Detection and characterization of drizzle cells within marine stratocumulus using AMSR-E 89 GHz passive microwave measurements

Introduction Conclusions References


Introduction
Marine stratocumulus clouds cover vast regions of the eastern subtropical oceans and are a source of net cooling in the earth's radiation budget (Hartmann et al., 1992). The radiative characteristics of marine stratocumulus are closely tied to cloud frac-25 tion which differs dramatically between regions of continuous unbroken cloud versus 4572 der and Vonder Haar, 1995). The emission by liquid water at 89 GHz has been used previously in the retrieval of liquid water path (LWP) when cloud ice or precipitation ice are not present. Petty (1994, his Eq. 9) describes a method using 85.5 GHz data from Special Sensor Microwave Imager (SSM/I) to estimate LWP as a function of the normalized polarization difference (Petty andKatsaros, 1990, 1992). This method was 15 devised for general precipitation cases and was specific to SSM/I sensors. The Petty method also required an estimation of hypothetical clear-sky brightness temperature values. The calculation of clear-sky values can be challenging for regions of broad cloudiness, like stratocumulus regions, without relying on climatology data or data from multiple sensor platforms. Crewell and Lohnert (2003) used a vertically pointing (up-20 ward) 90 GHz radiometer data to improve the accuracy of retrieved LWP of cloud liquid water and drizzle.
In the absence of ice, satellite-observed 89 GHz brightness temperatures are the net result of column-integrated emission and scattering of upwelling radiation by the oceansurface, water vapor, and liquid hydrometeors. The variances in observed brightness 25 temperature are primarily a function of the background ocean-surface emission (primarily related to sea surface temperature (SST) and wind speed), gas absorption, water vapor profile, beam filling, and LWP, including cloud and precipitable water (Westwater et al., 2001;Crewell and Lohnert, 2003;Greenwald et al., 2007;Horváth and 4573 limitations. The key limitation of the Moderate Resolution Imaging Spectroradiometer (MODIS) Cloud Liquid Water Path product (King et al., 2003) is that it requires information from MODIS visible wavelength channels and is not available at night. For the purposes of this study, we take the MODIS Cloud Water Path as LWP. For stratocumulus observations, the two parameters should be similar and we use the terms 10 interchangeably. While drizzle does occur in marine stratocumulus regions during daylight hours, more frequent and intense marine stratocumulus drizzle typically occurs at night when LWP values are higher (Bretherton et al., 2004;Comstock et al., 2005;Sharon et al., 2006). Any method attempting to ascertain the regional characteristics of drizzle while excluding nighttime drizzle would miss the mode of the drizzle cell 15 distribution and hence would not produce a robust representation of regional drizzle characteristics. Current passive microwave LWP products Meissner, 2000, 2004;Kida et al., 2009) have too coarse a spatial resolution (14 × 8 km) to adequately resolve small and intense drizzle cells, which are typically 1-10 km in horizontal dimension. The CloudSat radar (Stephens et al., 2002(Stephens et al., , 2008Haynes and Stephens, 20 2007) has a minimum sensitivity of −28 dBZ and provides information on the vertical structure of clouds. CloudSat observations of cloud-top heights and the profile of reflectivity are important physical characteristics of marine stratocumulus. However, CloudSat's swath width of 1.4 km is not adequate to obtain information on the horizontal mesoscale organization of drizzle. The Tropical Rainfall Measuring Mission (TRMM) 25 Satellite's Precipitation Radar can only detect the very strongest drizzle cells that have radar reflectivities above the radar's minimum sensitivity of ≈17 dBZ at 5 × 5 km spatial resolution (Kummerow et al., 1998 The goal of this study is to use empirical data to demonstrate the feasibility of using 89 GHz Advanced Microwave Scanning Radiometer -Earth Observing System (AMSR-E) passive microwave brightness temperature data to detect heavily drizzling cells within marine stratocumulus. A binary heavy drizzle classification product is described that can be used to determine areal and feature statistics of drizzle cells within 5 major marine stratocumulus regions.

Motivation
The Variability of the American Monsoon System (VAMOS) Ocean-Cloud Atmospheric-Land Study (VOCALS) Regional Experiment (REx) (Wood et al., 2011b), which took place in the southeastern Pacific Ocean during October and November 2008, provided an opportunity for the intercomparison of satellite and surface-based measurements of marine stratocumulus clouds and drizzle. The C-band wavelength radar on the National Oceanic and Atmospheric Administration (NOAA) ship Ronald H. Brown scanned a 120 km diameter circle centered on the ship with a minimum sensitivity of radar re-15 flectivity ≈0 dBZ. The mesoscale structures of heavy drizzle in the region scanned by the radar can be compared to those retrieved by various satellite algorithms. Figure 1 shows collocated observations near 21 • S and 85 • W within the southeast Pacific stratocumulus region from MODIS LWP, AMSR-E LWP Meissner, 2000, 2004), AMSR-E 89 GHz horizontally polarized (H) brightness temperature (T 89 H ), 20 and NOAA ship radar's C-band radar reflectivity. The horizontally polarized 89 GHz brightness temperature channel is used in this study because it contains less noise than the vertically polarized channel. In Fig. 1d LWP values and gradients at the cell edges. The heavy drizzle classification algorithm described in this paper stems from the observations in Fig. 1c. Drizzle cells corresponding to those observed by the C-band radar are evident as distinct local maxima in T 89 H (Fig. 1c). Since strong drizzle cells with Z > 0 dBZ are usually between 1 km and 10 km in scale, these features are more distinct in the finer resolution T 89 H data 5 at 6 × 4 km spatial resolution compared with the AMSR-E LWP product at 14 × 8 km.
The disadvantage of an 89 GHz based method is that it will not work where clouds containing ice are present. The inversion-topped boundary layers in the subtropical regions constrain marine stratocumulus cloud top altitudes to well below the 0 • C level, precluding the existence of drizzling clouds containing ice, and making these clouds a 10 prime candidate for the exploitation of 89 GHz emission. Within marine stratocumulus regions characterized by strong inversions, any middle level or high level ice clouds can be easily distinguished from liquid-phase clouds by their cloud top temperatures.

Binary heavy drizzle cell classification algorithm
For the purpose of this paper, we define heavily drizzling cells as areas ≥6 × 4 km with 15 C-band Z > 0 dBZ. C-band Z > 0 dBZ is approximately equivalent to LWP ≥ 200 g m −2 (Wood et al., 2011a). This definition corresponds to the subset of drizzle features within marine stratocumulus with sufficiently high drizzle intensity that precipitation usually reaches the surface (Comstock et al., 2004). A variety of independent studies suggest that LWP ≥ 200 g m −2 is a reasonable threshold to identify heavy drizzle. Ship-20 based upward-looking microwave radiometer data from the EPIC experiment in 2001 (Bretherton et al., 2004) show that there is a dramatic increase in the frequency of occurrence of heavy drizzle (>0 dBZ) for LWP values ≥ 200 g m −2 (Zuidema et al., 2005).
Fifty percent of drizzle and 90 % of heavy drizzle occurred when LWP was ≥200 g m −2 (Zuidema et al., 2005) within marine stratocumulus and its interrelations with changes in cloud fraction (de Szoeke et al., 2010). We sidestep the complexities of evaluating drizzle based on a quantitative estimation of LWP by instead creating a binary heavy drizzle detection product based on a background-adaptive threshold of T 89 H . Unlike the global MODIS LWP and AMSR-E 10 LWP products, our heavy drizzle product is only applicable to geographic areas with strong inversions and persistent marine stratocumulus clouds. Our binary drizzle detection product complements, rather than replaces, MODIS LWP and AMSR-E LWP algorithms. 15 Marine stratocumulus clouds can cover areas as large as 30

Drizzle detection against a variable background
• of latitude and 25 • of longitude. One would not expect the background clear-sky brightness temperature to be uniform across such large areas or among the marine stratocumulus regions (e.g. southeast Pacific, southeast Atlantic, northeast Pacific, and northeast Atlantic). An adaptive method is needed for determining the background, cloud-free, brightness 20 temperatures. Figure 2a shows the strong correlation between AMSR-E IWV (Wentz and Meissner, 2004) and AMSR-E T 89 H for a data set of more than 285 000 cloud-free AMSR-E pixels randomly distributed in space and time between 60 • N and 60 • S latitude. Cloud-free pixels were defined as pixels with a cloud fraction <0.01 for MODIS MYD06 L2 data 25 interpolated to match the resolution of the AMSR-E 89 GHz brightness temperature product. According to Petty's (1994) Eqs. (5e-g), much of the remaining variation in AMTD 5,2012 Detection and characterization of drizzle cells T 89 H -that is, the range of T 89 H values for a given IWV value -can be explained as a function of wind speed. While wind speed is important to precisely estimate cloudfree brightness temperature, we found that including wind data as part of our detection algorithm produced a negligible improvement, and we did not include it in our method. Liquid phase cloud, drizzle, and cloud-free pixels are all included in the scatter den-5 sity plot of AMSR-E IWV versus AMSR-E T 89 H data (Fig. 2b) obtained within marine stratocumulus regions in the southeast Pacific, southeast Atlantic, and northeast Pacific mostly during months of peak stratocumulus frequencies: September and October for the Southern Hemisphere regions and June for the northeast Pacific. The most noticeable difference between the clear-sky and marine stratocumulus distributions is that 10 the one for the stratocumulus regions is "two-pronged". One branch of the distribution is similar to that of the cloud-free pixels in Fig. 2a. The other branch has higher values of T 89 H for a given IWV value and corresponds to cloudy and drizzling pixels. The number of pixels with AMSR-E LWP ≥ 200 g m −2 for each pair of T 89 H and IWV values is shown as contours on top of the shaded scatter density plot. The relative values and shape of 15 the contoured distribution are more important than the absolute magnitudes of the contours, which are a function of the size of this particular data set (≈4.4 × 10 6 pixels). Of particular note is the local maximum in number of pixels with LWP ≥ 200 g m −2 located at IWV ≈ 15 g m −2 and T 89 H ≈ 245 K. Drizzling marine stratocumulus will emit more radiant energy and have a higher T 89 H 20 than nearby cloud-free and cloudy non-drizzling areas. Based on this physical principle, we can estimate an offset in T 89 H as a function of IWV that separates drizzling from non-drizzling pixels. We start with an empirically-derived curve fit between the cloud-free T 89 H and IWV values in Fig. 2a curve is somewhat subjective. Our goal was to find a curve that worked consistently among different marine stratocumulus geographic regions and would err on the side of underestimating the number of heavy drizzle pixels. The resulting empirically derived T 89 H threshold for classifying an AMSR-E pixel as drizzle or non-drizzle as a function of IWV is:

Algorithm application
The inputs to the algorithm are AMSR-E 89 GHz brightness temperatures, AMSR-10 E IWV, MODIS cloud top temperature (MYD06 L2), and AMSR-E SST. All products are interpolated to match the 6 × 4 km AMSR-E 89 GHz pixels. In the first step, pixels are removed from consideration based on cloud top temperature and SST. Pixels with cloud-top temperatures less than 273 K were classified as non-drizzle to remove icephase cloud. Pixels with SST values less than 14 • C and greater than 30 • C are also 15 classified as non-drizzle to reduce contamination from non-stratocumulus features that can be observed in scenes along the boundaries of stratocumulus regions. For the remaining pixels, the IWV value is converted to T 89 H, threshold using Eq. (1). If the pixel's T 89 H is greater than the T 89 H, threshold for that pixel, it is identified as heavy drizzle. The identification of discrete contiguous areas of detected drizzle is done using standard 20 connected component algorithms for raster data (Rosenfeld and Pfaltz, 1966;Haralick and Shapiro, 1992). Once the discreet drizzle features are identified, several statistics can be calculated, including, but not limited to, cell area, aspect ratio, orientation, and frequency of occurrence per unit area.

Caveats
The algorithm was designed with the intention of identifying heavy drizzle in marine stratocumulus regions with large scale subsidence and shallow boundary layers such as those found off the western coasts of the Americas, Africa, and the Canary Islands. The algorithm was not intended for drizzle classification in thermally driven areas of 5 stratocumulus (cold air outbreaks) more common to high latitude oceans, nor has it been tested in these areas. There is a potential for misclassification of drizzle in cases where the environmental properties lie outside the implicit climatology used in the empirical derivation. Anomalously high winds or atypical cloud properties could yield a skewed relationship between IWV and T 89 H , which would lead to misclassification.
However, the climatology of marine stratocumulus regions is such that these cases should be rare occurrences. Classification error due to beam filling is an obvious concern and a key reason why we chose to err on the side of underestimating heavy drizzle areas. The ship C-band radar data during VOCALS show that most drizzle features with Z > 0 dBZ have an 15 area of approximately one T 89 H pixel (6 × 4 km) or larger. Based on EPIC data, Comstock et al. (2007) found that heavy drizzle events coincide with the presence of drizzle cells >20 km 2 . Cloud and radiation model simulations by Lafont and Guillemet (2004) suggest that 89 GHz brightness temperatures values for stratiform clouds decrease as sub-pixel cloud fraction decreases. The effect increases as a function of LWP since 20 higher LWP clouds contribute more to the observed brightness temperature than low LWP clouds. Cloud homogeneity and 3-D beam effects can also impact 89 GHz brightness temperature estimations. The Lafont and Guillemet (2004) study suggests that heavy drizzle is most likely to be missed by our algorithm when the drizzle features have an area ≈80 % or less than the 6 × 4 km footprint of the T 89 H data. 5,2012 Detection and characterization of drizzle cells

Comparison with VOCALS ship-based radar data
Drizzle detection from twelve AMSR-E overpasses is compared with coincident C-band radar data from the NOAA ship Ronald H. Brown collected during the VOCALS project (Fig. 3). The radar data are overlaid atop the binary heavy drizzle classification prod-5 uct. In each case, the drizzle classification algorithm captures the mesoscale structure of the drizzle observed. These twelve images display all the AMSR-E and NOAA ship radar coincidences where drizzle was present in the ship radar volume during the VO-CALS campaign. Some of the ship radar data obtained near the South American coast were contaminated with second-trip echoes and are excluded.

Comparison among satellite algorithms
We compare our binary heavy drizzle detection algorithm to the well-established MODIS LWP and AMSR-E LWP products in daylight scenes containing marine stratocumulus from the Southeast Pacific, the Southeast Atlantic, and the Northeast Pacific (Figs. 4-7 and Table 1 higher spatial resolution MODIS LWP assuming that individual drizzle elements have an area equal to or larger than the 6 × 4 km pixel size of the 89 GHz channel. Microwave and optical methods of estimating LWP generally agree with each other and perform well for marine stratocumulus regions (Lin and Rossow, 1994;Borg and Bennartz, 2007). More specifically, Seethala and Horváth (2010) state that AMSR-E 5 and MODIS estimations of LWP agree best for stratocumulus regions and for other overcast regions when cloud fraction values are greater than 0.9. Within high cloud fraction conditions of marine stratocumulus regions, they found correlations up to 0.95 between AMSR-E LWP and MODIS LWP. For lower cloud fraction values, AMSR-E overestimates domain-mean LWP compared to MODIS. The changes in relative perfor-10 mance as a function of cloud fraction are likely a result of 3-D cloud effects, differences in the spatial resolution of the two products, and beam filling uncertainties.
Co-registered 89 GHz brightness temperature data, our 89 GHz based binary detection product, AMSR-E LWP, and MODIS LWP data are shown in Figs. 5-7. Figure 5 shows an area over the Southeast Pacific on 27 October 2008. Satellite data from a 15 scene with drizzle over the Southeast Atlantic off the coast of Africa for 26 June 2007 are shown in Fig. 6. There are some high, mostly transparent, cirrus clouds in this scene with cloud-top temperatures <273 K in a southwest to northeast strip at approximately 23 • S latitude. The marine stratocumulus common to this region have never been the subject of a major field observation campaign. Satellite data for an area over 20 the Northeast Pacific for 29 July 2009 that contains drizzle are shown in Fig. 7. This scene is a complex example with a strong SST gradient. Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | spatial distribution of the drizzle. This mesoscale spatial structure of heavy drizzle is an important component in understanding the underlying physical processes (Comstock et al., 2007). Recall from Sect. 1, that the primary advantage of our algorithm is that it provides higher spatial resolution information than AMSR-E LWP at night, when drizzle is more intense and frequent, and when the MODIS LWP product is not available.

5
Another artifact of the difference in spatial resolution between AMSR-E 89 GHz and MODIS LWP is that the coarser resolution 89 GHz data will tend to slightly shift the size distribution of contiguous areas of drizzle toward larger areas. The impact of this artifact is illustrated in Fig. 8, which shows a histogram of discrete drizzle cell area from our 89 GHz-based algorithm and MODIS LWP for the example scene from the Southeast Pacific in Fig. 5. Although the total area identified by the 89 GHz-based algorithm as drizzle is less than that from MODIS LWP, our algorithm detects contiguous drizzle cells >400 km 2 which are not present in the MODIS LWP data. Beam filling causes our algorithm to overestimate the size of the largest contiguous cells. This shifting of the size distribution has to be taken into account when examining results from our 15 algorithm. Table 1 summarizes the drizzle area statistics for the different products in each scene. As expected, our binary heavy drizzle detection product identifies more drizzle area (∼50 % more) than the 200 g m −2 threshold applied to AMSR-E LWP. The 200 g m −2 threshold applied to MODIS LWP yields the largest drizzle areas with the 20 exception of the Southeast Atlantic scene where the 89 GHz based binary heavy drizzle product identified ≈7 % more drizzle area. This discrepancy is likely due to sizes of the drizzle elements. The MODIS data panels in Fig. 5 illustrate that the drizzle elements are very small. We theorize that drizzle features that are a cluster of cells each smaller than the 6 × 4 km 89 GHz pixel size are detected as one large mass. The result 25 of this is that when there are many drizzle elements smaller than the 89 GHz pixel size, overestimation of area is possible with respect to MODIS data since the space between the small elements is also flagged as drizzle as part of an error due to resolution limitations.

Conclusions
Drizzle plays a key role in the evolution of marine stratocumulus and the transitions between closed-and open-cell states. The binary heavy drizzle classification product based on AMSR-E data described in this paper represents a new approach that provides additional information on the frequency of occurrence and spatial characteristics 5 of drizzle. Current satellite methods are either lacking in resolution (AMSR-E LWP), night coverage (MODIS LWP), or the broad areal coverage (CloudSat) necessary to observe the horizontal mesoscale structure of drizzling marine stratocumulus. Use of high-frequency passive microwave observations to detect drizzle in marine stratocumulus environments enables consistent observations at resolutions sufficient for resolving 10 individual heavily drizzling cells and their mesoscale structure. Emission by precipitation liquid water from drizzle cells in marine stratocumulus regions yields local maxima in brightness temperature in AMSR-E 89 GHz data. Colocated AMSR-E IWV data are used to determine an 89 GHz background brightness temperature threshold to which these local maxima can be compared. Once drizzle 15 is identified, standard connected-component algorithms are used to identify discrete drizzle cells and calculate various spatial statistics.
Even without an accompanying quantitative estimate of drizzle LWP, the satellitebased binary identification of heavy drizzle cells within marine stratocumulus regions will permit analysis of seasonal and regional drizzle cell occurrence and the interrela-20 tion between drizzle and changes in cloud fraction. Several characteristics of contiguous drizzle cell features can be documented: the number of drizzle cells per unit area, their sizes and shape, and the distances between cells. In ongoing work, we are using these properties as comparison metrics among the different marine stratocumulus regions. The information on marine stratocumulus drizzle from the 89 GHz based drizzle Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Lafont, D. and Guillemet, B.: Subpixel fractional cloud cover and inhomogeneity effects on microwave beam-filling error, Atmos. Res., 72, 149-168, doi:10.1016/j.atmosres.2004.03.013, 2004 Fig. 8. Histograms comparing drizzle area calculated from the 89 GHz based binary heavy drizzle classification product, AMSR-E LWP, and MODIS LWP for the time and region from Fig. 6. The upper panel is the area histogram for contiguous drizzle pockets detected using 89 GHz brightness temperature data. The middle panel is the area histogram for contiguous drizzle pockets defined as contiguous areas with AMSR-E LWP ≥ 200 g m −2 . The lower panel is the area histogram for contiguous drizzle pockets, defined as contiguous areas with MODIS LWP ≥ 200 g m −2 . The total area identified as drizzle is given for each product. Contiguous areas are identified based on 4-connectivity.