precipitating clouds tend to be deeper, which are also likely to be wider, and their fractions larger. As nonprecipitating clouds grow larger, they are likely to become precipitating. Thus, a wider range of sizes and fractions for precipitating clouds than nonprecipitating clouds measured by the MMCR. The mean and 75th percentile lines are very close to each other for the nonprecipitating clouds, indicating their occurrences with very high fractions are rare. Again, the simulated cloud fraction peaks spuriously in a thin layer between 955 and 932 hPa (Figure 12d). The variability of CAM5 low (precipitating) clouds is smaller than the observed except at the layer of the erroneous peak.
We now compare rain rate distributions for different types of clouds in the simulations and observations. The mismatch between a model grid box and a point measurement discussed in section 2 would be more of a concern for rain rate comparisons. It turns out that precipitation from the low, deep, and all clouds observed by the rain gauges is almost uniformly distributed over a broad range of rain rates (0.01–10 mm h−1) with only weak peaks near 0.16 mm h−1 (Figure 13a). There is no clear distinction between rain rates from low and deep clouds, and their contributions to the total rain. This suggests that rain rates from the rain gauges and cloud fraction from MMCR are not suitable for distinguishing characteristics of rain rates from shallow and deep clouds.
In contrast, deep clouds in the simulations tend to produce higher rain rates (peak at 0.7 mm h−1) than low clouds (peak at 0.06 mm h−1) as expected (Figure 13b). The deep scheme produces much higher rain rates for deep clouds than shallow clouds, and the shallow scheme produce very light rain (drizzle) for both low and deep clouds (Figure 13c). What not expected is that the deep scheme produces a wide range of rain rates for shallow clouds. Although the shallow scheme is active ~ 50% for low-cloud events, deep scheme accounts for most (>95%) of the total accumulated rain from low clouds.
4 Summary and Discussion
- In the tropics, the effect of attenuation on low-cloud tops detected by cloud radar can be neglected for time averages greater than 2 h. For studies that demand cloud top estimates at high time resolution (less than an hour), a correction of rain-attenuated cloud tops would be necessary.
- CAM5 underestimates total low-cloud events (13% of total cloud hours versus 48% in the observations), overproduces precipitating low clouds fraction (93% versus 27% in the observations), and overproduces deep clouds (86% versus 25% in the observations). The shallow scheme produces precipitation without rain from the deep scheme for 50% of the total low-cloud events. This suggests that both deep and shallow schemes are responsible for the overproduction of precipitation from low clouds.
- The observed low cloud fraction shows the typical mean profile that decreases with height. In a stark contrast, the simulated low cloud fraction peaks spuriously at 955–932 hPa. This spurious peak of large cloudiness exists persistently and consistently in many different realizations of the data (daily and monthly time series, the diurnal cycle, mean fractions, and PDF profiles, etc.). By no means, should this be interpreted as a result of mismatch between point observations and grid box output of the simulations. It is an unmistakable model error.
- The spurious peak of low clouds in the CAM5 simulations coincides with excessive humidity near the top of the mixed layer. It is suggested that the erroneous peak in the low cloud fraction is produced by the cloud macrophysics scheme in response to the excessive humidity that stems from insufficient vertical mixing and transport.
- The absence of the observed diurnal cycle in both low cloud fraction and boundary layer humidity at Manus in the simulations cannot be attributed to the unresolved island because of the coarse grid spacing. It must be an indication of deficiencies in model physics.
This study demonstrates the utility of the observations of the ARM western Pacific sites in model validation. While results from this study are consistent with those from previous ones using limited record of ARM observations [McFarlane et al., 2007; Comstock and Jakob, 2004; Boyle et al., 2005], the need of special care when using the ARM observations in model validation was emphasized in this study. Two examples are the consistent definition of cloud type for observations and simulations, and averaging times needed to address the mismatch between point observational measurement and grid box output from model simulations. While our attempt at addressing these issues is rudimentary, the robust signals of model errors are not related to any imperfectness of our method. It would be interesting to explore whether the same biases and errors exist in other global models.
Acknowledgments
The authors thank Louise Nuijens and two anonymous reviewers for their comments and suggestions on the earlier version of this manuscript. We would also like to thank Karen Johnson from Brookhaven National Laboratory and Connor Flynn from Pacific Northwest National Laboratory for sharing the details on ARSCL missing data at Manus site. This study was supported by the U.S. Department of Energy Atmospheric Science Research program through grant DE-SC0006808. The observational data are available from the U.S. Department of Energy SGP ARM Climate Research Facility (http://www.archive.arm.gov). The CAM5 simulations were performed under the Department of Energy Cloud-Associated Parameterizations Testbed (CAPT: http://www-pcmdi.llnl.gov/projects/capt/index.php) protocol. Their data are available on request from CAPT coinvestigators Stephen A. Klein, [email protected], and Hsi-Yen Ma, [email protected].
