1. Introduction
The Global Precipitation Measurement (GPM) Microwave Imager (GMI) measures brightness temperatures Tb in channels ranging from 10 to 183 GHz (Hou et al. 2014). The highest frequency channels (166 and 183 GHz), with wavelengths between 1 and 2 mm, are sensitive to the scattering of upwelling radiation by precipitation and large cloud hydrometeors. A Tb depression in these channels can indicate thick cloud and precipitation (Hong et al. 2005). The 37- and 89-GHz channels, with wavelengths between 3 and 8 mm, have less sensitivity to small particles. Over a radiometrically cold water surface, emission by raindrops raises Tb in these channels, but scattering by precipitation-sized ice reduces Tb relative to values associated with liquid precipitation. The Tb values indicative of scattering signatures may not be cooler than the nonprecipitating, water background, but it will be noticeably cooler than the profiles associated with liquid precipitation (Spencer et al. 1983, 1987, 1989). The lowest frequency channels (10 and 19 GHz), with wavelengths between 1 and 3 cm, have little sensitivity to cloud and small precipitation hydrometeors. They have traditionally been used for precipitation retrievals because emission by liquid raindrops raises their Tb relative to a radiometrically cold water surface (e.g., Wilheit et al. 1977; Kummerow et al. 1998; Hou et al. 2014), and ice hydrometeors above the rain layer are usually not large and dense enough to have an offsetting scattering effect on these longer-wavelength channels. Very large particles, such as hail, can produce a scattering signature in any of these channels (Leppert and Cecil 2015, 2019), and the 19- and 37-GHz channels are sometimes used for hail detection (Cecil 2009; Cecil and Blankenship 2012; Mroz et al. 2016; Bang and Cecil 2019).
Research by Leppert and Cecil (2015) used passive microwave Tb measurements and dual-polarization-derived hydrometeor types from a field experiment in Oklahoma and Kansas to identify the Tb values most commonly associated with different hydrometeor species. Their research combined high-resolution passive microwave Tb values from airborne radiometers to cover the same frequency range as GMI. For all frequencies, the presence of hail or large raindrops (interpreted as melted hail) was associated with Tb values colder than any other hydrometeor class. They found significant overlap in Tb values for high-density and low-density graupel, as identified by polarimetric radar. They found that 10 GHz had little sensitivity to varying hydrometeor species, but small depressions were associated with hail, big drops, and high-density graupel. As frequency increased, there were more notable depressions in Tb and higher sensitivity to different hydrometeors. For the highest frequencies, there was little distinction between hail and graupel, but there was a distinction between profiles containing these versus profiles without these hydrometeor types. With the Leppert and Cecil (2015) study being conducted over a radiometrically warm land surface, no lower frequencies were particularly useful for identifying liquid precipitation without the presence of ice aloft. The highest frequencies were able to distinguish between profiles containing liquid precipitation and nonprecipitating profiles, but there were concerns; this was due to the presence of small ice aloft that could not be detected by the radar, and it may be difficult to separate snow, small ice, rain, and drizzle. Their research findings indicate the possibility to identify various hydrometeor species using Tb measurements from the GMI, which provides the foundation for this research.
The goal of this research is to develop an empirical method to identify hydrometeor types based on passive microwave Tb values. GMI Tb values matched with ground-based hydrometeor identification (HID) data from the GPM Validation Network (GVN) dataset are used to develop the empirical method. This research identifies the Tb values most commonly associated with different hydrometeor classifications and expands upon the research of Leppert and Cecil (2015). Empirical lookup tables are created to estimate the likelihood that a specific hydrometeor type will be present in a profile given the combination of Tb values from selected GMI channels. The empirical lookup tables are then tested on an independent testing dataset and applied to an individual storm case with level 1C GMI Tb values for further analysis.
2. Data
This research uses 8 years of GVN data (January 2015–December 2022) for training and testing the HID retrieval. The data consist of precipitation events from all seasons, but many of the radars are located in the eastern and southeastern portions of the United States. The GVN data (Peterson 2021) include collocated level 1C Tb values from GMI (Berg 2022) and HID from ground-based dual-polarization radars (Schwaller and Morris 2011; Gatlin et al. 2020), primarily located in the United States, with 48 of those radars used here. The GVN processing matches ground-based radar data to GMI Tb values along the radiometer’s 53° incidence angle, rather than using a true vertical profile. This matching process requires 100 precipitating profiles to be located within 100 km of the ground radar, which helps limit the effects of beam broadening. The radar-based HIDs are retrieved following Dolan and Rutledge (2009) for each gate at the radar’s native resolution, as part of the GVN processing Dolan and Rutledge (2009). Those radar gates are then matched to the GMI footprint at each radar elevation angle. Those matched radar gates are then subjected to our hierarchy, which assigns a single HID to each GMI footprint. There is a possibility that HIDs high in the vertical profile will be missed by the ground radar when located close to the radar site. This research uses the reflectivity product from the GVN data to filter out nonprecipitating profiles. Only profiles with the precipitation echoes in the GVN radar data, matched to nearly coincident GMI Tb data, are included in this analysis.
The GMI Tb measurements from the horizontally (H) and vertically (V) polarized channels are combined into polarization corrected temperature (PCT) for 10, 19, 37, and 89 GHz (Cecil and Chronis 2018). At 166 GHz, the vertically polarized channel (V166) and the polarization difference between the 166-GHz channels (V166 − H166), referred to as Diff166, are used. The difference between two bands centered on a water vapor line at 183.3 GHz, (V183 ± 7) − (V183 ± 3) GHz, is referred to as Diff183. Finally, a pseudochannel combination is formed as [(PCT10 − PCT19) − Diff183 GHz] and referred to as Diff10_19_183. These channels and channel combinations will be used to form the empirical relationships with HID and to generate the lookup tables. Motivation for using these channel combinations is discussed along with the empirical results in section 3.
The HID in the GVN data uses a fuzzy-logic method, which assigns the most likely HID based on how well that HID matches the observed NEXRAD level 2 radar products, following the methods of Dolan and Rutledge (2009). The GVN radar data are subjected to a quality control process to identify and exclude radar gates experiencing beam blockage or other issues. We require radars included in the training data to have 14 elevations for volume coverage patterns (VCP) in order to better represent the vertical profile observed by the GMI. The GVN data include the radar-based HID that is assigned for each ground radar range gate and the total number of gates associated with each HID. This research uses a hierarchical approach, similar to that of Leppert and Cecil (2015), to assign a single HID to each GMI footprint. The hierarchy prioritizes species that are encountered least often and tend to have greater impact on the radiometer observed Tb. Leppert and Cecil (2015) found that big drops and hail were most often associated with the largest reduction of Tb. Gatlin et al. (2015) found that big drops are often associated with melting hail and graupel aloft. This research combines hail and big drops into a single hail class. In our training database, over 200 000 GMI footprints include the big drops HID category, but all those footprints also contain hail. Nearly 100 000 additional footprints contain hail without big drops. The GVN data included separate classifications for high- and low-density graupel, but for this analysis, they are combined into a single graupel class to avoid uncertainty about that distinction. Based on the results from Leppert and Cecil (2015), several of the smaller ice species, vertically oriented ice, ice crystal, aggregates, and snow, have been combined into a single snow/small ice class. Finally, rain and drizzle have been combined into a single rain class. The final hierarchy used in this research is hail, graupel, snow, and then rain. The hierarchical order and the amount of each class within the training data are given in Table 1. This hierarchical order will prioritize hail, and by association more convective weather. In practice, a profile containing hail almost always included all the other HID categories somewhere within the profile, but the inverse is not true. A profile identified as graupel likely also contains snow and rain, but does not include hail. The vast majority of profiles do include liquid rain, but in our hierarchy, a profile identified as liquid rain does not include the signatures of any of the ice categories. Assigning a profile as hail, graupel, or snow does not necessarily mean that those species reach the surface before melting—those profiles usually do have liquid rain at the surface. It is also important to note that the hail HID does not include information about particle size; small hail may be common aloft in thunderstorms that would not be expected to produce large hail reaching the surface.
GVN HID separated according to the hierarchy and the number of each HID within the training data. Graupel is made up of high- and low-density graupel. Snow is made up of aggregates/dry snow and vertically oriented ice/ice crystals. Rain consists of both rain and drizzle.
Individual orbits of level 1C GMI Tb are used to demonstrate the GMI-based HID retrieval on locations and dates outside of the GVN development and testing sample. The probability of precipitation (PoP) and 2-m temperature (2mT) are obtained from the GMI level 2A GPROF files (Kummerow et al. 2015) for the corresponding 1C GMI orbits. When PoP is less than 100%, it is used to normalize the GMI-predicted probabilities of each HID type lower. A 2mT threshold of greater than or equal to 5°C is used to filter out potential contamination by surface snow and ice cover.
3. Single frequency relationships with hydrometeor type
For the 19-, 37-, and 89-GHz PCT, the percentage of profiles with hail increases as PCT decreases (Fig. 1). For the 89-GHz channel (Fig. 1d), a negligible fraction of the profiles with PCT values around 275 K include hail, and only around 10% include graupel. The remainder are primarily snow, with an increasing fraction of rain, without ice, for the warmest Tb values. For decreasing PCT values, first the likelihood of graupel increases and then the likelihood of hail increases. Around 250 K, over half the profiles contain graupel, but more are still identified as snow than as hail. The likelihood of hail steadily increases for decreasing 89-GHz PCT values below 200 K, with hail approaching 100% for the coldest Tb bins. At PCT values less than 180 K, virtually all the profiles contain graupel and hail.
Fraction of ground-radar-derived HIDs as a function of GMI PCT. (a) 10-GHz PCT, (b) 19-GHz PCT, (c) 37-GHz PCT, and (d) 89-GHz PCT, with each bar representing a 5-K bin.
Citation: Journal of Applied Meteorology and Climatology 63, 7; 10.1175/JAMC-D-23-0196.1
With longer wavelengths, the increased likelihood of hail is more abrupt in the 37-GHz (Fig. 1c) and 19-GHz (Fig. 1b) channels. For the 37-GHz PCT around 230 K, over 80% of profiles include hail, and almost all of the remainder include graupel. For 19-GHz PCT around 240 K, nearly 80% of the profiles contain hail. This pattern does not continue to the 10-GHz channels (Fig. 1a), where the peak likelihood of hail (only around 10%) is near 300 K and the peak likelihood of graupel (around 30%) is around 285 K. The large footprint size of the 10-GHz channels (19 km × 32 km, compared to 11 km × 18 km for 19 GHz and smaller sizes for higher frequencies: Hou et al. 2014) may explain the relative insensitivity to hydrometeor type in our results for the 10-GHz channels. Leppert and Cecil (2015) did find a rapid increase in hail and graupel with decreasing 10-GHz Tb, using much higher-resolution (3 km × 3 km) airborne data.
Using the higher-frequency channels of the GMI, more distinction between graupel and hail can be observed in Tb values (Fig. 2). The V166 GHz channel (Fig. 2a) is more sensitive to scattering by smaller ice, including graupel. The Tb range between 155 and 235 K contains a higher percentage of graupel than any of the other HIDs. Only below 155 K does hail become the most likely HID. Similar to results from the 89-GHz channels, the likelihood of hail steadily increases as Tb values decrease, with the coldest bins approaching 100%. The V166 and Diff183 channels do show a higher fraction of rain at higher Tb values, but snow is still the dominant class.
As in Fig. 1, but for GMI frequencies (a) V166 GHz, (b) Diff166 GHz, (c) Diff183 GHz, and (d) Diff10_19_183 GHz. The bin sizes are 5 K for V166 and Diff10_19_183 GHz and 2.5 K for Diff166 and Diff183 GHz.
Citation: Journal of Applied Meteorology and Climatology 63, 7; 10.1175/JAMC-D-23-0196.1
Previous research supports using the polarization difference at 166 GHz to identify different types of small ice (Gong and Wu 2017). In Fig. 2b, 166-GHz polarization differences near zero are usually accompanied by snow with small fractions of graupel or hail. As the polarization difference increases to around 15 K (a vertically polarized Tb substantially greater than the horizontally polarized Tb), the likelihood of graupel increases to over 50%, while the likelihood of hail increases, but remains quite low.
Hong et al. (2005) examined deep convective clouds using differences between the narrowband and wideband channels centered on the 183.3-GHz water vapor line. Deep clouds with large ice reduce Tb in the wideband channels more so than the narrowband channels. An increasingly negative Diff183 in Fig. 2c reflects this, with the lower wideband Tb indicating a greatly increased likelihood of the presence of graupel. Positive values of Diff183 in precipitating profiles, that is, higher Tb in the wideband channel than in the narrowband channel, are usually associated with snow and rain, as shown in Fig. 2c. The results in Fig. 2c indicate the ability to separate profiles containing hail or graupel from those containing only small ice or rain. The overlapping results for hail and graupel indicate the Diff183 GHz relationship, by itself, will not be useful for separating hail and graupel.
Based on single-frequency results such as those in Figs. 1 and 2a–c, we constructed a “pseudochannel” Diff10_19_183 from multiple frequencies. Since the 10-GHz PCT shows little sensitivity to the HID types in Fig. 1a, we treat it as a background value and subtract the 19-GHz PCT to essentially form a scattering index. The 19-GHz PCT showed high sensitivity to hail in Fig. 1b. Large values of the PCT10 − PCT19 difference (around 40 K and greater) should indicate the high likelihood of hail. Increasingly negative values of Diff183 between 0 and −40 K in Fig. 2c indicate the increasing likelihoods of graupel and hail. Subtracting Diff183 from the PCT10 − PCT19 difference retains the signal for a high likelihood of hail at high values of Diff10_19_183, but adds a signal for a substantial likelihood of graupel without hail for values between about 10 and 35 K (Fig. 2d). Diff10_19_183 values near zero have a substantial likelihood of snow. Increasingly negative values of Diff10_19_183 are associated with high likelihoods of snow or small ice without signatures of hail or graupel. The Diff10_19_183 GHz relationship provides an ability to separate between graupel, hail, and smaller ice or rain.
4. Multifrequency relationships with hydrometeor type
Based on the results from Figs. 1 and 2, and the suggestions of previous research (Spencer and Santek 1985; Leppert and Cecil 2015; Ferraro et al. 2015; Bang and Cecil 2019), two-dimensional histograms of Tb values are constructed for each HID to determine the likelihood of each HID type as a function of the combinations of Tb. The percentages that are calculated are then used as empirical lookup tables for multifrequency classification of HID. The lookup tables identify the combination of Tb values that are most commonly associated with the different HIDs. Six frequency combinations are used, after initially testing 26 different combinations and removing combinations that provided little to no extra benefit. The six chosen frequency combinations include 37-GHz PCT versus 89-GHz PCT (3789), 37-GHz PCT versus V166 (37V166), 37-GHz PCT versus Diff183 (37Diff183), 37-GHz PCT versus Diff166 (37Diff166), 37-GHz PCT versus Diff10_19_183 (37Diff10_19_183), and 89-GHz PCT versus Diff183 (89Diff183). The Tb bins remain the same as in the single frequency barplots (Figs. 1 and 2), i.e., 3789 has bins 5 K × 5 K, while 89Diff183 has bins 5 K × 2.5 K. The line contours in Figs. 3–5 and Figs. S1–S3 in the online supplemental material represent the histograms of the total number of precipitating profiles for each Tb combination.
PCT 3789 empirical lookup table used for multifrequency classification. The fraction of (a) hail in each 5 K × 5 K bin, (b) graupel in each bin, (c) snow in each bin, and (d) rain in each bin, value represented by color scale on the right side of plots. The contours represent the samples size increasing from 1, 10, 100, 1000, and 10 000. The 10 sample contour is represented by the black, dashed contour, and the 100 sample contour is represented by solid black.
Citation: Journal of Applied Meteorology and Climatology 63, 7; 10.1175/JAMC-D-23-0196.1
As in Fig. 3, but for 37Diff10_19_183 GHz.
Citation: Journal of Applied Meteorology and Climatology 63, 7; 10.1175/JAMC-D-23-0196.1
As in Fig. 3, but for 89Diff183 GHz.
Citation: Journal of Applied Meteorology and Climatology 63, 7; 10.1175/JAMC-D-23-0196.1
The value in this multifrequency approach can be seen by considering differences along a vertical or horizontal line in Fig. 3, holding one frequency’s PCT constant. Recall from Fig. 1c that for 37-GHz PCT around 230 K, over 80% of profiles include hail. Along a vertical line for 230 K 37-GHz PCT in Fig. 3a, the percentage of profiles with hail increases from around 60% for very low values of 89-GHz PCT (below about 150 K) to over 90% for values of 89-GHz PCT between 150 and 200 K and then abruptly decreases for 89-GHz PCT greater than 200 K. Consider the same 230 K 37-GHz PCT value in Fig. 3b. The very low values of 89-GHz PCT (below 150 K) correspond to an increased likelihood of graupel. This is consistent with a deep layer or high concentration of graupel having a greater scattering effect on the 89-GHz PCT channels than on the longer-wavelength 37-GHz channels. In Fig. 3c, higher values of 89-GHz PCT (greater than 225 K) are associated with an increased likelihood of snow (without graupel or hail) when the 37-GHz PCT is around 230 K.
Considering a higher 37-GHz PCT where hail is highly unlikely, a similar comparison between panels Figs. 3b–d can make some distinction between snow and graupel signatures. For 37-GHz PCT around 270 K, values of 89-GHz PCT below 250 K have an increased likelihood of graupel, and snow is favored for 89-GHz PCT greater than 260 K. The relatively few profiles identified as rain without ice have higher PCT values in both channels.
Observing the full parameter space in Fig. 3, Figs. 3a and 3b show an obvious left-right difference between high likelihood of hail for low 37-GHz PCT and higher likelihood of graupel as 37-GHz PCT increases above 250 K. The bins containing the highest percentages of snow (Fig. 3c) are shifted toward warmer 89-GHz PCT than either hail or graupel. The highest fraction of rain (Fig. 3d) occurs in the warmer Tb bins, where there is an even higher fraction of snow. Some of the highest percentages of each HID do occur in bins that contain few samples, so they should be used with caution.
The 37Diff10_19_183 parameter space in Fig. 4 is appealing because the higher probabilities for each HID type appears more distinctly separated from the others. The upper-left portion of Fig. 4a (with Diff10_19_183 greater than about 20 K and 37-GHz PCT below about 250 K) is dominated by high percentages of profiles with hail. Along the lower (Diff10_19_183 near 10 K) and rightward (37-GHz PCT between about 255–270 K) margins of that region, there is an increased likelihood of graupel in Fig. 4b. The highest likelihood of snow (Fig. 4c), other than extremely low-sampled bins, has Diff10_19_183 GHz Tb values that are near 0 K or negative. Toward the lower-right corner of the observed parameter space, some percentage of the profiles with 37-GHz PCT near 300 K and Diff10_19_183 near 0 K are identified as rain without ice, but a higher percentage is identified as having snow aloft. This frequency combination separates hail and graupel from snow and rain by identifying if the Diff10_19_183 GHz Tb values are negative or positive and also separates hail with 37-GHz PCT below about 250 K from graupel with higher 37-GHz PCT.
The percentages of profiles for each HID are shown for 89Diff183 in Fig. 5. The general pattern is somewhat similar to that for 3789 (Fig. 3), with the lower frequency channel (in this case 89 GHz) making a distinction between graupel and hail. But unlike in Figs. 3a and 4a where hail probabilities approached 100% for a concentrated portion of the parameter space, the highest percentage with hail is scattered across a large range of values for Diff183 GHz. Ignoring the lower-sampled bins, the highest percentage occurs around −15 K. The 89Diff183 combination is more sensitive to scattering by smaller ice, such as graupel, which leads to a larger area of high percentages of graupel. The highest percentage with graupel occurs at higher 89-GHz PCT (greater than 200 K) than that of hail, but occurs at Tb < −20 K for Diff183 GHz. The highest percentage with snow occurs for 89-GHz PCT between 200 and 250 K, and at Diff183 GHz Tb > 0 K. The highest percentage with rain occurs at higher 89-GHz PCT than the other HIDs, and at similar Diff183 GHz to that of snow. The bins containing the highest fraction of rain are still dominated by snow. This combination provides the ability to separate the profiles of hail and graupel from the profiles of snow and rain, with the highest sensitivity to graupel of all the channel combinations.
The percentages of profiles with hail, graupel, snow, and rain in Figs. 3–5 and Figs. S1–S3 form the basis of empirical lookup tables, for estimating the likelihood of each particle type, given an input set of measured Tb. For application beyond those Tb values encountered in the training sample, the probabilities of each HID are extrapolated to an expanded range of Tb values. The extrapolation was performed using the Python script Scipy.Interp2d.GriddedData, and a nearest neighbor approach was used for extrapolation. To avoid unreasonable results associated with small sample sizes, a minimum of 10 samples in a Tb bin (the dashed contours in Figs. 3–5 and Figs. S1–S3) were required to compute the likelihood of each HID. Furthermore, a rolling average smoothing was applied to the bins with less than 100 samples (the black contour in Figs. 3–5 and S1–S3) before extrapolating to Tb bins that had insufficient sampling. The nearest neighbor approach then extrapolates from these smoothed values inside the black contour to Tb bins located outside the black contour line. This ensures that when an unusual Tb combination is encountered (as in the case study shown later, with extremely low Tb values for an intense thunderstorm), the lookup table will provide values for the probability of each HID. On the other hand, it does also fill the parameter space with “solutions” even where Tb combinations are not physically plausible. Some of the resulting patterns in the extrapolated probabilities do look like numerical artifacts, but we do not focus on the artifacts since they primarily affect unrealistic Tb combinations.
The smoothing and sample size requirement tends to reduce the highest percentages of hail for many of the lookup tables, since the lowest Tb have the highest likelihood of hail but are not frequently observed. Similar to the raw results shown in Figs. 3–5 and Figs. S1–S3, Figs. 6 and 7 and Figs. S4–S7 show the empirical lookup tables for probabilities of hail, graupel, snow, and rain after smoothing and extrapolation. The outer portions of the plots, with Tb combinations that make little physical sense and should not be observed in precipitation (e.g., an extremely low 37-GHz PCT with a high 89-GHz PCT in the upper-left portions of each panel in Fig. 6), do catch the eye but should be disregarded or used with extreme caution. Closer to the portions of the parameter space that have been encountered by GMI, the smoothed and extrapolated lookup tables do retain the key patterns seen in the raw data. For example, Figs. 6a and 6b show a generally left-right distinction between hail with low 37-GHz PCT and graupel with higher 37-GHz PCT, as in Figs. 3a and 3b.
Extrapolated values for 3789 empirical lookup tables with contours of sample size from 1, 10, 100, 1000, and 10 000. The fraction of (a) hail, (b) graupel, (c) snow, and (d) rain after extrapolating values to bins not represented in the training data are shaded.
Citation: Journal of Applied Meteorology and Climatology 63, 7; 10.1175/JAMC-D-23-0196.1
As in Fig. 6, but for 37Diff10_19_183.
Citation: Journal of Applied Meteorology and Climatology 63, 7; 10.1175/JAMC-D-23-0196.1
In Fig. 7, hail and graupel are generally separated from snow and rain based on whether the Diff10_19_183 is positive or negative, as they were in Fig. 4. For positive values of Diff10_19_183, a 37-GHz threshold around 255 K then marks a distinction between hail being likely at lower values and graupel being likely at higher values. The extrapolation does fill in unrealistic combinations of very low 37-GHz PCT and very low Diff10_19_183 as being indicative of hail (in the lower-left portions of Figs. 7a–d); otherwise, this parameter space would have a straightforward appearance with hail in the upper-left, graupel in the upper-right, and snow in the bottom portion.
To convey the likelihood that a Tb combination represents hail, graupel, or either snow or rain in a single plot (instead of separate figure panels), the empirical lookup tables are combined into a red–green–blue (RGB) format. The value for red (ranging between 0 and 255) is scaled based on the predicted probability of hail. The value for blue is scaled based on the predicted probability of graupel. Green is scaled by the combined probability of snow and rain because the Tb combinations consistent with rain have higher percentages of profiles with snow. A set of Tb values associated with a high probability of hail would have a bright red shade, graupel would be blue, and snow or rain would be green. There can also be mixtures of colors associated with similar predicted probabilities, such as purple for an equally high likelihood of hail and graupel. The RGB versions of the lookup tables (Fig. 8 and Figs. S8–S12) are used for visual representation, and for plotting individual cases. The actual values of probability of each HID type, including the distinction between snow and rain, are retained in the lookup tables. This RGB depiction of the 3789 lookup table is shown in Fig. 8 as an example, but for examination of individual cases, we use the RGB depiction based on the means of the six retrieved probabilities for each HID type.
RGB version of 3789 lookup table. The red shade indicates a high probability of hail, blue indicates graupel, purple indicates a mix of hail and graupel, and green indicates a high probability of snow, rain, or a mix of the two. RGB lookup tables are used for visual purposes, but actual values are taken from the smoothed extrapolated lookup tables.
Citation: Journal of Applied Meteorology and Climatology 63, 7; 10.1175/JAMC-D-23-0196.1
One GPM orbit (number 1497, on 3 June 2014) from outside the training or testing sample is used to demonstrate examples of brightness temperatures in the extrapolated portions of the 3789 lookup tables. In the left panels of Figs. 9a, 9c, 9e, and 9g, black dots represent the nonraining pixels and white dots represent the raining pixels (using the GPROF probability of precipitation > 10% to distinguish these). In each row, a group of yellow pixels is highlighted, marking GMI observations from this orbit that are outside the dashed contour of the training sample’s histogram. The lookup tables for these involve extrapolation as described above. In Figs. 9a and 9b, the very low values of PCT37 and PCT89 are located inside convective cores where hail is expected. The extrapolation works exactly as intended in this case. In Fig. 9c, pixels with relatively high PCT37 but substantially lower PCT89 have mixed likelihoods of hail or graupel. These same pixels in Fig. 9d are located adjacent to convective cores, possibly in regions of anvil overhang. The signature is ambiguous, so the extrapolated lookup tables do not strongly favor any particular particle type. Very high PCT37 and PCT89 values (over 300 K) in Fig. 9e correspond to warm, arid land surfaces in Fig. 9f. A cluster of GMI pixels with PCT37 and PCT89 both around 200 K, but PCT89 > PCT37, in Fig. 9g aligns with measurements taken over surface ice near Antarctica in Fig. 9h. These bottom two rows (Figs. 9e–h) demonstrate the need to use an external precipitation/no-precipitation determination because our lookup tables always assign hydrometeor probabilities that sum to 100%.
(a),(c),(e),(f) RGB depiction of the 3789 parameter space as in Fig. 8, but with observed brightness temperatures from GPM orbit 1497 on 3 Jun 2014. Black dots have GPROF PoP < 10%, white dots have PoP > 10%, and yellow dots highlight selected GMI pixels corresponding to extrapolated portions of the 3789 lookup tables. The red shade is likely hail in the lookup table, the blue shade is likely graupel, and the green shade is likely snow or rain. (b),(d) RGB depiction of a mesoscale convective system in Nebraska and Iowa, observed by this GPM orbit. Yellow dots are those marked in (a) and (c). (f) The full orbit, with yellow in the United States, Mexico, and Arabian Peninsula, and Horn of Africa for the yellow marked in (e). (h) The full orbit, with yellow near Antarctica for the yellow marked in (g).
Citation: Journal of Applied Meteorology and Climatology 63, 7; 10.1175/JAMC-D-23-0196.1
5. Application and testing of empirical lookup tables
The empirical lookup tables depicted in Figs. 6 and 7 and Figs. S4–S7 are applied to an independent, 2 years (2021 and 2022) testing dataset. The testing data contains the collocated GMI Tb values and dual-polarization radar-based HIDs from the GVN. Observations in the testing dataset are binned based on their GMI-predicted probability of each HID, which is compared to the observed percentage having that HID according to the polarimetric radar retrieval from the GVN. Such comparisons are made for each of the six channel combinations separately in Figs. S13–S18. Since using six different retrievals of probabilities of four different HIDs becomes unwieldy, we take the arithmetic mean of the six retrievals for each HID type, given an input set of observed Tb.
For example, in Fig. 10a, among the GMI pixels with a mean prediction of 35%–40% likelihood of hail, slightly over 40% are identified as hail by the radar-based retrieval, and nearly 80% have either hail or graupel, and a small portion are identified as snow. For most hail probability bins in Fig. 10a, the likelihood of hail is slightly underestimated by our GMI-based approach (the red bars surpass the dashed one-to-one line). For these bins of 5% increments in GMI-predicted probability of hail, there is a 0.98 linear correlation with the observed percentages of pixels identified as hail by the GVN, −11% bias, and 11.4% mean absolute error. For bins where the GMI-predicted probability of hail exceeds 10% in Fig. 10a, a very high percentage—usually around 80% or more—of the observations are identified as having either hail or graupel by the GVN. When the GMI-predicted probability of hail is less than 10%, the GVN mostly identifies snow. So besides the GMI-based lookup tables providing an accurate probabilistic diagnosis of hail, the 10% probability of hail threshold can also be thought of as indicating high confidence that there is at least graupel present.
Observed vs predicted HIDs from the mean of the six lookup tables. Bars represent the fraction of pixels observed as hail (red), graupel (blue), snow (orange), or rain (green) for each 5% bin of GMI-predicted probability of (a) hail, (b) graupel, (c) snow, or (d) rain. The gray dashed line is a 1:1 line, which would indicate a perfect prediction. If the bars are higher than the 1:1 line, the predicted probability is lower than observed HID percentages. Where the bars do not reach the line, the predicted probabilities are higher than observed HID percentages. Black dots represent the sample size within each bin.
Citation: Journal of Applied Meteorology and Climatology 63, 7; 10.1175/JAMC-D-23-0196.1
The GMI-based probabilistic retrievals of graupel in Fig. 10b also correlate very well with the radar-based identifications of graupel, with a 0.99 correlation coefficient, −3.8% bias, and 3.8% mean absolute error. When the GMI-predicted probability of graupel is below 40%, most of the GVN radar-based retrievals indicate snow instead. As the GMI-predicted probability of graupel increases, the fraction of observations with hail also slowly increases, and the fraction with only snow or rain decreases. Just as the graupel percentages in Figs. 3b and 8b rarely approached 80%, the GMI-based retrievals never identify greater than 75% likelihood of graupel (without hail).
The GMI-based retrievals of snow (Fig. 10c), that is, profiles with frozen particles in the column but not hail or graupel, performs similarly to graupel, with decreasing percentages of hail and graupel with increasing prediction of snow. The correlation between the predicted and observed portion of snow profiles is 0.99. Overall, the bias for the snow probabilities is 1% (a very slight overestimation of snow occurrence, in contrast to the underestimation of graupel and hail occurrence), and the mean absolute error is 3.3%. When the measured Tb lead to an extremely low GMI-retrieved probability of snow, there is usually hail or graupel present instead.
Recalling that liquid rain is present in most profiles, but that our rain category specifically indicates an absence of ice signatures aloft, the rain category comprises a small portion of our training sample. As such, the predicted probability of rain rarely exceeds 15% and it is difficult to say much about verification of this category. The vast majority of profiles in the testing data have less than 5% predicted likelihood of rain, and of those, only 0.1% do belong to the rain category. A prediction of 5%–10% likelihood of rain turns out to be an overprediction, with only 4.2% of those profiles belonging to the rain category. On the other hand, profiles predicted to have a 10%–15% likelihood of rain are diagnosed more accurately, with 12.6% falling into the rain category.
6. Mesoscale convective system example
The 3 June 2014 Nebraska–Iowa mesoscale convective system from Figs. 9b and 9d is examined more closely here. The predicted probabilities are then scaled by the GPROF probability of precipitation to reduce the values for nonprecipitating profiles. The corresponding Tb combinations used as inputs to the lookup tables are in Fig. 12. The predicted HIDs identify several areas with hail, with very low values of PCT37, PCT89, and V166 and high values of Diff10_19_183. These align well with areas where hail is retrieved by the ground-based dual-polarization radars in Fig. 13, although we diagnose a larger area of hail from GMI than from ground-based radar. The composite HID in Fig. 13 was derived by applying the same hierarchy as the GMI-based HID for all available elevation angles and assigning a single HID to each vertical profile at the radar’s native resolution. If the two ground-based radars identified different HIDs for a similar location, the one highest in the hierarchy was used. The HID figure was created using Colorado State University (CSU) RadarTools and Python ARM radar toolkit (PyArt) (Helmus and Collis 2016).
There is a large region identified by the GMI lookup tables as a hail, graupel mix in Fig. 11, and several areas identified as rain or snow. The areas diagnosed as likely hail, graupel mix have larger values of Diff166 (polarization difference) and Diff183 (greater ice scattering in a wideband channel than nearer a water vapor line) in Fig. 12, compared to those diagnosed as likely hail, snow, or rain. The ground-based radar retrieval in Fig. 13 depicts a much smaller region of graupel than suggested by GMI in Fig. 11, although the shades of blue-green and dark purple in Fig. 11 indicate uncertainty in distinguishing between graupel, snow, rain, and hail. Much of the area with blue-green shades in Fig. 11 is retrieved as vertically oriented ice by the dual-polarization radars in Fig. 13.
GMI-predicted HIDs for orbit 1497, which was primarily located over Nebraska and Iowa. This orbit was observed on 3 Jun 2014. Colors are as in Figs. 9 and 10: the red shade represents a high likelihood of hail, blue represents graupel, purple represents a graupel/hail mix, and green represents rain, snow, or a mix of the two. Dark regions indicate a low PoP.
Citation: Journal of Applied Meteorology and Climatology 63, 7; 10.1175/JAMC-D-23-0196.1
GMI-observed Tb for orbit 1497 (Fig. 11). The blue shade represents colder Tb values, and the red shade represents higher Tb values. The channels shown are (a) PCT37, (b) PCT89, (c) V166, (d) Diff166, (e) Diff183, and (f) Diff10_19_183.
Citation: Journal of Applied Meteorology and Climatology 63, 7; 10.1175/JAMC-D-23-0196.1
Radar-derived composite HID for similar time and location as the GMI HIDs from Fig. 12, near 2140 UTC 3 Jun 2014. KOAX (Omaha, NE) and KDMX (Des Moines, IA) HID derived using CSU RadarTools and plotted using PyArt.
Citation: Journal of Applied Meteorology and Climatology 63, 7; 10.1175/JAMC-D-23-0196.1
The regions identified as likely snow or rain (green shades) in Fig. 11 tend to align with vertically oriented ice or snow in the dual-polarization radar retrieval (Fig. 13). These regions have GMI signatures suggestive of precipitation in all the channel combinations in Fig. 12, without having particularly strong signatures in any of those channel combinations. In this warm-season convective case over subtropical land, there is hardly any liquid rain that lacks ice precipitation aloft—the narrow ring of green shades surrounding the precipitation area in Fig. 13 is mostly a plotting artifact.
7. Discussion–interpretation of HID signatures
A key reason for using empirical lookup tables instead of more advanced machine learning approaches is the desire to retain straightforward physical interpretation of the relationships between the Tb and hydrometeor type. The simplest relationship to interpret from the lookup tables is that very low Tb in the lower frequency channels tend to result from scattering of the upwelling radiation by hail. This is not at all surprising based on prior research, but the quantitative relationships involving combinations of microwave channels is a new result of this research. With very low Tb being indicative of hail, a first-order expectation would be that somewhat higher Tb values in precipitation are associated with graupel. The multichannel analysis here refines that expectation, with Fig. 3b showing that a low PCT89 (150 K) paired with a substantially higher PCT37 (250 K) can be a signature of graupel, since a deep column or high concentration of millimeters-sized graupel can produce a more robust signature in the 89-GHz (3.3 mm) channel than in the 37-GHz (8 mm) channel. Conversely, a higher PCT89 value (200 K) paired with increased scattering in the 37-GHz channel (with PCT37 235 K) can suggest hail (Fig. 3a), even though such a high PCT89 value by itself would not distinguish between hail and graupel signatures (Fig. 1d).
Incorporating the higher-frequency GMI channels helps distinguish between graupel and lower-density snow, or profiles without detectable precipitation ice. In Figs. 2c and 5, for example, negative values of Diff183 tend to correspond to hail or graupel profiles and positive values correspond to higher fractions of snow or rain. The negative values of Diff183 result from a lower Tb in the wideband 183.3 ± 7 GHz channel than in the 183.3 ± 3 GHz water vapor channel because of strong ice scattering in both but emission (warmer Tb) in the water vapor channel. The positive values of Diff183 for snow and rain profiles result from warmer emission by the land surface (without sufficient ice aloft) in the wideband channel, giving a higher Tb than the emission by colder water vapor aloft in the water vapor channel. With this distinction between (hail or graupel) versus (snow or rain) made by the high-frequency 183-GHz channels, lower frequency (19, 37, or 89 GHz) channels can help make the (hail vs graupel) and (snow vs rain) distinctions, although our lookup tables do have substantial overlap between snow and rain.
Distinction between liquid and solid hydrometeors for passive microwave radiometers benefits from having a water background surface, and our database is primarily located over land with high emissivity. This limits the capability of distinguishing between the two. This frozen versus liquid distinction we are making does not necessarily mean it is occurring at the surface, so our approach is not well suited for winter precipitation-type application.
This research considers the dual-pol derived HIDs to be the “truth.” The radar-based retrievals are widely available, but there are uncertainties within the derived HIDs. These uncertainties, along with the results from Leppert and Cecil (2015), influenced our decision to combine high- and low-density graupel and small ice classifications into more broad classes.
8. Conclusions
Using GMI Tb measurements and GV ground-based HID data, an empirical lookup table approach was used to predict the probability of specific HIDs being located within a vertical profile. The categories are profiles with hail (and any other hydrometeors), profiles with graupel but no hail, profiles with snow but no hail or graupel, and profiles with rain but no detectable precipitation ice aloft (Table 1). Using single frequencies, the percentage of profiles with each HID as a function of Tb was found. The results show that 19- and 37-GHz PCT are the best for identifying profiles with hail, with the percentage of observed hail increasing rapidly with decreasing Tb. The 89- and 166-GHz frequencies are better suited to separate profiles of hail and graupel from profiles consisting of snow or rain, with the possibility to differentiate between hail and graupel. The relationships Diff166 and Diff183 GHz are useful for separating profiles containing hail and graupel from the remaining HIDs, but with little distinction between hail and graupel. The relationship Diff10_19_183 GHz is useful for identifying profiles containing hail and graupel, with increased likelihood of hail with increasing value. There is little distinction between profiles containing higher percentages of snow and rain. Many of the frequencies indicate the ability to identify the presence of larger ice hydrometeors, but there is limited ability to separate profiles of snow and rain.
Based on the single frequency results, we identify and test many multifrequency relationships, which are chosen to differentiate between the HID species. The multifrequency approach led to the development of six empirical lookup tables, and the results are extrapolated to expand the range of Tb values represented by the lookup tables. The lookup tables were then used to test the GMI HID retrieval on 1 year (2018) of independent GVN matched data. The percentage of observed HIDs was compared to the GMI-predicted probability of each HID, and statistical analysis was performed. The results show the GMI-prediction performs with good accuracy (Fig. 10). The correlation between the predicted and observed percentages of profiles with hail and graupel both exceed 0.98. The mean absolute errors in the predicted probabilities of hail and graupel are 11% and 3%, respectively, and both have negative biases (−11% and −3%, respectively), indicating overall underprediction of hail and graupel. Despite the statistical tendency for underprediction of hail in the full validation dataset, analysis of a strong mesoscale convective system (Figs. 11–13) and other cases not shown does indicate a tendency for the area of hail cores to be overestimated in strong convective systems.
The approach for matching GMI measurements to HID retrievals from dual-polarization radar only required a single pixel of the specified HID type, anywhere in the vertical profile. Profiles identified as hail almost always have a greater number of high-resolution radar gates with other HID types elsewhere within the same GMI footprint, than the number of gates identified as hail. Future work may include requiring a greater number of radar gates with a given HID type during the observational GMI HID matching. This would probably reduce the size of the predicted hail cores in cases such as Fig. 11, but at the expense of exacerbating the underprediction of hail that is already present in the lookup tables.
The empirical lookup tables struggle to distinguish between profiles with snow aloft and those with only rain (no detectable precipitation ice aloft). The predicted probabilities of rain are never greater than those of snow. Because of this, these two categories are combined when plotting individual cases (Fig. 11), although the individual probabilities of the two categories are retained. The occurrence of snow profiles is overestimated by 1% in the validation.
The training and testing samples were entirely from cases near ground-based radars in the United States. Analysis of subsets of those cases (not shown here) did not suggest differences in accuracy as a function of location or season, but the relevance of these lookup tables to precipitation regimes very different from those in the United States sample cannot be evaluated. When separating the dataset by location and season, the Tb values most commonly associated with each HID remained similar, but the frequency in which specific Tb values were observed increased or decreased according to the frequency of precipitation and convection. It is likely that inclusion of more island or coastal radar sites could improve the distinction between snow and rain profiles, at least for scenes over water surface types.
The GMI HID lookup tables derived here are available in the supplemental material, for application to individual cases as in Fig. 11 or other types of analysis. A major caveat is that these lookup tables do not distinguish precipitating from nonprecipitating profiles. They were developed from a training database entirely composed of profiles with precipitation. In Figs. 9 and 11, we multiply the predicted probability of each HID type by the GPROF probability of precipitation to account for this.
The GMI-based HID retrieval was applied to an individual orbit of level 1C GMI Tbs. The case shown here was chosen to showcase the ability to identify varying storm features, within close proximity to ground-based radars for comparison. The results show the GMI HID retrieval is capable of identifying realistic storm structures and often display similar features represented by the ground-based radar-derived composite HIDs. The large GMI footprints are not capable of identifying the small-scale details shown by the ground-based radar but allows for HID analysis for near-global coverage. Based on these results, the GMI HID retrieval will be useful for studying the global distribution and seasonality of precipitation.
Acknowledgments.
This research work was supported by NASA’s Precipitation Measurement Mission Science Team. The GPM Validation Network data are processed by NASA Wallops Flight Facility and NASA Marshall Space Flight Center, supported by the GPM Ground Validation program.
Data availability statement.
The GMI calibrated brightness temperatures and the 2A GPROF data can be found by creating an account and downloading the data at https://arthurhouhttps.pps.eosdis.nasa.gov. The matched GMI-ground radar data can be found at https://pmm-gv.gsfc.nasa.gov/pub/gpm-validation/data/gpmgv/netcdf/geo_match/GPM/GMI/2AGPROF/ and is maintained by the GPM Ground Validation team. The empirical lookup tables are available in the supplemental material as S19–S42. The lookup tables, S19–S42, use small font to contain them on individual pages, but the text can be copied and pasted into a spreadsheet by potential users. The level 2 NEXRAD data used for ground-radar HID derivation can be found at https://www.ncei.noaa.gov/access/search/data-search/weather-radar-level-ii.
REFERENCES
Bang, S. D., and D. J. Cecil, 2019: Constructing a multifrequency passive microwave hail retrieval and climatology in the GPM domain. J. Appl. Meteor. Climatol., 58, 1889–1904, https://doi.org/10.1175/JAMC-D-19-0042.1.
Berg, W., 2022: GPM GMI common calibrated brightness temperatures collocated L1C 1.5 hours 13 km, v07. Goddard Earth Sciences Data and Information Services Center (GES DISC), accessed 10 September 2023, https://doi.org/10.5067/GPM/GMI/GPM/1C/07.
Cecil, D. J., 2009: Passive microwave brightness temperatures as proxies for hailstorms. J. Appl. Meteor. Climatol., 48, 1281–1286, https://doi.org/10.1175/2009JAMC2125.1.
Cecil, D. J., and C. B. Blankenship, 2012: Toward a global climatology of severe hailstorms as estimated by satellite passive microwave imagers. J. Climate, 25, 687–703, https://doi.org/10.1175/JCLI-D-11-00130.1.
Cecil, D. J., and T. Chronis, 2018: Polarization-corrected temperatures for 10-, 19-, 37-, and 89-GHz passive microwave frequencies. J. Appl. Meteor. Climatol., 57, 2249–2265, https://doi.org/10.1175/JAMC-D-18-0022.1.
Dolan, B., and S. A. Rutledge, 2009: A theory-based hydrometeor identification algorithm for X-band polarimetric radars. J. Atmos. Oceanic Technol., 26, 2071–2088, https://doi.org/10.1175/2009JTECHA1208.1.
Ferraro, R., J. Beauchamp, D. Cecil, and G. Heymsfield, 2015: A prototype hail detection algorithm and hail climatology developed with the advanced microwave sounding unit (AMSU). Atmos. Res., 163, 24–35, https://doi.org/10.1016/j.atmosres.2014.08.010.
Gatlin, P. N., M. Thurai, V. N. Bringi, W. Petersen, D. Wolff, A. Tokay, L. Carey, and M. Wingo, 2015: Searching for large raindrops: A global summary of two-dimensional video disdrometer observations. J. Appl. Meteor. Climatol., 54, 1069–1089, https://doi.org/10.1175/JAMC-D-14-0089.1.
Gatlin, P. N., W. A. Petersen, J. L. Pippitt, T. A. Berendes, D. B. Wolff, and A. Tokay, 2020: The GPM Validation Network and evaluation of satellite-based retrievals of the rain drop size distribution. Atmosphere, 11, 1010, https://doi.org/10.3390/atmos11091010.
Gong, J., and D. L. Wu, 2017: Microphysical properties of frozen particles inferred from Global Precipitation Measurement (GPM) Microwave Imager (GMI) polarimetric measurements. Atmos. Chem. Phys., 17, 2741–2757, https://doi.org/10.5194/acp-17-2741-2017.
Helmus, J. J., and S. M. Collis, 2016: The Python Arm Radar Toolkit (Py-ART), a library for working with weather radar data in the Python programming language. J. Open Res. Software, 119, e25, https://doi.org/10.5334/jors.119.
Hong, G., G. Heygster, J. Miao, and K. Kunzi, 2005: Sensitivity of microwave brightness temperatures to hydrometeors in a tropical deep convective cloud system at 89–190 GHz. Radio Sci., 40, RS4003, https://doi.org/10.1029/2004RS003129.
Hou, A. Y., and Coauthors, 2014: The global precipitation measurement mission. Bull. Amer. Meteor. Soc., 95, 701–722, https://doi.org/10.1175/BAMS-D-13-00164.1.
Kummerow, C., W. Barnes, T. Kozu, J. Shiue, and J. Simpson, 1998: The Tropical Rainfall Measuring Mission (TRMM) sensor package. J. Atmos. Oceanic Technol., 15, 809–817, https://doi.org/10.1175/1520-0426(1998)015<0809:TTRMMT>2.0.CO;2.
Kummerow, C. D., D. L. Randel, M. Kulie, N.-Y. Wang, R. Ferraro, S. Joseph Munchak, and V. Petkovic, 2015: The evolution of the Goddard profiling algorithm to a fully parametric scheme. J. Atmos. Oceanic Technol., 32, 2265–2280, https://doi.org/10.1175/JTECH-D-15-0039.1.
Leppert, K. D., II, and D. J. Cecil, 2015: Signatures of hydrometeor species from airborne passive microwave data for frequencies 10–183 GHz. J. Appl. Meteor. Climatol., 54, 1313–1334, https://doi.org/10.1175/JAMC-D-14-0145.1.
Leppert, K. D., II, and D. J. Cecil, 2019: Sensitivity of simulated GMI brightness temperatures to variations in particle size distributions in a severe hailstorm. J. Appl. Meteor. Climatol., 58, 1905–1930, https://doi.org/10.1175/JAMC-D-19-0031.1.
Mroz, K., A. Battaglia, T. J. Lang, D. J. Cecil, S. Tanelli, and F. Tridon, 2016: Hail-detection algorithm for the GPM core observatory satellite sensors. J. Appl. Meteor. Climatol., 56, 1939–1957, https://doi.org/10.1175/JAMC-D-16-0368.s1.
Peterson, W. A., 2021: GPM Ground Validation Validation Network (VN). NASA Global Hydrometeorology Resource Center DAAC, accessed 11 June 2023, https://doi.org/10.5067/GPMGV/WFF/MULTIPLE/DATA101.
Schwaller, M. R., and K. R. Morris, 2011: A ground validation network for the global precipitation measurement mission. J. Atmos. Oceanic Technol., 28, 301–319, https://doi.org/10.1175/2010JTECHA1403.1.
Spencer, R. W., and D. A. Santek, 1985: Measuring the global distribution of intense convection over land with passive microwave radiometry. J. Climate Appl. Meteor., 24, 860–864, https://doi.org/10.1175/1520-0450(1985)024<0860:MTGDOI>2.0.CO;2.
Spencer, R. W., W. S. Olson, W. Rongzhang, D. W. Martin, J. A. Weinman, and D. A. Santek, 1983: Heavy thunderstorms observed over land by the Nimbus 7 scanning multichannel microwave radiometer. J. Climate Appl. Meteor., 22, 1041–1046, https://doi.org/10.1175/1520-0450(1983)022<1041:HTOOLB>2.0.CO;2.
Spencer, R. W., M. R. Howland, and D. A. Santek, 1987: Severe storm identification with satellite microwave radiometry: An initial investigation with Nimbus-7 SMMR data. J. Climate Appl. Meteor., 26, 749–754, https://doi.org/10.1175/1520-0450(1987)026<0749:SSIWSM>2.0.CO;2.
Spencer, R. W., H. M. Goodman, and R. E. Hood, 1989: Precipitation retrieval over land and ocean with the SSM/I: Identification and characteristics of the scattering signal. J. Atmos. Oceanic Technol., 6, 254–273, https://doi.org/10.1175/1520-0426(1989)006<0254:PROLAO>2.0.CO;2.
Wilheit, T. T., A. T. C. Chang, M. S. V. Rao, E. B. Rodgers, and J. S. Theon, 1977: A satellite technique for quantitatively mapping rainfall rates over the oceans. J. Appl. Meteor., 16, 551–560, https://doi.org/10.1175/1520-0450(1977)016<0551:ASTFQM>2.0.CO;2.
