The Application of the Convective–Stratiform Classification Algorithm for Feature Detection in Polarimetric Radar Variables and QPE Retrieval During Warm-Season Convection

Abstract

Feature detection is one of the hot topics in the weather radar research community. This study employed a convective–stratiform classification algorithm to detect features in polarimetric radar variables and Quantitative Precipitation Estimation (QPE) retrieval during a heavy precipitation event in Crossville, Tennessee, during warm-season convection. Analysis of polarimetric radar variables revealed that strong updrafts, mixed-phase precipitation, and large hailstones in the radar resolution volume during the event were driven by the existence of supercell thunderstorms. The results of feature detection highlight that the regions with convective–stratiform cores and strong–faint features in the reflectivity field are similar to those in the rainfall field, demonstrating how the algorithm more effectively detects features in both fields. The results of the estimates, accounting for uncertainty during feature detection, indicate that an offset of +2 dB overestimated convective features in the northeast in both the reflectivity and rainfall fields, while an offset of −2 dB underestimated convective features in the northwest part of both fields. The results highlight that convective cores cover a small area with high rainfall exceeding 50 mmh⁻¹, while stratiform cores cover a larger area with greater horizontal homogeneity and lower rainfall intensity. These findings are significant for nowcasting weather, numerical models, hydrological applications, and enhancing climatological computations.

1. Introduction

Convective–stratiform classification during Quantitative Precipitation Estimation (QPE) in operational weather radar is a prominent topic within the weather radar research community [1,2,3]. This is due to the fact that accurate QPE from weather radar provides essential data that enhances weather forecasting, flood management, agriculture, water resource management, disaster response, and public safety. It also supports long-term climate monitoring and improves radar systems and operational decision-making [4,5,6]. Therefore, research on feature detection in radar reflectivity and rainfall fields to provide accurate QPE and Quantitative Precipitation Forecasting (QPF) has significantly advanced and fostered developments in machine learning and radar technology [2,7,8]. Following the development of dual-polarization radar technology, there has been significant effort to enhance the accuracy and reliability of QPE [9,10]. One of the notable efforts is the Joint Polarization Experiment (JPOLE) carried out by the NOAA Severe Storm Laboratory (NSSL) team, which focused on developing algorithms for the use of polarimetric radar variables in QPE retrieval and hydrometeor classification [11,12,13]. The polarimetric radar variables retrieved from dual-polarization radar offer significant advantages for QPE, particularly in improving the accuracy and reliability of rainfall measurements [14,15,16]. These radar parameters, including horizontal reflectivity (ZH), differential reflectivity (ZDR), differential phase (ΦDP), specific differential phase (KDP), and cross-correlation ratio ($\mathsf{\rho }$HV), provide additional information about the size, shape, and composition of precipitation particles. This leads to more precise rainfall estimates, especially in challenging weather conditions [16,17].

As explained above, critical feature detection during operational weather radar observations is crucial for reliable QPE [2]. To ensure this idea is practically applied in operational weather radar, various machine learning algorithms have been developed for feature detection [3,8,18]. A study conducted by [3] employed a machine learning algorithm, K-nearest neighbor (KNN). The results showed that KNN achieved a 95% probability of detection, an 8% false alarm rate, and an 87% cumulative success index for stratiform classifications. For convective classifications, KNN produced a 78% probability of detection, a 13% false alarm rate, and a 69% cumulative success index, suggesting that KNN has great potential in precipitation-type classification. Similarly, the study conducted by [18] on feature detection classification applied a neural network algorithm and highlighted the probability of detection, critical success ratio, and false alarm rate, which were 75%, 67%, and 18%, respectively. Furthermore, advanced machine learning models, especially tree-based ensemble methods like XGBoost and random forests, have been shown to perform better in feature detection [1]. A notable algorithm for feature detection developed by [8,19,20] demonstrated strong performance on feature detection. The algorithm separates convective–stratiform features by calculating the value that exceeds the background average by a certain amount within the reflectivity field [8]. The key difference between this algorithm and those previously mentioned is its ability to classify features in different fields, namely reflectivity, rain rate, and snow rate [2]. Its performance is assessed by considering its ability to account for uncertainty in the observed data, with a 5 ± 2 dB reflectivity calibration correction [8]. Our study focuses on identifying convective–stratiform classification in both reflectivity and QPE retrieval; therefore, we selected this algorithm to apply in our study. It is notable that the algorithm can detect features in reflectivity, rain rate, and snow rate. Previous research developed this algorithm to determine feature detection in the reflectivity field [8,19,20], while [2] applied the algorithm in developing thresholds for detecting strong and faint echo features in snow rate during the cool season. However, the application of this algorithm for detecting features in rain rate remains a gap in the literature. Our current study addresses this gap by detecting features during a heavy precipitation event in Crossville, Tennessee, during warm-season convection. Therefore, we apply the algorithm and extend our analysis to detect features in both reflectivity and rain rate (QPE retrieval), which is the main focus of this research. Although QPE needs to be retrieved after the analysis of feature detection, since our study focuses on feature detection, we first retrieve QPE and apply the algorithm to detect features in both reflectivity and QPE retrieval. Understanding feature detection is important in improving accurate QPE, which, in turn, provides data that are used directly in nowcasting, model simulations, hydrological applications, and climatological computations. The structure of this research is organized as follows: after describing the employed data and methods in Section 2, Section 3 presents the results by analyzing the inter-relationships among the polarimetric radar variables used to discriminate precipitation types and by examining meteorological and non-meteorological echoes detected in the radar resolution volume during the event. Following the correction of the base data to eliminate non-meteorological echoes, the QPE retrieval is then determined. Finally, we detect features in both reflectivity and QPE retrieval. Section 4 describes the discussion while Section 5 presents the major findings observed during the study.

2. Data and Methods

2.1. Data

To analyze feature detection in warm-season convection, we utilized Next Generation Weather Radar (NEXRAD) Level-II data provided by the National Centers for Environmental Information (NCEI) [21]. The data were downloaded from KOHX NEXRAD S-band radar during the heavy precipiation event on 11 June 2023 in Crossville, Tennessee. The data are available as open source at https://www.ncdc.noaa.gov/nexradinv/ (accessed on 15 December 2024) and have been used in various studies [2,5,22]. The data system consists of S-band Doppler weather radars with 160 high-resolution networks, including the KOHX NEXRAD used in this study. In addition to ZH, mean radial velocity (V), and spectrum width (SW), the system includes dual-polarization base data, such as ZDR, $\mathsf{\rho }$HV, and ΦDP. The data are saved in files, with each file generally containing 4, 5, 6, or 10 min of base data, depending on the Volume Coverage Pattern (VCP) scanning strategy. Each file starts with a 24-byte volume scan header record, followed by several 2432-byte base data and message records. These base data contain non-meteorological echoes and therefore need to be processed before use.

Furthermore, weather radar data are crucial for monitoring and analyzing atmospheric conditions, particularly precipitation patterns, storm structures, and atmospheric dynamics [23]. The NEXRAD level-II data source is available for the various acquisition systems, such as the Plan Position Indicator (PPI) and the Range Height Indicator (RHI), which are used to acquire weather information based on different observational needs. The PPI provides a horizontal cross-section of radar data at a fixed elevation angle. The radar continuously rotates 360° while maintaining a set tilt angle above the horizon. The RHI provides a vertical cross-section of the atmosphere by varying the elevation angle while keeping the azimuth angle fixed. This generates a profile of the atmosphere in a specific direction. Therefore, both the PPI and RHI complement each other, offering a more comprehensive understanding of weather systems. (Note: Operational observations do provide RHI data, but with a set of tilt rotational observations using the VCP11/VCP21 scan strategy. VCS can be extracted and presented to show the vertical structure of a specific storm.)

2.2. Methodology

2.2.1. Polarimetric Radar Variables Description

The KOHX weather radar, located near Crossville, Tennessee, is part of the NEXRAD network, which provides S-band Doppler radar data, including polarimetric variables. These variables enhance meteorological analysis by offering detailed information on precipitation types, sizes, and distributions. The variables are retrieved from the returned power reflected by hydrometeor particles, and the reflected power, P_r is expressed as:

$${\mathrm{P}}_{\mathrm{r}}={\displaystyle \frac{{\mathsf{\pi }}^3}{1024\times \ln 2}}\left[{\displaystyle \frac{{\mathrm{P}}_{\mathrm{t}}{\mathrm{hG}}^2\mathsf{\theta }\mathsf{\phi }}{{\mathsf{\lambda }}^2\mathrm{L}}}\right]{\displaystyle \frac{{\mathrm{K}}^2\mathrm{Z}}{{\mathrm{R}}^2}}$$

(1)

where P_t represents the radar’s peak power, h represents the pulse length, G represents the radar’s antenna gain, θ and φ represent horizontal and vertical beam-widths, respectively, λ represents the wavelength of the radar, R represents the target distance from the radar, and L is the loss factor for the radar beam attenuation, with a value around 1 [24]. Since the beam attenuation is often not known, it is often ignored. $\mathrm{K}$ represents the dielectric constant of hydrometeors and is given by the relation:

$${\lceil \mathrm{K}\rceil }^2={\left|{\displaystyle \frac{m^2-1}{m^2+2}}\right|}^2$$

(2)

where m represents the complex refractive index of hydrometeors; usually ${\left|{\displaystyle \frac{m^2-1}{m^2+2}}\right|}^2$for liquid water is taken as 0.93. The microphysical parameter, the radar reflectivity factor symbolized in Equation (1) by the letter Z, is defined and calculated by the following formula:

$$\mathrm{Z}={\displaystyle \frac{{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{N}}{\mathrm{D}}^6\mathrm{n}\left({\mathrm{D}}_{\mathrm{i}}\right)}{{\mathrm{V}}_{\mathrm{c}}}}\hspace{1em}\hspace{1em}\hspace{1em}$$

(3)

where $\mathrm{n}\left({\mathrm{D}}_{\mathrm{i}}\right)$ is the total number of rain drops in the radar resolution volume, ${\mathrm{D}}_{\mathrm{i}}$ is the rain drop diameter [24], ${\mathrm{V}}_{\mathrm{c}}\hspace{1em}$ is the contributing volume, and its logarithmic unit is dBZ. KOHX polarimetric radar transmits and receives signals in two orthogonal polarizations (horizontal and vertical) in a mode called Simultaneous Transmit and Receive (STAR mode). This dual transmission allows the radar to collect different scattering information from hydrometeors and retrieve ZH, ZDR, ΦDP, KDP, and $\mathsf{\rho }$HV (hereafter referred to as polarimetric radar variables), among others. These variables are calculated by the equations listed below:

$$\mathrm{ZDR}=10{\log }_{10}\left[{\displaystyle \frac{\mathrm{ZH}}{\mathrm{ZV}}}\right]$$

(4)

$$\mathsf{\rho }\mathrm{HV}={\displaystyle \frac{{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}\left({\mathrm{S}}_{\mathrm{i}}^{\mathrm{VV}}-\overline{{\mathrm{S}}^{\mathrm{VV}}}\right)\left({\mathrm{S}}_{\mathrm{i}}^{\mathrm{HH}}-\overline{{\mathrm{S}}^{\mathrm{HH}}}\right)}{\sqrt{{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{S}}_{\mathrm{i}}^{\mathrm{VV}}-\overline{{\mathrm{S}}^{\mathrm{VV}}}\right)}^2{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{S}}_{\mathrm{i}}^{\mathrm{HH}}-\overline{{\mathrm{S}}^{\mathrm{HH}}}\right)}^2}}}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}$$

(5)

where ZH and ZV in Equation (4) denote the radar reflectivity factors obtained by the horizontal and vertical polarization channels, respectively. S is the backscattering of waves; horizontal and vertical polarization amplitudes of backscattered waves are represented by superscripts in Equation (5). ZDR is measured in dB while $\mathsf{\rho }$HV is unitless with values ≤ 1.

2.2.2. Specific Differential Phase Retrieval

NEXRAD Level-II data do not directly include KDP as a primary base product, so it must be retrieved. To handle the noisy and non-meteorological artifacts present in ΦDP, we calculated KDP using the Vulpiani Method [25]. A filter was applied to smooth out the noise in ΦDP. After filtering ΦDP, the range derivative of ΦDP was calculated to derive the KDP using the following formula:

$$\mathrm{KDP}={\displaystyle \frac{1}{2}}{\displaystyle \frac{\mathrm{d}\left(\mathsf{\Phi }\mathrm{DP}\right)}{\mathrm{dr}}}$$

(6)

where dr is the range resolution distance between radar gates, measured in deg/km.

2.2.3. Description of Meteorological and Non-Meteorological Echoes

Initially, since many echoes in the radar base data do not originate from hydrometeors, the polarimetric radar variables were analyzed from the base data to capture both hydrometeor and non-hydrometeor echoes [26]. These echoes can interfere with or distort the radar data, making it more challenging to interpret meteorological phenomena. Common sources of non-meteorological echoes include ground clutter, sea clutter, biological clutter (birds and insects), aircraft or other moving objects (like chaff), radar interference from other radar systems (radio frequency interference), electrical and mechanical noise (radar system noise), pollution or industrial emissions, sunlight or external light reflections (sun clutter), and terrain (mountains and hills). These non-meteorological echoes must be removed (corrected) for better interpretation of the radar data.

2.2.4. Quality Control of the Base Data to Filter Non-Meteorological Echoes

Since weather radar data are frequently used in automated weather applications, it is essential to filter out non-meteorological contaminants, such as ground clutter, bioscatter, and instrument artifacts, from the data [27,28]. Therefore, efforts to improve radar systems must be carried out by a multidisciplinary team composed of technicians and meteorologists, each with experience related to weather radar in their respective fields [29]. As explained above, weather radar systems provide valuable data for QPE retrieval and hydrometeor classification. However, base radar data are frequently affected by several factors that introduce errors and biases. Therefore, correcting these errors is crucial to ensure accurate analysis and reliable QPE. For this reason, filtering was performed using the pyart.correct. GateFilter module, implemented in the Python ARM Radar Toolkit (Py-ART), a library for working with weather radar data in the Python programming language [30,31]. This process involves building boolean arrays to filter gates in the radar fields based on a set of typical conditions. In this regard, the functions exclude above, exclude below, and exclude invalid were applied to the gate filter module. The exclude above function excludes gates where a given field exceeds a specified value, the exclude below function excludes gates where a given field falls below a specified value, and the exclude invalid function excludes gates where an invalid value occurs in a field (e.g., NaNs). During filtering, we excluded values above ZH > 80 dBZ, ZDR > 5 dB, and $\mathsf{\rho }$HV > 1, as well as values below ZH ˂ 0 dBZ, ZDR ˂ −1 dB, and $\mathsf{\rho }$HV ˂ 0.75. Therefore, invalid and masked data were excluded, leaving only valid values within the following ranges: 0 ≤ ZH ≤ 80 dBZ, −1 ≤ ZDR ≤ 5 dB, and 0.75 ≤ $\mathsf{\rho }$HV≤ 1. The selection of these thresholds for discriminating non-meteorological echoes have also been suggested by other studies [32].

2.2.5. QPE Retrieval

After correcting the base data, the polarimetric radar variables were used to retrieve the QPE during the event. A key objective of polarimetric radar research is to incorporate polarization variables into precipitation estimation algorithms [13,16,33]. This effort is primarily driven by the fact that ZDR and KDP are less affected by radar calibration errors or partial beam blockage, providing a more reliable basis for rainfall estimation across a broader range of conditions compared to relying solely on radar reflectivity [11,12,13].

The QPE algorithms examined in this study utilized filtered Z, ZDR, and retrieved KDP, and are listed below:

$$\mathrm{R}\left[\mathrm{Z}\right]={\mathrm{a}}_{\mathrm{Z}}.{\left[\mathrm{Z}\right]}^{{\mathrm{b}}_{\mathrm{Z}\hspace{1em}}}$$

(7)

$$\mathrm{R}\left[\mathrm{KDP}\right]={\mathrm{a}}_{\mathrm{KDP}}.{\left[\mathrm{Z}\right]}^{{\mathrm{b}}_{\mathrm{KDP}}}.\mathrm{sign}\left[\mathrm{KDP}\right]$$

(8)

$$\mathrm{R}\left[\mathrm{Z},\mathrm{ZDR}\right]={\mathrm{a}}_{\mathrm{DR}}.{\left[\mathrm{Z}\right]}^{{\mathrm{b}}_{\mathrm{DR}}}.{\left[\mathrm{ZDR}\right]}^{{\mathrm{c}}_{\mathrm{DR}}}$$

(9)

$$\mathrm{R}\left[\mathrm{KDP},\mathrm{ZDR}\right]={\hspace{1em}\mathrm{a}}_{\mathrm{DP}}.{\left[\mathrm{KDP}\right]}^{{\mathrm{b}}_{\mathrm{DP}}}.{\left[\mathrm{ZDR}\right]}^{{\mathrm{c}}_{\mathrm{DP}}}.\mathrm{sign}\left[\mathrm{KDP}\right]$$

(10)

For the synthetic algorithm R[Z,KDP,ZDR]

$$\begin{gathered} \mathrm{if}\hspace{1em}\mathrm{R}\left[\mathrm{Z}\right]<6{\hspace{1em}\mathrm{mm}\hspace{1em}\mathrm{h}}^{-1},\hspace{1em}\mathrm{then} \\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\mathrm{R}={\displaystyle \frac{\mathrm{R}\left[\mathrm{Z}\right]}{{\left[0.4+5.0\left|\mathrm{ZDR}-1\right|\right]}^{1.3}}} \end{gathered}$$

(11)

$$\begin{gathered} \mathrm{if} 6<\mathrm{R}\left[\mathrm{Z}\right]<50{ \mathrm{mm} \mathrm{h}}^{-1},\mathrm{then} \\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\mathrm{R}={\displaystyle \frac{\mathrm{R}\left[\mathrm{KDP}\right]}{{\left[0.4+3.5 \left|\mathrm{ZDR}-1\right|\right]}^{1.7}}} \end{gathered}$$

(12)

$$\begin{gathered} \mathrm{if} \mathrm{R}\left[\mathrm{Z}\right]>50{ \mathrm{mm} \mathrm{h}}^{-1},\mathrm{then} \\ \hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\hspace{1em}\mathrm{R}=\mathrm{R}\left[\mathrm{KDP}\right] \end{gathered}$$

where a_Z, b_Z, a_KDP, b_KDP, a_DR, b_DR, c_DR, a_DP, b_DP, and c_DP are unitless empirical coefficients. Z represented in Equations (7)–(11) is in linear units, with ZDR being in dB, and KDP in (deg/km). The sign (KDP) represents a sign function that returns either +1, 0, or −1, depending on the sign of KDP. This term ensures that the rain rate reflects the correct direction of phase shift, maintaining the correct sign for the rain rate. For $\mathrm{R}\left[\mathrm{Z},\mathrm{KDP},\mathrm{ZDR}\right]$, a synthetic algorithm proposed during JPOLE was used in QPE [11,12]. The validation of QPE retrieved from the coefficients contained in these equations by previous studies showed some improvement, providing a reason for adoption of these equations in this study [12]. The coefficients are defined to make sure that the resulting estimated rainfall rate is in ${\mathrm{mm} \mathrm{h}}^{-1}$. For convenience, the coefficient values are provided in

Table 1. Empirical coefficient values used for QPE retrieval in Equations (6)–(9).

Coefficients	QPE Retrieval Formulas	Coefficient Values	Reference
az;bz	$\mathrm{R}\left[\mathrm{Z}\right]={\mathrm{a}}_{\mathrm{Z}}.{\left[\mathrm{Z}\right]}^{{\mathrm{b}}_{\mathrm{Z}}}$	$1.7\times {10}^{-2};0.714$	[12]
akdp;bKDP	$\mathrm{R}\left[\mathrm{KDP}\right]={\mathrm{a}}_{\mathrm{KDP}}.{\left[\mathrm{Z}\right]}^{{\mathrm{b}}_{\mathrm{KDP}}}.\mathrm{sign}\left[\mathrm{KDP}\right]$	$44.0;0.822$	[12]
aDR;bDR;cDR	$\mathrm{R}\left[\mathrm{Z},\mathrm{ZDR}\right]={\mathrm{a}}_{\mathrm{DR}}.{\left[\mathrm{Z}\right]}^{{\mathrm{b}}_{\mathrm{DR}}}.{\left[\mathrm{ZDR}\right]}^{{\mathrm{c}}_{\mathrm{DR}}}$	$1.42\times {10}^{-2};0.770;-1.67$	[13]
aDP;bDP;cDP	$\mathrm{R}\left[\mathrm{KDP},\mathrm{ZDR}\right]={\mathrm{a}}_{\mathrm{DP}}.{\left[\mathrm{KDP}\right]}^{{\mathrm{b}}_{\mathrm{DP}}}.{\left[\mathrm{ZDR}\right]}^{{\mathrm{c}}_{\mathrm{DP}}}.\mathrm{sign}\left[\mathrm{KDP}\right]$	$136;0.968;-2.86$	[13]

, and the simplified research flow chart is presented in Figure 1.

2.2.6. Interpolation of Radar Data into Cartesian Grid

To identify severe weather conditions during the event, the radar data, which are in antenna coordinates, were interpolated into a Cartesian grid, as shown in Figure 4 and Figure 9. Additionally, we interpolated the radar data into a Cartesian grid for improved analysis and visualization of feature detection. During interpolation, the grid shape used was (1, 201, 201), and the grid limits were set as follows: ((0, 15000), (−200000, 200000), (−200000, 200000)). This means that in all feature detection figures, the x and y axes span from −200,000 m to 200,000 m, covering a 400 km by 400 km region in both directions. The resulting grid size from this interpolation is 2 km, which is used as the grid size in feature detection.

2.2.7. Feature Detection Algorithm

After QPE retrieval, we extended our analysis to feature detection of enhanced polarimetric radar variables during the event in both horizontal reflectivity and QPE retrieval. We applied a convective–stratiform classification algorithm to detect features during a heavy precipitation event in warm-season convection [8,20]. According to [2], the algorithm can detect features in different fields, like reflectivity (primary goal of the algorithm), rain rate, and snow rate. The first step was to calculate the background radius by considering the grid dimensions. For the algorithm to run, the background radius must be two times (or three times is recommended) the grid size. Therefore, in our study, with a grid size of 2 km, a 6 km background radius size was used.

We also applied a 0.75 calculation threshold for calculating the background average value. This means that if the data are less than 75% within the footprint circle, the algorithm does not run. The algorithm classifies the point field as convective if the always core threshold value is greater than the background value by a certain value. This value is determined by either a cosine scheme or scalar. This means that if the cosine scheme is to be used, then the cosine function must be true, and the maximum difference and zero maximum cosine value come into use, as shown in Figure 2a,b.

The maximum difference indicates the maximum difference between the field values and the background value for a feature to be identified (a point where this cosine function crosses the y-axis or x-axis = 0). The zero-difference cosine value is where the difference between the field and the background is zero (a point where the cosine function crosses the x-axis or y-axis = 0). If the zero-maximum cosine value is less than the always core threshold, then all values greater than zero-maximum cosine value are considered as convective cores, or if the zero-maximum cosine value is greater than the always core threshold, then all values greater than the always core threshold will be convective cores. For reflectivity, 40 dBZ was considered the always core threshold value [8], while 25 mmh⁻¹ was used as the always core threshold for rainfall, as suggested by [2]. We also applied the algorithm to capture weak and strong features during the event by examining features that show a clear distinction compared to the background. Features with a significant distinction are classified as strong features, while those with little difference are regarded as faint features. Finally, we investigated uncertainty in both reflectivity and rainfall estimates by applying a reflectivity calibration correction of 5 ± 2 dB [8]. Therefore, we applied an offset of ±2 dB of reflectivity during feature detection. Running the algorithm with a +2 dB offset is considered an overestimate, while a −2 dB offset is considered an underestimate of the observed features during the event.

Table 2. Parameter values used in feature detection during the event.

Parameters	Reflectivity	Rainfall Rate (QPE)
dB_averaging	True	False
always_core_thres	40 dBZ	25 mmh−1
bkg_rad_km	6 km	6 km
calc_thres	0.75	0.75
use_cosine	True	True
max_diff	5 dBZ	2.5 mmh−1
zero_diff_cos_val	55 dBZ	13 mmh−1
weak_echo_thres	10 dBZ	0.5 mmh−1
estimate_flag	True	True

provides a summary of the parameters used in feature detection. More details about the algorithm can be found in [2,8].

3. Results

3.1. Interrelationships Among the Polarimetric Variables to Discriminate Different Precipitation Types in the Radar Resolution Volume During the Event

The results of PPI scans, which display the investigation of polarimetric radar variables from the base data during the event, are shown in Figure 3. These results indicate that most of the scatterer particles within the radar resolution volume were concentrated within a 300 km range (Figure 3a–f). Figure 4 provides a zoomed-in view of the features, where non-meteorological echoes are clearly marked within the black dotted circles (Figure 4a–f).

The interrelationships between the polarimetric variables are evident, highlighting valuable information about specific hydrometeors and offering complementary insights into the objects within the radar resolution volume. Figure 4 shows areas with high ZH > 50 dBZ, low ZDR (nearly 0 dB), and varying values of ΦDP and KDP, in the ranges of 100 < ΦDP < 150 degrees and 1 < KDP < 2 deg/km, respectively, indicating a storm containing large hail. This demonstrates how KDP can aid in distinguishing between regions of pure hail and areas with a mix of hail and rain at S-band frequencies. Areas with high ZH, low ZDR, and low KDP suggest a low concentration of oblate liquid drops, the presence of more spherical liquid drops, or an absence of liquid drops altogether. On the other hand, areas with high KDP and low ZDR indicate the presence of melting hailstones. Changes in ZDR and KDP within the storm’s high-reflectivity core (above 55 dBZ) align with low values of $\mathsf{\rho }$HV, in the range of 0.8 < $\mathsf{\rho }$HV < 0.9. Additionally, radar variables at an azimuth of 270° along the RHI are displayed in Figure 5. The data reveal a variety of precipitation types in this direction. Noticeable heavy rain is observed within the 75 to 125 km range, with 40 < ZH < 50 dBZ, while moderate rain is marked within the 125 to 150 km range (30 < ZH < 40 dBZ) below 4 km above the ground. Drizzle, very light rain, and light rain are observed above 4 km from the ground within the 110–175 km range.

The results show a signature of low ΦDP, moderate to high ZDR, and moderate ZH within the 75–100 km range. Near the surface, precipitation transitions to rain mixed with hail, as indicated by high ZH, high ZDR, and lower $\mathsf{\rho }$HV within the 55–60 km range. However, the combination of high ZH, low ZDR, and elevated ΦDP exhibits a characteristic hail signature. Additionally, the results of the Range–Azimuth and Gates–Rays analyses are shown in Figure 6. The radar detects high ZH > 50 dBZ around 50 km from the scan across all directions (0–360°), as shown in Figure 6a–d. This indicates a strong precipitation event with thunderstorms in all directions of the radar site along the range, concentrated between 250 and 1000 gates from the radar (Figure 6e–h). However, other regions are characterized by low ZH values, nearly 10 dBZ, between 100 and 200 km, indicating either light precipitation or a dry area near the radar site.

The presence of high ZDR values from 0–100 km typically indicates large, non-spherical particles such as hail, while lower ZDR values beyond 100 km suggest raindrops. The presence of thunderstorms around the 50 km range is further evidenced by high ΦDP values in the range of 60 < ΦDP < 160 degrees, demonstrating how the radar wave’s phase changes as it travels through different precipitation particles. High $\mathsf{\rho }$HV values (0.9 < $\mathsf{\rho }$HV < 1.0) typically indicate uniform, liquid precipitation such as rain.

Furthermore, the presence of large hailstones within a supercell thunderstorm is evident from 75–100 km in Figure 4a and Figure 5a, where ZH > 60 dBZ, as well as in Figure 4b and Figure 5b, where ZDR is near 0 dB. This highlights the presence of strong updrafts, mixed-phase precipitation, and large hailstones within the radar resolution volume during the event.

3.2. The Role of Polarimetric Radar Variables in Distinguishing Observed Meteorological and Non-Meteorological Echoes in the Radar Resolution Volume During the Event

One of the most interesting details indicated by the polarimetric radar variables is their ability to distinguish between meteorological and non-meteorological echoes within the radar resolution volume. Figure 4a–e clearly indicate non-meteorological echoes detected during the event (areas marked with black dotted circles). This observation aligns with Figure 3 and Figure 7a–e within the 0–75 km range. According to the PPI scans in Figure 3, these non-meteorological echoes are concentrated within a 100 km range from the radar.

Figure 7 shows different characteristics of the returned radar signals, exhibiting varying values of polarimetric radar variables. Below the 75 km range, significantly lower ZDR values reaching −15 dB and lower $\mathsf{\rho }$HV < 0.4 are evident. Such low, negative ZDR values and reduced $\mathsf{\rho }$HV suggest that these echoes were non-meteorological. Although tornadoes are typically associated with supercell thunderstorms, Figure 3 reveals that even at high resolution, a hook echo—characteristic of a tornadic signature—was not detected.

On the other hand, beyond the 75–150 km range, hydrometeor particles are evident (Figure 7). This region is characterized by $\mathsf{\rho }$HV ≈ 1.0, indicating that rain dominated the area. In contrast, within the 150–200 km range, $\mathsf{\rho }$HV values between 0.75 and 0.8 suggest the presence of hail. Additionally, KDP values fluctuate near zero between the 75–150 km range, indicating difficulty in detecting KDP in light precipitation. However, within the 150–200 km range, KDP values range from 0.5 to 1.0 deg/km, demonstrating that KDP is more easily detected in heavy precipitation. This observation is consistent with the high ZH values exceeding 40 dBZ in the 150–200 km range, which correspond with clearly detectable KDP values.

3.3. The Results of QPE Retrieval During the Event

After identifying both meteorological and non-meteorological echoes in the radar resolution volume, a quality control process was implemented, as shown in Figure 8. All non-meteorological echoes were filtered out, leaving the data clean and ready for QPE retrieval. QPE was successfully retrieved, and the results are presented in Figure 9. The results show consistency in rainfall intensity among R(ZH), R(ZH,ZDR), and R(ZH,KDP,ZDR). However, observations indicate that R(KDP) and R(KDP,ZDR) exhibited higher water content compared to the other methods during the event, highlighting the reliability and value of KDP for QPE retrieval, particularly in high-intensity rainfall where R(ZH) fails to provide accurate estimates. This confirms that KDP offers a more direct measure of precipitation liquid water content (LWC) and is relatively unaffected by particle size or shape.

However, in low-to-moderate intensity rainfall, KDP can be noisy and may require smoothing or additional processing for accuracy. KDP is more sensitive to strongly convective conditions, making it less reliable in stratiform or weakly convective environments. The QPE retrieval results indicate that low-to-moderate intensity rainfall dominated the radar field during the event, as evidenced in Figure 9. Therefore, due to this and the consistency observed in QPE retrieval, we selected QPE estimated from R(ZH) and R(ZH,KDP,ZDR) to demonstrate the application of the convective–stratiform classification algorithm for feature detection analysis during the event.

3.4. The Results of Feature Detection During the Event

After QPE retrieval, we extended our analysis to include feature detection in both ZH and QPE estimated from R(ZH) and R(ZH,KDP,ZDR). An investigation of feature detection in QPE retrieval was conducted, and the results of feature detection in R(ZH) and R(ZH,KDP,ZDR) produced reliable outcomes similar to the existing technique of detecting features in reflectivity (ZH). The results of feature detection are presented in Figure 10, Figure 11, Figure 12 and Figure 13. The convective–stratiform features detected by the algorithm are clearly shown within the ZH, R(ZH), and R(ZH,KDP,ZDR) fields, as shown in Figure 10. Distinct strong convective cores (ZH > 40 dBZ) are evident in the ZH field, extending from the southwest to the northeast, as well as in some parts of the east and west of Figure 10b. The surrounding regions of these convective cores are characterized by stratiform cores with ZH < 40 dBZ. The remaining areas of the ZH field were not categorized due to backscattered weak signal, which resulted in unavailable data.

Feature detection in the R(ZH) and R(ZH,KDP,ZDR) fields highlights convective cores (R(ZH) > 13 mmh⁻¹) extending from the southwest to the northeast, as well as in some parts of the east and west, as shown in Figure 10d,f. Similarly, the neighboring regions of the convective cores in R(ZH) and R(ZH,KDP,ZDR) correspond to stratiform cores (R(ZH) < 13 mm h⁻¹). The results indicate that convective cores occupy smaller areas with intense rainfall exceeding 50 mmh⁻¹, as shown in Figure 10c,e. In contrast, stratiform cores have a larger horizontal coverage and lower rainfall intensity (<13 mmh⁻¹), as shown in Figure 10c,d.

In addition to analyzing convective and stratiform cores, the study extended its analysis to strong and faint features to provide insight into the intensity and nature of the features in ZH, R(ZH), and R(ZH,KDP,ZDR), as shown in Figure 11. The results indicate that strong features are associated with convective cores, while faint features are observed within stratiform cores. This is evident in the northeast, west, southwest, and northwest regions of the ZH, R(ZH), and R(ZH,KDP,ZDR) fields. The analysis of strong–faint features reveals that most of the convective cores detected in the ZH field are strong features (Figure 11b), while faint features are very few within the stratiform core, indicating little difference from the background values. Similarly, in R(ZH) and R(ZH,KDP,ZDR), the convective cores detected in the field are also strong features, while faint features are scarce within the stratiform core, as displayed in Figure 11d,f.

Furthermore, the results of overestimation and underestimation, which account for uncertainty in both ZH and R(ZH) fields, are presented in Figure 12 and Figure 13. An offset of +2 dB causes the radar to overestimate convective cores in the ZH field in the northeast, as shown in Figure 12d, displaying more convective cores than those depicted in the best feature detection of Figure 12b. Conversely, an offset of −2 dB causes the radar to underestimate convective cores in the ZH field in the northwest (Figure 12c), showing weaker convective cores than those in the best feature detection. Similarly, an offset of +2 dB causes the radar to overestimate convective cores in the R(ZH) field in the northeast, as shown in Figure 13d, displaying more convective cores than those shown in the best feature detection (Figure 13b). On the other hand, an offset of −2 dB leads to an underestimation of convective cores in the R(ZH) field in the northwest (Figure 13c), showing weaker convective cores than those depicted in the best feature detection (Figure 13b).

The observation shows that the regions exhibiting convective and stratiform cores in the ZH field are the same regions where these features are depicted in the R(ZH) and R(ZH,KDP,ZDR) fields. This similarity is also evident in the strong–faint features of the ZH, R(ZH) and R(ZH,KDP,ZDR) fields. However, there is a significant difference in the signal intensity of convective cores between the ZH and both R(ZH) and R(ZH,KDP,ZDR) fields in both convective–stratiform and strong–faint feature detection. The intensity of convective cores is higher in ZH than in both R(ZH) and R(ZH,KDP,ZDR). Convective cores are typically more intense, producing localized heavy rainfall and strong updrafts, whereas stratiform cores are more widespread, characterized by lighter, steady rainfall and weaker updrafts.

4. Discussion

The study initially analyzed the interrelationship among polarimetric radar variables within the radar resolution volume during the event. It demonstrates how polarimetric radar can more effectively differentiate among meteorological phenomena. This is evidenced in Figure 7, which shows that $\mathsf{\rho }$HV is a powerful tool for distinguishing between meteorological and non-meteorological echoes. Furthermore, tornadic debris, a major signature typically associated with supercell thunderstorms during warm-season convection, was not clearly depicted in the images. This is consistent with the Climavision report at https://climavision.com/blog/june-2023-tornadic-events-highlight-need-for-more-low-level-radar-coverage/ (accessed on 10 January 2025), which highlighted that the tornadic signature during the event was not captured by the radar display system, resulting in no warning being issued. According to the report, the reason for this was that the storm structure was below the radar beam. This circumstance might also be attributed to a lack of clear rotation within the storm system, as shown in Figure 4f. Some tornadoes may not exhibit distinct rotational signatures due to weak or diffuse circulations that are difficult to detect. If a tornado is not embedded within a well-defined mesocyclone or if the storm’s rotation is weak, the radar might fail to capture its signature. It should be noted that not all tornadoes are associated with large, well-organized supercells that have strong mesocyclones. Tornadoes can also form within more disorganized storms, making them harder to detect on radar.

Furthermore, the results of feature detection highlight the existence of strong convective cores (ZH > 40Z dBZ) extending from the southwest to the northeast, as well as in some parts of the east and west within the ZH field (Figure 10b). These regions are surrounded by stratiform cores (ZH < 40 dBZ). Observations indicate that the convective and stratiform cores in the ZH field closely resemble the features depicted in the R(ZH) and R(ZH,KDP,ZDR) fields. This similarity is also evident in the strong–faint features of the ZH and both R(ZH) and R(ZH,KDP,ZDR) fields in Figure 11b,d,f.

However, differences in the intensity of convective cores between the ZH and both R(ZH) and R(ZH,KDP,ZDR) fields are observed in both convective–stratiform and strong–faint feature detections. The intensity of convective cores is higher in the ZH field than in the R(ZH) and R(ZH,KDP,ZDR) fields. This discrepancy might be due to uncertainties in the Z-R and R(ZH,KDP,ZDR) conversion relationships, leading to variations in the intensity of the features detected across the fields. The results also indicate that stratiform precipitation dominates over convective precipitation in the R(ZH) and R(ZH,KDP,ZDR) fields.

A previous study conducted by [20] indicated that within stratiform precipitation, the same rain rate can result from either a drop spectrum dominated by many small droplets (lower ZH) or a few large droplets (higher ZH), meaning that ZH values for a given rain rate can vary by as much as 9 dB. Additionally, when comparing the convective–stratiform cores of ZH, R(ZH) and R(ZH,KDP,ZDR) fields, it is evident that convective cores cover a smaller area with high rainfall rates exceeding 50 mmh⁻¹, while stratiform cores cover a larger area with greater horizontal homogeneity and lower rainfall intensity. This key characteristic of convective–stratiform precipitation is consistent with previous research [1,3,7,18], which observed widespread stratiform cores compared to the relatively smaller convective cores.

5. Conclusions

This study investigated feature detection in a heavy precipitation event using polarimetric radar variables and QPE retrieval during warm-season convection. The analysis of polarimetric radar variables revealed that strong updrafts, mixed-phase precipitation, and large hailstones within the radar resolution volume during the event were associated with the presence of supercell thunderstorms.

The feature detection results indicated that regions exhibiting convective–stratiform cores and strong–faint features in the reflectivity field were similar to those in the rainfall field. However, there was a difference in the intensity of convective cores between the reflectivity and rainfall fields, with stronger convective cores observed more frequently in the reflectivity field than in the rainfall field. This discrepancy can be attributed to uncertainty in the reflectivity and polarimetry–Rain rate conversion, which results in varying magnitudes of reflectivity dB values.

The features detected in the reflectivity field better represent the reality of convective–stratiform cores in the weather radar field, demonstrating the improved accuracy of feature detection using reflectivity data. Estimates to account for uncertainty during feature detection indicated that a +2 dB offset overestimated convective features in the northeast in both the reflectivity and rainfall fields, while a −2 dB offset underestimated convective features in the northwest. The results highlight that convective cores cover smaller areas with high rainfall rates exceeding 50 mmh⁻¹, while stratiform cores cover larger, more homogeneous areas with lower rainfall intensity.

Although different machine learning algorithms have been employed in feature detection, challenges in convective–stratiform feature detection still exist. These challenges are primarily due to the complexities of precipitation structures, limitations in radar technology, rapid storm evolution, and the need for sophisticated algorithms to accurately classify and track these features. These factors make it difficult to achieve precise and timely identification of convective and stratiform areas, which is essential for accurate forecasting and severe weather prediction. However, our results highlight how the algorithm can detect features in QPE retrieval and are beneficial in several aspects. Convective–stratiform detection is crucial for nowcasting, especially for severe weather warnings, as convective systems can develop rapidly and unexpectedly. It also enhances weather forecasting by identifying whether a precipitation event is convective or stratiform, allowing meteorologists to more accurately predict storm behavior, intensity, and duration. Lastly, in hydrological modeling, differentiating between convective and stratiform rainfall is important for flood forecasting, as convective rain tends to lead to more sudden and localized flooding, while stratiform rainfall may result in slower, more widespread runoff. The next part of the research will focus on the validation of QPE retrieval and QPF.