Separation of Stratiform and Convective Rain Types Using Data from an S-Band Polarimetric Radar: A Case Study Comparing Two Different Methods ^†

Abstract

Data from an S-band polarimetric radar located at a mid-latitude, coastal location are used to compare two different methods for identifying stratiform and convective rain regions. The first method entails the retrievals of two (main) parameters of the rain drop size distributions using the radar reflectivity and the differential reflectivity. The second technique is a well-known texture-based method which utilizes the radar reflectivity and its spatial variability. A widespread event with embedded line convection was used as a test case. The two methods were compared using 500 m by 500 m pixel resolution gridded data constructed from the radar volume scans. Only 12% of the pixels showed disagreement between the two methods.

1. Introduction

There are several different methods available in the literature to classify and separate stratiform and convective rain types. These include (a) the use of radar reflectivity ‘texture’ from weather radar scans, e.g., [1], (b) using ground in situ measurements including surface disdrometers, e.g., [2,3], (c) using profiler observations, e.g., [4], and (d) based on the magnitude of up- and downdrafts, e.g., [5]. Additionally, the estimated or retrieved characteristics of rain drop size distributions (DSD) have also been used to identify the two rain types [6,7], and were compared with the reflectivity-texture based method in [8,9] using C-band radar observations in Darwin, Australia, which is a tropical region. In this paper, we perform similar comparisons between the same two methods (i.e., between the reflectivity texture-based method and the DSD-based method), but in the current study, we utilize data from an S-band polarimetric radar (NPOL, e.g., [10]) at a mid-latitude coastal region, namely Wallops Island, Delmarva peninsula, USA.

In a very recent paper [11], the DSD-based separation method was tested using data from two collocated disdrometers based at the Wallops site. The two disdrometers were (i) a Meteorological Particle Spectrometer (MPS; [12]), which provided accurate measurements of drop concentrations for small and tiny drops (particularly for drop diameters below 1.2 mm), and (ii) a 2D video disdrometer (2DVD; [13,14]) for drop concentrations above 1 mm. By combining both sets of measurements, the full DSD spectra were constructed and a rain-type classification was made using 1- and 3-min DSDs. For each of these DSDs, the mass-weighted mean diameter, D_m, and the normalized intercept parameter, N_W, were derived, and depending on where each point was in the N_W versus D_m domain, the corresponding rain type was determined. The classification was then ‘visually’ compared against RHI scans from the NPOL radar, made over the disdrometer site. Over 20 h of 1- and 3-min DSD data (and the corresponding RHI scans) were used for verification.

As a follow on, in this paper, we examine the application of the same basic technique to the S-band NPOL radar data. An improved retrieval method is used to estimate N_W and D_m from the radar reflectivity and differential reflectivity measurements. Gridded data from −60 to +60 km along north-to-south and east-to-west are used at various altitudes from 1000 m to 3000 m above sea level, in steps of 500 m. The (constant-altitude) gridded data were constructed from the NPOL radar volume scans taken during a relatively wide-spread event with embedded line convection which passed over the Wallops site on 30 April 2020. The lowest gridded level was used for pixel-by-pixel comparison against the reflectivity texture-based method from [1].

2. Estimating N_W and D_m from NPOL Radar Data

The equations used for retrieving (or estimating) N_W and D_m from the NPOL reflectivity and differential reflectivity (Z_h and Z_dr, respectively) were previously given in [11], and hence, they are briefly described here. The estimation of D_m is a two-step procedure, the first step involving the estimation of an intermediary parameter, D_m’, which depends on two chosen reference moments [15]. If we denote these as M_i and M_j, then D_m’ is given by:

$${D_m}^{’}={\left(\frac{M_j}{M_i}\right)}^{\frac{1}{\left(j-i\right)}}$$

(1)

At S-band, simulations have shown that D_m’ can be directly estimated from Z_dr within reasonable/acceptable accuracy. Figure 1a shows the simulation points using 3-min measured DSDs (i.e., full DSD spectra, but from two different locations, Huntsville, Alabama, which is a sub-tropical region, and Greeley, Colorado, which is a mid-latitude continental climate) as well as the fitted curve, given by:

$$D_m^{’}=0.0822 Z_{dr}^3-0.4841 Z_{dr}^2+1.7515 Z_{dr}+0.628$$

(2)

Next, we determine D_m from D_m’. Once again, S-band simulations have shown that this can be done within good accuracy, as shown in panel (b) of Figure 1. The fitted curve is given by:

$$D_m=0.7977D_m^{’}+0.0883$$

(3)

Note that both panels (a) and (b) show a monotonic increase; hence, no ambiguity will arise when using Equations (2) and (3).

The third and final step is to determine N_W. As shown in [11], N_W is given by:

$$N_W=\left(\frac{4^4}{6}\right){N_0}^{’}$$

(4)

where:

$${N_0}^{’}=M_i{}^{\frac{\left(j+1\right)}{\left(j-i\right)}} M_j{}^{\frac{\left(i+1\right)}{\left(i-j\right)}}$$

(5)

To derive N_W, we need to make use of both the retrieved D_m as well as the radar reflectivity. Panel (c) shows the variation of N_W/Z_h(linear) versus D_m, where Z_h(linear) is the radar reflectivity in linear units. The fitted curve (monotonic decrease) is given by:

$$\frac{N_W}{Z_{h \left(linear\right)}}=39.446D_m^{’}{}^{-6.839}$$

(6)

In summary, Equations (2), (3), and (6) are used to estimate N_W and D_m for each of the gridded pixels from the S-band scans.

3. NPOL Data and the Event on 30 April 2020

On 30 April 2020, a slow-moving cold front passed over the WFF region. A NW/SE-oriented line of strong convection with heavy rain moved through the region. Reflectivity within the line was as high as 60 dBZ in areas. The convective line was embedded in stratiform with reflectivity in the range 25 to 35 dBZ. As recorded in NASA rain gauges at Wallops, approximately 20 mm accumulated between 19 to 22 h UTC, the majority of which fell within 30 min associated with the convective line.

Figure 2a,b show the gridded data constructed from the volume scans taken at 21:05 UTC: (a) reflectivity and (b) differential reflectivity, both at 1000 m above sea level with a pixel resolution of 500 m × 500 m. The line convection shows reflectivities as high as 55 to 60 dBZ and differential reflectivities of >2 dB in some regions. Note also that, at azimuths of around 170°, some beam-blockage problems exist (which need to be omitted from the classification procedure).

Panels (c) and (d) of Figure 2 show the estimated N_W and D_m, respectively, derived using Equations (2), (3), and (6). Once again, the effect of beam-blockage at ~170° azimuth is evident in the retrievals. Within the line convection itself, D_m values > 2 mm are seen in some regions.

4. Rain Type Classification

The gridded data ranging from −60 km to +60 km both in the north-south and the east-west directions were extracted and the classification based on the N_W-D_m values for each pixel was determined. As mentioned in [11], “a simple ‘index’ parameter, i (empirically derived), was used to indicate whether the N_W versus D_m lie above or below the separation line”. The value of i for each pixel is given by:

$$i=lo{g}_{10}\left(N_W^{est}\right)-lo{g}_{10}\left(N_W^{sep}\right)$$

(7)

where:

$$lo{g}_{10}\left(N_W^{sep}\right)=c_1{D}_m^{est}+c_2$$

(8)

In Equation (7), N_W^est is the estimated N_W for the specific pixel and D_m^est the (corresponding) estimated D_m. Note that, in [11], Equations (7) and (8) were applied to disdrometer-based DSD data. Note also that “values of c₁ and c₂ may vary somewhat depending on the location, but to be consistent with our previous study, they were set to −1.682 and 6.541, respectively”. If, for a given pixel, i is negative, it is categorized as stratiform rain, and when i is positive, it is categorized as convective rain. Additionally, we introduce another category, namely, ‘Mixed’ (or Uncertain/Transition), when the magnitude of i is less than 0.05, i.e., −0.05 ≤ i ≤ 0.05. Such a category was initially introduced in [16] based on measurements from large squall lines with trailing stratiform rain. The disdrometer measurements showed that they were able to identify the convective line and the stratiform rain areas quite easily but there was a transition region between the convective and the stratiform which showed a different Z-R than for pure convection versus pure stratiform. We use the term ‘transition’ here even if the storm type does not belong to the squall lines described in [16]. Based on the latter, a third category was introduced in [8], using the C-band CPOL radar data from Darwin, Australia (although in that study, a wider range for i was used, viz. −0.1 ≤ i ≤ 0.1). The third category has been termed transition, mixed or uncertain.

Panel (a) of Figure 3 shows the classification for the 1000 m altitude gridded data in Figure 2. The orange/red color represents convective rain category, and the cyan/turquoise color for stratiform rain. The purple color represents the mixed category, which appears to be predominantly in regions surrounding convective rain areas. By comparison, panel (b) of Figure 3 shows the texture-based classification from the texture-based method. Both plots show somewhat similar features, but to compare the classifications on a pixel-to-pixel basis, we show in panel (c) the matched/mismatched pixels. The colors represent the following:

Light blue/cyan	when both methods classify as stratiform rain.
Red	when both methods classify as convective rain.
Orange	when the DSD-based method classifies as convective rain
	and the texture method as stratiform rain.
Green	when the DSD-based method classifies as stratiform rain
	and the texture method as convective rain.
Purple	when the DSD-based method classifies as mixed type.
Black	when Zdr is <0 dB, which is omitted from the classification procedure.

In terms of percentage pixels, our comparison resulted in: (a) 56% of the radar pixels being categorized as stratiform rain by both methods; (b) 21% as convective rain by both methods; and (c) a further 11% as the ‘mixed’ category from the DSD-based method. For the remaining 12% of the pixels, there was disagreement between the two methods, which largely occurred in regions adjacent to (b).

A small but significant improvement was obtained when co-polar attenuation and differential attenuation were included in the radar-data correction procedures. Although at S-band, the attenuation effects are largely negligible, it was found that for this particular case event, the differential propagation phase shift along certain azimuth angles (e.g., 220°) were sufficiently high to cause around 0.5 dB co-polar attenuation beyond the line convection. After applying the correction procedures, it was found that (a) 59% of the radar pixels were categorized as stratiform rain by both methods; (b) 20% as convective rain by both methods; and (c) a further 10% as the ‘mixed’ category from the DSD-based method. The percentage of ‘mismatched’ pixels reduced from 12% to 11%.

One of the main uses of the DSD-based separation technique is that it enables the ratios of the stratiform rain volume to convective rain volume to be determined from the gridded data. Such an analysis is very important from the viewpoint of latent heat estimation, since convective and stratiform rain have different heating rates [17]. Furthermore, since large-scale numerical weather prediction models assume different ratios of convective to stratiform area and rain volume, the retrieved products demonstrated here will form an important validation tool.

5. CFADs

Contoured Frequency-by-Altitude Diagrams (CFADs; [18]) are useful for examining vertical structures. For the Figure 2 case event, these were constructed separately for stratiform and convective rain regions (after applying the DSD-based separation), as well as for mixed precipitation. They are shown in Figure 4. The left panels show the reflectivity contours and the right panels show the differential reflectivity contours. One important aspect to note is that for convective rain (middle panels), both Z_h and Z_dr decrease from 3 down to 1 km, which in turn indicates that drop break-up is the dominant process [19]. On the other hand, for stratiform rain (top panels), Z_h is almost uniform from 3 to 1 km a.g.l. but Z_dr decreases, indicating, once again, the possible occurrence of drop break process but also indicating an increase in number concentration (per unit volume). These features were also observed in the stratiform rain regions of the outer rain-bands of Category-1 Hurricane Dorian (again over Wallops; [20]). A 1D MonteCarlo microphysical model using the super-particle concept (named McSnow; [21]) together with radiosonde data as model input also showed the importance of drop break-up even in light to moderate rain rates, being consistent with the radar observations [20].

The corresponding CFADs for N_W and D_m are shown in Figure 5. The freezing height on this day was at around 3 km; hence, the retrieved N_W and D_m should be neglected above this height. In the rain region below, one main feature to be noted is that for both stratiform and convective rain, D_m decreases from 3 to 1 km, but for the latter, the rate of decrease (with decreasing height) is noticeably higher, indicating that the break-up is more severe. For mixed precipitation, the rate of decrease is more similar to that for convective rain.

6. Summary

The case study considered here (of line convection embedded within a larger system) has clearly shown that there is considerable agreement between the texture-based method in [1] and the DSD-based method in [6,8]. Only 12% of the gridded radar data pixels showed a classification mismatch. When a simple attenuation-correction method (based on differential propagation phase shift) was applied to the S-band data, the percentage of mismatch reduced to 11%.

The DSD-based method utilized previously derived retrievals for the two DSD parameters N_W and D_m. Stratiform and convective rain was based on where the N_W-D_m points lay in relation to the well-established separation line. Although the separation line was determined based on disdrometer data, we have shown that it can also be used for the gridded S-band NPOL radar data. A third category was also introduced to represent the mixed region. They tended to be in areas immediately surrounding the convective rain regions.

Contoured Frequency-by-Altitude Diagrams (CFADs) were also generated for the stratiform and convective rain separately. They indicate drop break-up to be a dominant process in the rain region below the freezing height. The rate of decrease (with decreasing height) of D_m was higher in the case of convective rain, implying that the drop break-up is more severe.

One caveat of the DSD-based technique is that there are a number of error sources which need to be considered when applying this technique. These include (i) radar measurement errors; (ii) retrieval algorithm errors (for example the scatter seen in Figure 1); and (iii) small uncertainties in the assumed separation line. Out of these, (i) is likely to be the most significant source of error. Nevertheless, from this case study, it seems likely that the DSD-based technique can be used for the S-band NPOL gridded data with reasonable accuracy. Furthermore, it can be applied not just to the lowest gridded data but also to higher altitudes, i.e., up to the nominal freezing height.