1. Introduction

Global measurement of ice precipitation is needed to fully understand the water cycle and climate. In particular, knowing its vertical profiles is important in studies of latent heating and energy transport. However, because of sparse observation stations in high-latitude regions, little knowledge has been accumulated about the global distribution of snowfall or ice precipitation. Measurement from space is the most promising way to change the state of lacking sufficient observation data.

Measurements of ice precipitation from space have been carried out in the past with microwave radiometers and the W-band cloud profiling radar (CPR) of CloudSat (Hiley et al. 2011; Kulie et al. 2016; Liu 2008; Kulie and Bennartz 2009; Skofronick-Jackson et al. 2013). The CloudSat radar has the highest sensitivity for snow detection and brought us a reliable observational distribution of falling snow globally for the first time. Detection of ice precipitation with a spaceborne radiometer relies on the scattering signal at high-frequency channels. Combining microwave radiometer data with CloudSat data has accelerated the improvement of ice retrieval algorithms for microwave radiometers substantially (Liu and Seo 2013). Nevertheless, there still remains large uncertainty in the retrieval algorithms for both radar and radiometer to detect and quantify ice precipitation with sufficient accuracy. A major part of uncertainty originates in large variation of the scattering properties of ice particles that depend heavily on the kind of ice particles and the habit of snow crystals. In addition, in the CloudSat case, the effect of non-Rayleigh scattering and multiple scattering by ice particles makes the quantitative retrieval of precipitation rate very difficult in intense ice precipitation. The CloudSat’s very narrow swath and sun-synchronous orbit limit the sampling representativeness as well.

One of the objectives of the Global Precipitation Measurement (GPM) mission was to detect falling snow with both the GPM microwave imager (GMI) and the dual-frequency precipitation radar (DPR) on its Core Observatory spacecraft (Hou et al. 2014; Skofronick-Jackson et al. 2017). DPR consists of the Ku-band radar (KuPR) and the Ka-band radar (KaPR). KuPR is a radar very similar to the precipitation radar (PR) on the Tropical Rainfall Measuring Mission’s (TRMM’s) satellite. KaPR was added to increase the sensitivity of precipitation measurement and to estimate the particle size more accurately than a single-frequency radar so that we can improve the estimates of rainfall rate and identify snow precipitation regions. In fact, by using the difference in the scattering and attenuation properties of liquid and solid water particles between Ku- and Ka-band electromagnetic (EM) waves, it is possible to estimate the mean diameter of precipitation particles once an appropriate particle size distribution (PSD) model is chosen. Since the mean particle size for a given radar reflectivity differs substantially between liquid phase and solid phase particles, it is possible in principle to separate solid phase regions from liquid phase regions.

In a typical stratiform rain system, a bright band is observed in radar echoes so that we can easily separate the solid precipitation range from the liquid precipitation range. In a convective system, however, no bright band is observed. We need to take advantage of the difference in radar echoes of rain and snow between the Ku and Ka bands. One possible way to achieve this objective is to use the size information. By combining radar echoes from the common scattering volume at two different frequencies, we should be able to estimate two parameters of the PSD of precipitation. Once two independent parameters of PSD are obtained, we can not only convert radar reflectivity factor into rainfall rate accurately but also identify and separate solid precipitation regions from liquid precipitation regions.

In fact, if the attenuation correction is properly realized, the difference in effective radar reflectivity factor at the two different frequencies, which is called the dual-frequency ratio (DFR), gives particle size information because the size effect caused by non-Rayleigh scattering takes place at the higher-frequency radar echoes more severely than at the lower-frequency radar echoes. As a measure of the difference in effective radar reflectivity factor at the two different frequencies, we define DFR in this paper as follows:

View Expanded

where is expressed in decibels of reflectivity factor (dBZ) and the argument in the parentheses indicates the frequency band. Since ice particles and snowflakes are generally larger than raindrops, DFR becomes much larger for ice or snow precipitation than rain for a given value of (Ku). For example, if a typical particle size distribution is assumed, when (Ku) dBZ, then DFR from rain echo is about 0 to dB, whereas DFR from ice precipitation takes a value larger than a few decibels (3–10 dB) depending on the density of ice particles (Liao and Meneghini 2011). Therefore, for light to moderate precipitation, DFR can be used to identify the phase state of precipitating particles.

It is known that attenuation resulting from absorption by ice particles is very small at both Ka and Ku bands, but the Ka-band EM waves suffer from attenuation by liquid phase precipitation much more severely than the Ku-band EM waves. We define the measured DFR (DFR_m) as

View Expanded

where is the measured apparent radar reflectivity factor. If we observe light to moderate precipitation from above with a dual-frequency radar, then DFR_m takes a value close to DFR in the ice precipitation range because the attenuation effect is small. However, it may recover from a finite value to a value close to zero at the top of the rain range because light rain can produce only a small value of DFR. Then DFR_m increases in rain with range as a result of the attenuation difference between the Ka- and Ku-band channels that increases monotonically with range.

After the GPM Core Observatory satellite was launched, however, it became apparent that combining the dual-frequency precipitation echoes to extract PSD information is much more difficult than we anticipated. To obtain accurate DFR from measured radar echo data, we need to correct for the attenuation in and , but the accurate attenuation correction turned out to be very difficult. This is because, first, we need to know the phase state of the precipitating particles (solid or liquid) to calculate the attenuation from at each range bin, but the phase state can be estimated from DFR only after we make the attenuation correction. Thus, we may end up with a circular algorithm that may not correctly estimate the phase state. Theoretically, we can calculate the attenuation by assuming either a liquid or solid phase of the particles at each range bin, and choose the one that produces a more natural or reasonable solution. In reality, however, small errors in the estimates and fluctuation of signals make it very difficult to judge whether an increase in DFR_m in a particular range is caused by the size effect or by attenuation and to determine the phase state.

The second reason for difficulty in attenuation correction is that some assumptions in the simple radar equation used in the attenuation correction are not necessarily satisfied, especially in measurement of an intense convective storm. The most important assumption of this kind is the assumption of uniformity of precipitation within the scattering volume. Since typical horizontal dimensions of a convective system are smaller than or comparable to the radar beam diameter, uniformity of precipitation within the scattering volume may not be an appropriate assumption. Depending on the degree of nonuniformity and the spatial correlation of the distribution of precipitating particles between different range bins, the effective attenuation may change drastically even if the measured profiles are identical. The attenuation estimate given by the surface reference technique may also be severely biased if the rain distribution is not uniform within the beam.

In a very intense convective precipitation system with a substantial amount of ice aloft, multiple scattering by ice particles may further complicate the problem. Since the scattering cross section is larger at the Ka band than at the Ku band, the multiple-scattering effect appears first in Ka-band echoes, and the multiple scattered echoes often appear at ranges that are not the true ranges from the radar to the volumes where the multiple scattering dominantly takes place because of complicated echo paths of multiple scattering. Attenuation by cloud liquid water is also another factor of uncertainty.

The uncertainty factors mentioned above make it very difficult to estimate accurately from by attenuation correction at the two frequencies. Since the estimation of PSD needs an accurate estimate of DFR, small errors in attenuation correction have a fatal effect.

Given such a difficult situation, we decided to develop a simple algorithm that flags the pixels in two-dimensional space that contain relatively intense ice precipitation somewhere in the vertical column with certainty. The algorithm uses only the magnitudes of measured dual-frequency ratio (DFR_m) and radar reflectivity factors together with ancillary temperature data. The algorithm is not intended to identify all solid precipitation regions, but to detect intense ice precipitation that includes intense snow precipitation. Even if the algorithm misses many light ice precipitation, information about the distribution of intense ice precipitation will complement the ice measurements by microwave radiometers or cloud profiling radar like CloudSat and should be of use to understand the global climate related to ice precipitation. Just as CloudSat data helped the improvement of ice retrieval algorithms for microwave radiometer, The new ice flag from DPR is expected to help improving microwave radiometer’s retrievals of precipitation that contains intense convection. The flag created by this algorithm is called flagHeavyIcePrecip. Note that there is another new flag denoted by flagSurfaceSnowfall in the DPR V5 output that indicates falling snow at or near surface (Le et al. 2017). This paper describes the algorithm that creates the former flag and shows some examples of its performances.

2. Algorithm

KuPR scans its 49 beams in the cross-track direction with the swath width of 250 km. KaPR’s matched scan consists of 25 beams that match with the central 25 beams of KuPR (Kubota et al. 2014; Iguchi et al. 2017). We call this narrow central swath the inner swath. The swath of KuPR outside the inner swath is called the outer swath. The new algorithm flags those beams in which intense ice precipitation is detected above the C isotherm height. DFR_m is used in the inner swath where dual-frequency information is available, whereas in the outer swath of the DPR products and in the Ku-only or Ka-only products, measured radar reflectivity factors () are used to set the flag. The flags from Ku-only and Ka-only judgment are also stored in the inner swath of the DPR products.

The dual-frequency algorithm for detecting intense ice precipitation is based on the property that the dual-frequency ratio (DFR) becomes rather large when precipitation particles are large. Size effect makes smaller than always for ice particles and for medium to large liquid particles. Matrosov (1998) shows that DFR at Ka and X band (or other longer wavelengths such as C or S band) is nearly independent of the density of snow particles, but only depends on its median diameter (Liao and Meneghini 2011). Similar argument is also applicable to the combination of the Ka and Ku bands. In fact, the DFR caused by ice particles may exceed a few dB when the median or mean particle size is large (Liao et al. 2005). In particular, aggregates of snowflakes create very large DFRs (Tyynelä and Chandrasekar 2014). Since the ice precipitation particles are generally larger than raindrops, we can expect that DFRs are larger in ice precipitating regions than in raining regions (Liao and Meneghini 2011).

We decided to use DFR_m instead of DFR itself in order to avoid errors originating from unreliable attenuation correction. As a result, DFR_m may take a large value not only when precipitation particles are large but also when the attenuation in Ka-band echoes becomes significant. The latter situation may happen, for example, in rainfall regions, where decreases with range more rapidly than by attenuation. In solid precipitation layers, however, attenuation by absorption can be ignored because the imaginary part of dielectric constant of ice is negligibly small at both Ku and Ka bands. Attenuation by scattering can generally be ignored at the Ku band, whereas that at the Ka band may become nonnegligible and decrease the apparent reflectivity when ice precipitation is intense. This attenuation effect may further increase the value of DFR_m in addition to the size effect. If the ice precipitation is very intense, then multiple scattering at the Ka band may compensate for the attenuation effect and sometimes decrease DFR_m at far ranges to produce a so-called knee in the DFR_m profile (Battaglia et al. 2015). In such a case, because of large attenuation, DFR_m takes a rather large value at a range well before the multiple-scattering effect appears. Since we look only at the maximum DFR_m value in each profile, the flag is set because of the large DFR_m above the height where the knee appears.

Since spaceborne radar measurements begin at the storm top, it can be assumed that solid (ice) hydrometeors, if present, will generally occur at ranges closer to the radar than liquid (water) hydrometeors. Above the height of C, we can assume that large precipitation particles are solid and that only small liquid drops may exist in a supercooled state (Williams et al. 1989). The total attenuation caused by small liquid drops that include water cloud droplets is very difficult to estimate, but it is unlikely that the Ka-band attenuation down to the C height exceeds 3 dB (Liebe et al. 1989). Considering the size effect and the attenuation effect, therefore, if a DFR_m larger than approximately 5 dB is measured above the C isotherm, it must be caused by large ice particles or a thick layer of intense ice precipitation.

In the current case, we set the threshold to 7 dB in order to be on the safe side. This threshold value of DFR_m is determined by referring to Figs. 1 and 4 in Liao and Meneghini (2011) and by some trial and error with actual DPR data.

When the radar reflectivity factor is calculated from the received power, the background noise power is subtracted from the received power. Since these sampled powers suffer from fading noise, the calculated includes a relatively large fluctuation error when the received power is very close to the noise power. If (Ka) is negatively biased in such a situation, then the calculated DFR_m may be positively biased substantially. To ensure that DFR_m larger than 7 dB can be judged with certainty even when the KaPR’s received power is close to the noise level, we impose another condition that (Ku) must be larger than 27 dBZ at the same range bin in which the corresponding DFR_m is evaluated. KaPR’s relatively low sensitivity limits the detectability of ice precipitation with the DFR_m method. If the high-sensitivity beams of KaPR are matched with the KuPR beams (which is envisioned in future operations), since its minimum detectable is about 6 dB better than the current KaPR’s matched beams, then we should be able to lower the threshold for KuPR accordingly and to improve the identification of ice precipitation regions.

Note that in order to avoid erroneous flagging associated with large DFR_m caused by attenuation as a result of the liquid phase of precipitation, only data from the storm top to C height are looked at, and that if the abovementioned conditions are satisfied at any range bin within the interval, then the flag is set for the angle bin that contains this profile.

Heavy ice precipitation flags are provided for the Ku-only and Ka-only products and in the outer swath of the DPR products as well. Since only single-frequency radar echoes are available in these cases, the flag is set by looking at the maximum measured radar reflectivity above the C height based on the following reasoning.

It is known that intense precipitation is associated with formation of ice particles created in a strong updraft in a convective cell. In particular, thunderstorms and lightning are caused by ice particles. Such intense storms are judged by large reflectivity factors resulting from relatively large ice particles at altitudes above the freezing height. It is generally accepted that 40 dBZ of reflectivity at the C height is a good threshold for detecting lightning activities (Williams et al. 1989; Zipser and Lutz 1994; Liu et al. 2012). We adopt this criterion for flagging the existence of intense ice precipitation in the Ku-only single-frequency algorithm. The same height is selected as the lower bound in the DFR_m method in the DPR algorithm as well. It is known, however, that the relationship between the magnitude of the radar reflectivity factor and the lightning activity is rather complicated and differs substantially between oceanic storms and continental storms (Petersen et al. 1996). To deal with such variation and uncertainty, different values are set in the flag at 40 ± 5-dBZ thresholds for KuPR profiles as well.

Since the Ka-band reflectivity is generally a few decibels lower than the corresponding Ku-band reflectivity because of the size effect from large particles and the attenuation caused by scattering in an intense storm, the threshold levels for ice detection with the KaPR are lowered by 5 dB from the corresponding Ku-band thresholds.

With these properties in mind, we adopt the following conditions to set a flag called flagHeavyIcePrecip that is added to the output of version 5 of the DPR products. The adjective heavy used in the name of the flag means intense precipitation and not heavy or dense ice particles. The actual code was implemented in the classification module of the DPR, Ku-only, and Ka-only algorithms.

The flag uses 5 bits and takes values between and in a binary expression. We first define , , and according to the following rules by examining the data in each profile above the C isothermal height.

1. Condition on DFR_m and at the same range bin

   

   

   View Expanded

   
2. Condition on the maximum of KuPR’s

   

   

   View Expanded

   
3. Condition on the maximum of KaPR’s

   

   

   View Expanded

   
   The value of flagHeavyIcePrecip is the sum of the variables , and ,

   

   

   View Expanded

   

From this construction of the 5-bit flag, it is evident that the most significant bit represents the fulfillment of condition A, the next two bits condition B with KuPR’s , and the last two bits condition C with KaPR’s .

Note that DFR_m is examined only when KuPR’s is larger than 27 dBZ in order to avoid misjudgment resulting from the noisy signal of KaPR when its echo is weak. With the condition of dBZ for ice precipitation detection, most light ice precipitation will be missed by this algorithm. Nevertheless, the test results shown in the following two sections indicate that the DFR_m algorithm (condition A) can detect intense ice precipitation more frequently than the condition B or C. Note that conditions A and C are applicable only in the inner swath, whereas condition B is examined in the full swath.

3. Performance of the algorithm

We compared the distribution of pixels flagged by flagHeavyIcePrecip with ground-based measurements in three cases.

Figure 1 shows the case with a very intense hailstorm off the coast of Naples in Italy on 5 September 2015 (orbit number: 8630, scan: around 4938, time: 0848:11 UTC). Very low brightness temperatures measured at the high-frequency channels of GPM/GMI indicate the presence of a large amount of ice particles within the storm. Ground-based polarimetric radar data also support the presence of large graupel/hail particles. In fact, baseball-size hailstones were observed at the surface during this violent hailstorm event (Marra et al. 2017).

Radar echoes and hail detection off the coast of Naples on 5 Sep 2015. (a) Vertical cross sections of the KuPR (contour) and DFR_m (color) along the brown line shown in (b). Contour lines are drawn with 4-dB intervals. Broken line at about 6 km indicates the C height. (b) Heavy ice precipitation flagged by flagHeavyIcePrecip. Pixels flagged by conditions A, B, and C are respectively indicated by red, cyan, and yellow circles. When they overlap, the circles for B and C are reduced to a smaller circle and a dot, respectively. Circles with light gray indicate outer swath pixels. The star denotes the location of Naples.

Citation: Journal of Atmospheric and Oceanic Technology 35, 3; 10.1175/JTECH-D-17-0120.1

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Radar echoes and hail detection off the coast of Naples on 5 Sep 2015. (a) Vertical cross sections of the KuPR (contour) and DFR_m (color) along the brown line shown in (b). Contour lines are drawn with 4-dB intervals. Broken line at about 6 km indicates the C height. (b) Heavy ice precipitation flagged by flagHeavyIcePrecip. Pixels flagged by conditions A, B, and C are respectively indicated by red, cyan, and yellow circles. When they overlap, the circles for B and C are reduced to a smaller circle and a dot, respectively. Circles with light gray indicate outer swath pixels. The star denotes the location of Naples.

Citation: Journal of Atmospheric and Oceanic Technology 35, 3; 10.1175/JTECH-D-17-0120.1

• Download Figure
• Download figure as PowerPoint slide

Radar echoes and hail detection off the coast of Naples on 5 Sep 2015. (a) Vertical cross sections of the KuPR (contour) and DFR_m (color) along the brown line shown in (b). Contour lines are drawn with 4-dB intervals. Broken line at about 6 km indicates the C height. (b) Heavy ice precipitation flagged by flagHeavyIcePrecip. Pixels flagged by conditions A, B, and C are respectively indicated by red, cyan, and yellow circles. When they overlap, the circles for B and C are reduced to a smaller circle and a dot, respectively. Circles with light gray indicate outer swath pixels. The star denotes the location of Naples.

Citation: Journal of Atmospheric and Oceanic Technology 35, 3; 10.1175/JTECH-D-17-0120.1

• Download Figure
• Download figure as PowerPoint slide

Figure 1a shows the vertical cross sections of measured with contour lines and DFR_m with color. The broken line near 6 km indicates the C height. Since hail will produce large reflectivity factors, the conditions B and C are easily met. The regions where condition A is satisfied correspond to yellow-to-red areas above the C height in this figure. Figure 1b shows the output from flagHeavyIcePrecip. Red pixels are the pixels that satisfy condition A. In this particular case, the regions that satisfy condition A nearly correspond to the regions where condition B [ dBZ] or C [ dBZ] is satisfied. Note that the region with dBZ in Fig. 1a is always inside the region with DFR_m dB. This fact implies that condition A can be used to detect graupel and hail, since reflectivity factors larger than 40 dBZ at or above C height generally indicate the existence of hail or graupel (Williams et al. 1989).

Figure 2 shows an intense hailstorm case near Fort Worth, Texas, on 26 May 2015 (orbit number: 7052, scan: around 5150, time: 2224:23 UTC). This storm was measured with a ground-based polarimetric radar at almost exactly the same time (2225:19 UTC) as well. Figure 2a shows the hydrometeor types identified by a polarimetric method at the height of 4 km. Hail regions are denoted in red, and wet and dry graupel regions are shown in yellow and green, respectively. This storm was overpassed by the GPM Core Observatory satellite 1 min earlier. Figure 2b shows the output from the DPR product. In this case, all three conditions, A–C, are met at many pixels.

Hail detection in the thunderstorm near Fort Worth on 26 May 2015. (a) Hydrometeor identification by a ground-based polarimetric radar (provided by Dr. Cecil); BD = big drops/melting hail, HA = hail, WG = high-density graupel, DG = low-density graupel, VI = vertical ice, WS = wet snow, DS = dry snow, CR = ice crystals, RN = rain, DZ = drizzle, UC = unclassified. (b) Output from flagHeavyIcePrecip from DPR; pixels are color coded as in Fig. 1b. (c) Output from flagHeavyIcePrecip from KaPR; cyan: dBZ, dark blue: dBZ. (d) Output from flagHeavyIcePrecip from KuPR; cyan: dBZ, dark blue: dBZ, red: . Circles with light gray indicate outer swath pixels.

Citation: Journal of Atmospheric and Oceanic Technology 35, 3; 10.1175/JTECH-D-17-0120.1

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Hail detection in the thunderstorm near Fort Worth on 26 May 2015. (a) Hydrometeor identification by a ground-based polarimetric radar (provided by Dr. Cecil); BD = big drops/melting hail, HA = hail, WG = high-density graupel, DG = low-density graupel, VI = vertical ice, WS = wet snow, DS = dry snow, CR = ice crystals, RN = rain, DZ = drizzle, UC = unclassified. (b) Output from flagHeavyIcePrecip from DPR; pixels are color coded as in Fig. 1b. (c) Output from flagHeavyIcePrecip from KaPR; cyan: dBZ, dark blue: dBZ. (d) Output from flagHeavyIcePrecip from KuPR; cyan: dBZ, dark blue: dBZ, red: . Circles with light gray indicate outer swath pixels.

Citation: Journal of Atmospheric and Oceanic Technology 35, 3; 10.1175/JTECH-D-17-0120.1

• Download Figure
• Download figure as PowerPoint slide

Hail detection in the thunderstorm near Fort Worth on 26 May 2015. (a) Hydrometeor identification by a ground-based polarimetric radar (provided by Dr. Cecil); BD = big drops/melting hail, HA = hail, WG = high-density graupel, DG = low-density graupel, VI = vertical ice, WS = wet snow, DS = dry snow, CR = ice crystals, RN = rain, DZ = drizzle, UC = unclassified. (b) Output from flagHeavyIcePrecip from DPR; pixels are color coded as in Fig. 1b. (c) Output from flagHeavyIcePrecip from KaPR; cyan: dBZ, dark blue: dBZ. (d) Output from flagHeavyIcePrecip from KuPR; cyan: dBZ, dark blue: dBZ, red: . Circles with light gray indicate outer swath pixels.

Citation: Journal of Atmospheric and Oceanic Technology 35, 3; 10.1175/JTECH-D-17-0120.1

• Download Figure
• Download figure as PowerPoint slide

Since Ka-band echoes are available only in the inner swath of measurement, condition B is the only information we can use from flag flagHeavyIcePrecip in the outer swath. In this outer part of the swath, Fig. 2d shows that the distribution of dark blue and red pixels agrees pretty well with hail regions identified by the ground-based radar, confirming the premise that 40 dBZ at to above C height is a good threshold for the detection of hail.

In the inner part of the swath, the regions flagged by condition A and condition B or C match pretty well with the hail or graupel regions classified by the polarimetric radar. It seems that the regions flagged by condition A are somewhat wider than the regions with hail or graupel identified by the ground-based radar. This disagreement may partly come from the comparison of the flags that are determined from the profile data with the hydrometeor identification map at the particular height of 4 km.

The figure shows that the regions that satisfy condition A are mostly larger than the regions flagged by B or C, and that the dual-frequency method is more sensitive to the intense ice precipitation. In fact, a comparison of Figs. 2b and 2d shows that the regions flagged by condition A are closer to the regions flagged by condition B with the 35-dBZ threshold than those with 40 or 45 dBZ. Similarly, Fig. 2c shows that even in this rather intense hailstorm case, KaPR’s reflectivity factor exceeds 35 dBZ only at a few pixels and never exceeds 40 dBZ and that the regions flagged by condition C with the KaPR’s 30-dBZ threshold (light blue in Fig. 2c, yellow dot in Fig. 2b) agree pretty well with the regions flagged by condition B with the KuPR’s 40-dBZ threshold (dark blue and red pixels in Fig. 2d). The difference caused by different thresholds may be used to characterize the storms in future studies.

A widespread snowstorm hit the northeastern part of Japan on 18 January 2016 (local time). Figure 3a shows the vertical cross sections of Ku-band’s (contour lines) and DFR_m (color) along the brown line in Figs. 3b and 3c. The contour lines are drawn with 2.5-dB intervals, and the regions of DFR_m larger than 7 dB are indicated by warm colors (from yellow through red to pink) in the figure. Figure 3b shows the estimated attenuation-corrected radar reflectivity factor () at 2 km above sea level given by KuPR’s standard product (orbit number: 10725, scan: around 2935, time: 0001:56, 18 January 2016 UTC). It shows that the storm is widespread and distributes rather uniformly. There are about 100 stations in the Japan Meteorological Agency's (JMA) Automatic Meteorological Data Acquisition System (AMeDAS) within the area shown in this figure. Almost all of them recorded a temperature lower than 2.5C and only a few stations reported a temperature between 2.5° and 5°C. Either snow or sleet was observed at all of the JMA’s 13 ground observatories located in the part of the main island of Japan depicted in this figure.

Detection of intense ice precipitation over northern Japan on 18 Jan 2016. (a) Vertical cross sections of KuPR’s (contour lines with every 2.5-dB intervals) and DFR_m (color) along the brown line shown in (b) and (c). (b) Estimated radar reflectivity factor () in dBZ at 2 km. (c) Pixels flagged by flagHeavyIcePrecip in the DPR product. Color code as in Fig. 1. Circles with light gray indicate outer swath pixels.

Citation: Journal of Atmospheric and Oceanic Technology 35, 3; 10.1175/JTECH-D-17-0120.1

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Detection of intense ice precipitation over northern Japan on 18 Jan 2016. (a) Vertical cross sections of KuPR’s (contour lines with every 2.5-dB intervals) and DFR_m (color) along the brown line shown in (b) and (c). (b) Estimated radar reflectivity factor () in dBZ at 2 km. (c) Pixels flagged by flagHeavyIcePrecip in the DPR product. Color code as in Fig. 1. Circles with light gray indicate outer swath pixels.

Citation: Journal of Atmospheric and Oceanic Technology 35, 3; 10.1175/JTECH-D-17-0120.1

• Download Figure
• Download figure as PowerPoint slide

Detection of intense ice precipitation over northern Japan on 18 Jan 2016. (a) Vertical cross sections of KuPR’s (contour lines with every 2.5-dB intervals) and DFR_m (color) along the brown line shown in (b) and (c). (b) Estimated radar reflectivity factor () in dBZ at 2 km. (c) Pixels flagged by flagHeavyIcePrecip in the DPR product. Color code as in Fig. 1. Circles with light gray indicate outer swath pixels.

Citation: Journal of Atmospheric and Oceanic Technology 35, 3; 10.1175/JTECH-D-17-0120.1

• Download Figure
• Download figure as PowerPoint slide

As shown in Fig. 3c, condition A is met at several places, and the flagHeavyIcePrecip is set at many pixels in the swath. In this particular case, neither condition B nor C is satisfied except at one pixel at 38.3°N, 138.7°E, where condition B with the threshold 35 dBZ is barely fulfilled. A possible reason for this large difference in flagging by conditions A and B is that aggregates of snow crystals may be the dominant ice particles in this storm, since they can produce large DFR even for relatively small (Tyynelä and Chandrasekar 2014). This example proves that this flag is set not only by hail but also some intense snow or graupel. We can generally say that the cases satisfied by condition B are a subset of those satisfied by condition A. As a result, we can use the part of the flag as an indicator of intense ice or snow precipitation, whereas or can be used to detect thunderstorms.

4. Global distribution of intense ice precipitation

Since we have confirmed that the new flag is properly set at pixels with intense ice precipitation, we looked at the global distribution of the flagged pixels in 2015.

Figure 4a shows the percentage of observation pixels in which flagHeavyIcePrecip indicates that condition A is set in each 2° × 2° box in 2015. Figure 4b shows the corresponding percentage in which KuPR’s maximum is larger than 40 dBZ above the C height, that is, condition B with . This condition generally implies that there are enough graupel hydrometeors to create thunder and lightning. The fact that flagHeavyIcePrecip with condition A occurs more frequently than that with dBZ also implies that condition A has a higher sensitivity to intense ice precipitation.

Frequencies at which flagHeavyIcePrecip was set with different conditions over 2° × 2° boxes in 2015. (a) Percentage of occurrences in which flagHeavyIcePrecip is set by condition A [ dB, (Ku) dBZ above C level]. (b) Percentage of occurrences in which the maximum of of KuPR above the C height exceeds 40 dBZ (condition B with ). (c) Percentage of occurrences in which flagHeavyIcePrecip is set by condition A and the surface air temperature is below C. Percentages are calculated relative to the total number of the measured pixels that include no rain pixels.

Citation: Journal of Atmospheric and Oceanic Technology 35, 3; 10.1175/JTECH-D-17-0120.1

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Frequencies at which flagHeavyIcePrecip was set with different conditions over 2° × 2° boxes in 2015. (a) Percentage of occurrences in which flagHeavyIcePrecip is set by condition A [ dB, (Ku) dBZ above C level]. (b) Percentage of occurrences in which the maximum of of KuPR above the C height exceeds 40 dBZ (condition B with ). (c) Percentage of occurrences in which flagHeavyIcePrecip is set by condition A and the surface air temperature is below C. Percentages are calculated relative to the total number of the measured pixels that include no rain pixels.

Citation: Journal of Atmospheric and Oceanic Technology 35, 3; 10.1175/JTECH-D-17-0120.1

• Download Figure
• Download figure as PowerPoint slide

Frequencies at which flagHeavyIcePrecip was set with different conditions over 2° × 2° boxes in 2015. (a) Percentage of occurrences in which flagHeavyIcePrecip is set by condition A [ dB, (Ku) dBZ above C level]. (b) Percentage of occurrences in which the maximum of of KuPR above the C height exceeds 40 dBZ (condition B with ). (c) Percentage of occurrences in which flagHeavyIcePrecip is set by condition A and the surface air temperature is below C. Percentages are calculated relative to the total number of the measured pixels that include no rain pixels.

Citation: Journal of Atmospheric and Oceanic Technology 35, 3; 10.1175/JTECH-D-17-0120.1

• Download Figure
• Download figure as PowerPoint slide

Detection of intense ice precipitation does not necessarily mean that solid precipitation reaches the surface. In fact, most ice precipitation detected in the tropics exists only at very high altitudes. To see whether the ice particles are reaching the actual surface in those cases, only the cases in which flagHeavyIcePrecip is set are counted and at the same time the surface air temperature is approximately below C. This temperature was estimated from the value stored in “phase” variable at the surface range bin determined by the KuPR. The temperature data used to create standard GPM products are taken from the Global Analysis Dataset (GANAL) created by the JMA (Iguchi et al. 2017). The value stored in phase is calculated at the center of the range bin from the GANAL temperature data and rounded down to the nearest integer. As a result, assuming that the radar correctly identifies the surface range bin, the actual air temperature at the surface should be lower than approximately 1°C considering the round-off error and the maximum possible height bias of 62.5 m resulting from the quantization of range samples. Since surface snowfall occurs even when the surface air temperature is a few degrees above C, 1°C is considered as a conservative threshold (Sims and Liu 2015).

The results are shown in Fig. 4c. In these cases we can assume that ice precipitation reaches the surface without melting. An interesting observation is the relatively high occurrence of intense ice precipitation over the Antarctic Ocean, which is also found with CloudSat observations (Liu 2008; Kulie et al. 2016). Since only very intense ice precipitation is flagged in the current algorithm, the frequency of detecting ice precipitation is much lower with DPR than with CloudSat, but the patterns of frequency distribution have some similarity. However, relatively frequent detection of ice precipitation over the North Atlantic Ocean to the south of Greenland and over the Sea of Okhotsk and the Bering Sea is not seen in the snowfall fraction derived from CloudSat data (Fig. 6a in Kulie et al. 2016). Its distribution seems to resemble more the distribution of the mean annual liquid equivalent snowfall (Fig. 6b in Kulie et al. 2016). This is not surprising if intense ice precipitation contributes much to the total amount of liquid equivalent ice precipitation.

To see the seasonal variation, similar figures are drawn for the boreal winter (January, February, and December) and summer (June–August) months in 2015 (Fig. 5). Relatively high percentage regions in the tropics and subtropics, such as those in central Africa, South America, and Australia, in Figs. 5a and 5b correspond well with active lightning regions in those seasons. This observation is not surprising when we think that lightning is associated with the existence of ice particles, in particular graupel, at high altitudes.

Frequencies of occurrence of condition A [ dB, (Ku) dBZ above C level] relative to the total number of measured pixels for (a) boreal winter in 2015 and (b) boreal summer in 2015. (c) As in (a), but with the surface air temperature less than 1°C. (d) As in (b), but with the surface air temperature less than 1°C.

Citation: Journal of Atmospheric and Oceanic Technology 35, 3; 10.1175/JTECH-D-17-0120.1

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Frequencies of occurrence of condition A [ dB, (Ku) dBZ above C level] relative to the total number of measured pixels for (a) boreal winter in 2015 and (b) boreal summer in 2015. (c) As in (a), but with the surface air temperature less than 1°C. (d) As in (b), but with the surface air temperature less than 1°C.

Citation: Journal of Atmospheric and Oceanic Technology 35, 3; 10.1175/JTECH-D-17-0120.1

• Download Figure
• Download figure as PowerPoint slide

Frequencies of occurrence of condition A [ dB, (Ku) dBZ above C level] relative to the total number of measured pixels for (a) boreal winter in 2015 and (b) boreal summer in 2015. (c) As in (a), but with the surface air temperature less than 1°C. (d) As in (b), but with the surface air temperature less than 1°C.

Citation: Journal of Atmospheric and Oceanic Technology 35, 3; 10.1175/JTECH-D-17-0120.1

• Download Figure
• Download figure as PowerPoint slide

In the high-latitude regions, however, the regions flagged with condition A are not related to lightning regions. It is well-known that some lightning occurs east of Japan over the North Pacific Ocean in winter. However, the occurrence of condition A in the North Pacific Ocean and the North Atlantic Ocean in boreal winter is very frequent and out of proportion to lightning activities compared with other regions.

As inferred from the difference between Figs. 4a and 4b, the pixels flagged by condition A in high latitudes are not necessarily associated with intense storms. In other words, these pixels are flagged by large ice particles, such as large snowflakes or aggregates, but not necessarily by particles associated with strong convections.

To see how often such flagged ice precipitation actually reaches the surface, the data are filtered with the condition that the surface air temperature is below C. Figures 5c and 5d show the results. All flagged boxes in the tropics and subtropical regions in Figs. 5a and 5b disappear, and only boxes at high-latitudinal regions remain. The distribution of intense ice precipitation pixels in Fig. 5c shows that there are frequent surface-reaching snowfalls in marginal seas of the northwestern Pacific Ocean (the Sea of Okhotsk and the Bering Sea) and of the northwestern Atlantic Ocean to the south of Greenland, including Buffin Bay and Hudson Bay. It is also interesting to note that some ice precipitation reaches the sea surface without melting over the Antarctic Ocean even in austral summer. Less frequently but similar summer ice precipitation to the ocean surface can be seen in Fig. 5d between Canada and Greenland.

As described in the algorithm section, the intense ice precipitation flag is based on the echoes above the estimated C height. The surface air temperature is taken from the GANAL data stored in the standard GPM products. Although these temperatures may have some error, the error cannot be so large as to invalidate the abovementioned observation.

5. Summary

To detect intense falling snow, graupel, and hail, a new flag was introduced in the version 5 products of the GPM/DPR. The flag is set when a relatively strong echo is measured above the C isotherm height in both dual-frequency and single-frequency algorithms. In the dual-frequency products, the flag is also set when a DFR_m larger than 7 dB with the corresponding KuPR’s larger than 27 dBZ appears above the C isotherm height.

The performance of the algorithm is tested in typical hailstorm cases and in a relatively intense snowstorm. In the hailstorm cases in Italy and the United States, both the single-frequency and dual-frequency algorithms give very reasonable results. In the widespread snowfall case over Japan, the dual-frequency method that utilizes the DFR_m identifies the intense snowfall regions well.

These examples indicate that the new flag, in particular the flag set by the DFR_m method, is useful in determining the distribution of intense ice precipitation, even though it may miss many light ice precipitation because of the limited sensitivity of DPR.

A year of statistics of the flag is analyzed to see the distribution of intense ice precipitation. The results indicate that the dual-frequency algorithm flags more pixels than the single-frequency algorithm with the conditions that are often used to identify a thunderstorm. It detects intense ice precipitations not only in strong tropical convections but also in winter storms at high latitudes.

Because of the surface clutter, the DPR cannot obtain precipitation echoes very close to the actual surface. Existence of intense ice precipitation above the C isotherm height does not imply any solid precipitation at the surface. To see the fraction of ice precipitation that actually reaches the surface without melting, only the cases with the surface air temperature below C are counted. An interesting observation is that there remain quite a few pixels where the flagHeavyIcePrecip is set after being filtered by the surface air temperature condition. In these cases, ice precipitation particles are likely to reach the surface without melting. Considering the importance of heat transfer between the atmosphere and ocean, information about the existence of ice precipitation directly to the ocean surface must be invaluable.

Although CloudSat has a much better sensitivity of detecting ice precipitation than the DPR, it cannot quantify the amount of ice very well when ice particles are large or when ice precipitation is intense because of the complicated effect of non-Rayleigh scattering and multiple scattering at the W band. In this sense the DPR and CloudSat are complementary to each other for ice detection. As snow distribution data obtained by CloudSat helped improve ice retrieval algorithms for microwave radiometers, the flagHeavyIcePrecip flag from DPR could serve as useful validation data for identification of ice precipitation by microwave radiometers in a similar way that PR or DPR data were used to improve the hail identification algorithm of microwave radiometers (Cecil 2009, 2011; Cecil and Blankenship 2012; Mroz et al. 2017).

Improvement of the flag by adding more information, such as the height where the flag is set, will be a future task. Although it is a very simple flag with room for improvement, the new flag must be of help to researchers in the study of ice precipitation, in particular in combination with data from cloud radar or microwave radiometers.