Cloud detection, as part of quality control procedures to reject cloud-affected radiances, is crucial for proper assimilation of clear radiances measured by high-spectral- resolution infrared (IR) sounders ( McNally, 2002). There are two classes of cloud-detection schemes. One class inc-ludes schemes that use information from imagers (e.g., the Moderate Resolution Imaging Spectroradiometer (MODIS), the Advanced Very High Resolution Radiometer (AVHRR), and the Advanced Infrared Sounder/Visible and Near IR (AIRS/VisNIR) instrument) in the vis-ible and near infrared channels to exploit the potentially cloud-affected IR observations ( Saunders, 1988; Ackerman et al., 1998). Th-ese schemes require simultaneous ava-ilability of two sou-rces of independent data and have a high computational

cost. The schemes in the other class seek to detect clouds by comparing the measured radiance and the equivalent clear-sky radiance ( Eyre and Menzel, 1989; Smith and Richard, 1990; Huang et al., 2004; Li et al., 2004). Detecting channels unaffected by cloud according to their peak levels allows a much better use of available data and avoids discarding potentially useful information. McNally and Watts (2003) developed a clo-ud-detection scheme (MW03) by isolating the cloud signal in the vertically ranked departure vector, which is used operationally in the European Centre for Medium-Range Weather Forecasts (ECMWF) and the French glo-bal model Action de Recherche Petite Echelle Grande Echelle (ARPEGE), as well as in other research ( Guidard et al., 2011; Xu et al., 2013). The MW03 cloud-detection scheme is widely used because of its computational efficiency and algorithmic simplicity. All the inputs for the MW03 cloud-detection algorithm are efficiently provided by the radiative transfer models as part of the clear-sky calculation and the assimilation procedure. The Multivariate and Minimum Res-idual (MMR) scheme uses a different approach ( Auligné, 2007, 2014a, b; Xu et al., 2014), which identifies the channels that are affected by cloud by simulating the variation in the cloudy radiance to fit the observations. Different from the traditional IR cloud-detection schemes that discard the whole pixel when cloud is present, both the MW03 and MMR schemes try to identify cloud-free channels by identifying the signals of cloud height, rather than cloud masks, thereby retaining as much information as possible from radiance observations.

The MW03 codes developed by ECMWF are incorporated into the WRF Data Assimilation (WRFDA) ( Barker et al., 2004, 2012) system and used to explore the difference between the two cloud-detection schemes in identifying cloud-free channels. This research aims to provide some guidelines for setting up cloud-detection schemes before data assimilation. The remainder of the paper is structured as follows. In section 2, a brief introduction to the two methods for detecting cloud is provided. Section 3 presents results comparing the MMR and MW03 schemes for simulated cloudy radiance. Real observation experiments and results are presented in section 4, before conclusions are drawn in section 5.

The Observing System Simulation Experiment (OSSE) method is regarded as an efficient way to test new data assimilation techniques ( Huang et al., 2013; Wang et al., 2013). Since cloud detection is a pre-processing step of IR sounder assimilation, cloud-detection schemes are first evaluated by the OSSE. The AIRS ( Aumann et al., 2003) is one of the most advanced sensors, the radiances of which are often used to test cloud-detection schemes ( McNally and Watts, 2003; Auligné, 2007). Following those studies, this study first compares the MW03 scheme and the MMR scheme based on simulated AIRS radiances.

Simulated cloud-contaminated AIRS radiances, , are obtained from the CRTM with the mixing ratios of rainwater, snow, graupel, and ice hydrometeors in a reference atmosphere. The reference atmosphere is generated by conducting a one-hour WRF ( Skamarock et al., 2008) forecast with the Lin microphysics scheme ( Lin et al., 1983) valid at 0100 UTC 3 June 2012 (at which time enough hydrometeors have been generated as the input for the CRTM to obtain simulated cloudy radiance). The domain for this forecast has a 30-km grid spacing on a

100 × 100 horizontal grid and 40 vertical levels with the model top at 30 hPa. Usually, the thresholds of the MW03 cloud-detection algorithm are empirically determined in default settings. We choose three different thresholds in the MW03 scheme, as listed in Table 1, for comparison. Experiments denoted as P4 have larger thresholds than those of P2. The default thresholds are set to smaller values than P2 and P4. The for 281 AIRS channels are constructed over a uniform network (10 000 model grid points all together).

A parallel procedure for clear-sky simulated radiance, , is also conducted with no hydrometeor information for the CRTM. By comparing the two groups of sim-ulated radiances, and , with and without hyd-rometeor information in the input profiles, it is easy to see whether or not a particular channel for a particular pixel is cloud-free. For this simulation case there are 5514 cloudy profiles among the total 10 000 profiles.

Figure 1 shows the used-data counts after cloud-det-ection procedures with different configurations. The first model level corresponds to the model top, while level 40 refers to the surface. The peak level, c_ , derived in Eq. (3) for each channel is also displayed (pixel averaged). For the MW03 scheme, with small threshold values in the default setting, the criteria specified in section 2.1 are hard to satisfy for most channels, resulting in too many data being mistakenly rejected. The default threshold values are conservative to reduce the possibility of cloud-contaminated radiances being used ( Collard et al., 2009). The two cloud-detection schemes produce consistent data rejection distributions overall for channels with wavelengths from 6.4 μm to 10.0 μm, when the MW03 thresholds are raised. MMR generally retains more data for channels with wavelengths longer than 12 μm.

Table 1

Table 1

Table 1 Configurations for the MW03 threshold sensitivity tests, where P2 standing for paremeters with BT_Threshold=2 and P4 for BT_Threshold=4. Default P2 P4 BT_Threshold 0.5 2 4 BT_Grad_Threshold 0.02 0.05 0.06

Table 1 Configurations for the MW03 threshold sensitivity tests, where P2 standing for paremeters with BT_Threshold=2 and P4 for BT_Threshold=4.

Figure 1

	Figure Option ViewDownloadNew Window
	Figure 1 Data counts (×1000) of cloud-clear data from the MMR scheme (black curve) and MW03 scheme with different thresholds (see Table 1). Peak levels, c_k, in Eq. (3) for selected channels are also shown (black dashed line).

Whether or not a particular channel for a particular pixel is cloud-free is determined by comparing and so the results are easy to verify. Figure 2a shows the percentages of pixels for each channel that are wrongly flagged as clear by the cloud-detection procedures. Note that the MW03 scheme's performance is easily affected by the specified thresholds (Threshold and Grad_Thre-shold), especially for channels from 6.4 μm to 10.0 μm, indicating that cloud-detection results are more sensitive to thresholds for higher channels than for lower channels. Smaller thresholds lead to a cloud signal detected at a higher altitude, thereby leading to MW03 rejecting more data. Likewise, Fig. 2b shows the percentages of pixels for each channel that are wrongly flagged as cloudy. With the default settings, the MW03 scheme rejects most of the data including "good" data (~ 60% more). The two cloud- detection schemes produce consistent data rejection distributions overall for channels with wavelengths from 6.4 μm to 10.0 μm, when the MW03 thresholds are raised (roug-hly eight times the default thresholds).

To further evaluate the two cloud-detection schemes for more complicated cases, the schemes are applied to real AIRS data onboard Aqua. We perform one-month cloud- detection experiments using MW03 and MMR from 1800 UTC 20 July 2010 to 1800 UTC 20 August 2010 with a 60-km thinning mesh for AIRS data. The domain for the experiments has a 36-km grid spacing on a 330 × 200 horizontal grid and 43 vertical levels with the model top at 30 hPa.

4.1 Data usage and rejection rate

Figure 3 shows the monthly-averaged used-data counts after cloud-detection procedures with different configurations.

Figure 2

	Figure Option ViewDownloadNew Window
	Figure 2 The percentage of pixels wrongly flagged (a) clear and (b) cloudy for each channel.

Figure 3

	Figure Option ViewDownloadNew Window
	Figure 3 Month-averaged data counts (×1000) of cloud-clear data from the MMR scheme (black curve) and MW03 scheme with different thresholds (see Table 1). Peak levels, c_k, in Eq. (3) for selected channels are also shown (black dashed line).

There is a high level of consistency between the results from real AIRS data and those from the simulated data. As expected, both cloud-detection schemes have good performances in finding the cloud-contaminated channels based on the channels' peak levels. The cloud-detection schemes reject more pixels for the lower-peaking channels than for the higher-peaking channels.

In general, MMR and MW03-P4 retain comparable clear data for most channels and MMR retains fewer data for short-wave channels than MW03 does. One possible explanation is that for the real case, MMR easily interprets the mix of multiple error sources (such as, the system bias in radiative transfer models and the solar contamination for short channels) as cloud signals are compared for the ideal simulation case. P2 can reject more data than P4, especially for channels with high peak levels. The MW03 scheme using small threshold values results in conservative observations.

Figure 4 shows the time series of the rejection rates (the rejected pixel numbers over the total pixel numbers) for two different channels. We did not adopt the MW03 scheme with default thresholds for further comparison because there were too few data left. Here, we choose channel 1776 corresponding to 6.48 μm (the top panel) to represent the higher channels (peak level of ~ 200 hPa) and channel 950 corresponding to 10.37 μm (the bottom panel) for lower channels (peak level of ~ 850 hPa). All cloud-detection schemes reject more pixels for the lower channel (Fig. 4b) than for the higher channel (Fig. 4a). For the lower channel 950 (10.37 µm), MMR rejects more pixels than MW03 and P4 retains a few more pixels than P2. For the higher channel 1776 (6.48 µm), P2 rejects many more pixels than both P4 and MMR, which is consistent with the results in Figs. 2 and 3.

4.2 Verification against Vis/NIR channels

The AIRS instrument also includes four visible/near-IR (Vis/NIR) channels between 0.40 and 0.94 µm with a footprint of 2.3 km at nadir, sharing the scan mirror with the IR channels. The primary function of the Vis/NIR channels is the diagnostics of cloud in the IR field of view (13.5 km footprint at nadir), which is used in this study to evaluate the performance of the two cloud-detection sch-emes. Figure 5a shows the distribution of cloud amount

Figure 4

	Figure Option ViewDownloadNew Window
	Figure 4 Rejection rates from the MMR scheme and MW03 scheme for (a) channel 1776 (6.48 μm) and (b) channel 950 (10.37 μm).

from the Vis/NIR channels (Fig. 5a). Figures 5b-g show the cloud-free radiances detected from MMR (two top panels), MW03-P2 (two middle panels), and MW03-P4 (two bottom panels) at 1800 UTC 20 July 2010. The three left panels (Figs. 5b, 5d, and 5f) correspond to the cloud- detection results for the higher channel 1776 (6.48 µm), while the right panels (Figs. 5c, 5e, and 5g) are for the lower channel 950 (10.37 µm). In general, the cloud-free radiance distributions from all cloud-detection schemes are in good agreement with the cloud masks in Fig. 5a for both channels (for example, the regions in the black circles). Fewer pixels are clear for the lower channel 950 (10.37 µm) than for the higher channel 1776 (6.48 µm). To conclude, both cloud-detection schemes reasonably identify channels as clear or cloudy according to their peak levels. Generally, P4 obtains similar results to the MMR scheme.

The MW03 cloud-detection scheme codes developed by ECMWF for high-spectral-resolution IR sounders are incorporated into the WRFDA system in this study. The default values for the two thresholds in MW03 to search cloud signals are set quite small in ECMWF to reduce the possibility of cloud-contaminated radiances. This study provides some guidelines for setting up MW03 by comparing its results with those from MMR for AIRS data.

There is high consistency between the results from real AIRS data and those from the simulated data. Both cloud-detection schemes agree closely in the distribution patterns of clear pixels based on the channels' peak levels. The cloud-detection results from MW03 are sensitive to the prescribed brightness temperature innovation threshold and the brightness temperature gradient threshold, especially for higher channels. Smaller thresholds lead to

Figure 5

	Figure Option ViewDownloadNew Window
	Figure 5 (a) The distribution of cloud amount (color points) from AIRS Vis/NIR channels and distributions of cloud-free pixels left after cloud detections (black dots): (b) channel 1776 from MMR; (c) channel 950 from MMR; (d) channel 1776 from P2; (e) channel 950 from P2; (f) channel 1776 from P4; and (g) channel 950 from P4.

a cloud signal detected at a higher altitude, thereby rejec-ting more data. In general, MMR and the MW03 scheme, with empirically determined parameters (MW03-P4 in Table 1), can yield comparable results. For real AIRS data, it is found that results from both schemes are generally in better agreement with the visible/near-IR data.

Further work is planned on adopting adaptive thresholds through offline statistics generated over some spin- up periods for the MW03 scheme. More numerical experiments on comparing the two cloud-detection schemes will also be further investigated in the framework of data assimilation and weather forecasting. Other applications based on the MMR cloud-detection algorithm in terms of the cloud fractions retrieval technique is another interesting aspect in future work.

Acknowledgments. This work was jointly sponsored by the National Basic Research Program of China (973 Program, 2013CB-430102) and the Program of Scientific Innovation Research of College Graduate in Jiangsu Province (Grant Nos. CXZZ12-0490 and CXZZ11-0606). The authors would like to thank Dr. Junmei BAN for fruitful discussions on editing the manuscript. The National Center for Atmospheric Research is sponsored by the National Science Foundation. Any opinions, findings, and conclusions or recommendations expressed in this publication are those of the authors and do not necessarily reflect the views of the National Science Foundation.