This study included 15 isolated thunderstorm cells that produced lightning, which was observed via an S-band dual-polarization radar deployed in Guangzhou city (GZ radar) and a low-frequency E-field detection array (LFEDA) (Table 1). These isolated thunderstorms were selected in accordance with the rules in Zhao et al. (2021a, 2022). The initiation of such thunderstorms was defined as the first radar volume scan in cases where the composite reflectivity was ≥ 5 dBZ and the locations of thunderstorms were within the effective range of the radar and lightning array. When the maximum reflectivity of this cell starts to fade with a value of less than 30 dBZ later, the evolutionary process of a cell is marked as the end. Moreover, we required the subsequent lightning activity to be present after the first lightning occurrence in this study. The average 6-hourly convective available potential energy (CAPE) of these thunderstorms was obtained from ERA-Interim reanalysis data, as in Zhao et al. (2022). The detailed examination of lightning activity with related dynamics and microphysics in case #1 was conducted first, and then the statistical results of all cases were given. On 20 June 2016, one isolated thunderstorm cell was observed by the GZ radar (Fig. 1). The composite reflectivity data revealed that this thunderstorm (the boundary was determined via manual inspection; the black lines in Fig. 1) was nearly stationary and that the cloud life cycle lasted approximately 2 h. Lightning activity occurred when values of composite reflectivity > 35 dBZ were present; this phenomenon seemingly supported the results of Hayashi et al. (2021), highlighting the importance of graupel in cloud electrification.

The beam width of the GZ radar is ≤ 1°, and the azimuth and range resolutions are 1° and 250 m, respectively. A full radar volume scan lasted 6 min at nine elevation angles. The GZ radar data were processed via the Python ARM Radar Toolkit (Py-ART), including quality control (Helmus and Collis, 2016; Li et al., 2023; Li et al., 2024). The ZDR offset of the raw data was corrected via drizzle, and the calibration accuracy was expected to be 0.1–0.2 dB (Bringi and Chandrasekar, 2001). We utilized two methods to smooth the differential phase ΦDP, namely, “lightly filtering” (2 km) and “heavily filtering” (6 km), as in Park et al. (2009). Two estimates of KDP were subsequently obtained from a slope of a least-squares fit of the filtered ΦDP; a lightly filtered KDP was subsequently used in the case of horizontal reflectivity > 40 dBZ, and a heavily filtered KDP was selected otherwise (Ryzhkov and Zrnić, 1996). The radar data were gridded via Py-ART gridding routines on a Cartesian grid with a 0.25 km horizontal resolution and a 500 m vertical resolution (“Barnes2” method).

The lightning flashes within this thundercloud were detected via LFEDA, which is a 3D mapping detection system for intracloud and CG lightning, with 10 sensors. Previous studies (Fan et al., 2018; Shi et al., 2017) utilized information from triggered lightning flashes to evaluate LFEDA detection results; the results show the detection efficiency of the LFEDA can reach 100 %, and the mean location error is 102 m. Discharge pulse events were grouped into a lightning flash via the same method as described by Liu et al. (2020). A potential discharge pulse event of one lightning flash should occur within 0.4 s of the previous discharge pulse event and within 0.6 s and 4 km of any other discharge pulse event of this lightning flash. If any source within one lightning flash is below the 2 km height, this lightning flash is regarded as one CG flash, as suggested by Zhao et al. (2021a).

The lightning flashes were assigned to the isolated thunderstorm on the basis of the boundary of the thunderstorm, as well as a constraint every 6 min (according to the duration of a full radar volume scan). The life cycle of this isolated thunderstorm initiated from the first radar echo (i.e. the presence of a maximum horizontal reflectivity (ZH)≥5 dBZ in a full radar volume scan within this cloud was first detected by the GZ radar, as suggested by Zhao et al. (2021a, 2022, 2024)) and ended when the maximum ZH started to decrease and the ZH reached ≤ 40 dBZ. The distributions of these detection systems, including radar and lightning location system, were illustrated by in Zhao et al. (2024).

To estimate the precipitation-sized ice mass (e.g. graupel, hail, and frozen drops) and the content of supercooled raindrops within the mixed-phase zone, an approach on the basis of difference reflectivity (ZDP, dB) is applied (Carey and Rutledge, 2000; Straka et al., 2000): 1Zh=4λ4/π4Kw2〈|Shh|2〉,2Zv=4λ4/π4Kw2〈|Svv|2〉,3ZDP=10log⁡10Zh-Zv,forZh>Zv, where Sij refers to an element of the backscattering matrix of a hydrometeor (Zrnić, 1991). The first subscript i indicates the polarization of the backscattered field (h is horizontal, v is vertical), and the second subscript j indicates the polarization of the incident field; Kw=(ϵw-1)/(ϵw+2) is the factor associated with the dielectric constant of water, ϵw is the dielectric constant, and λ is the radar wavelength. The brackets indicate expectations expressed in terms of the distribution of mean hydrometeor properties such as shape, size, fall orientation, particle density, canting angle, dielectric constant, and others.

Pruppacher and Klett (1997) assumed that precipitation-sized ice particles were more spherically symmetrical or tumble. The low dielectric constant and significant canting behaviour of ice particles probably result in a near-zero ZDR (e.g. Seliga and Bringi, 1976). Therefore, the horizontal reflectivity and vertical reflectivity are equal for ice particles, as “effective spheres”, and ZDP is solely influenced by raindrops. If the relationship between horizontal reflectivity and ZDP (raindrops) is known, the horizontal reflectivity of raindrops can be derived. The relationship between the horizontal reflectivity of raindrops and ZDP (raindrops) is derived from 2-year disdrometer data in Guangdong Province (Li et al., 2019), which is suitable for the current radar used in this study: 4ZHrain=0.0044ZDP2+0.58054ZDP+16.591. Then, the ice echo intensity ZHice can be expressed as ZH-ZHrain. The standard error for the relationship between horizontal reflectivity and ZDP is consistently approximately 1 dB (Carey and Rutledge, 2000). If ZH-ZHrain<1 dB, which is below the melting layer, Zhice=0 mm⁶ m⁻³, ZHrain=10log⁡10Zhrain, and Zhrain=Zh. In contrast, above the melting layer, if ZH-ZHice<1 dB, then Zhrain=0 mm⁶ m⁻³, and Zhice=Zh (Carey and Rutledge, 2000). The estimates of rain mass (Mw, g m⁻³) and ice mass (Mice, g m⁻³) are derived via the following reflectivity-mass relationships (Chang et al., 2015; Zhao et al., 2022): 5Mw=3.44×10-3Zhrain4/7,6Mi=1000πρiN03/75.28×10-18Zhice7204/7, where ρi indicates the ice density (kg m⁻³) and N0=4×106 m⁻⁴. The estimated ice mass from the horizontal reflectivity of ice particles is proportional to the actual ice mass and depends on the variability in the intercept parameter of an assumed inverse exponential distribution for ice and the ice density; thus, the trends of the estimated ice mass are deemed sufficient to investigate lightning activity. Importantly, the value of ZDP can differentiate between ice and rain only if the ZH is sufficiently large (i.e. diameter ≥ 1 mm); under such conditions, the raindrop is characterized by significant oblateness (Carey and Rutledge, 2000; Green, 1975).

The estimated ice masses are assigned to graupel masses on the basis of scattering properties, that is, where the ZH values exceed 35 dBZ (Carey and Rutledge, 2000; Kumjian, 2013a, b; Zhao et al., 2021b). The threshold value is usually applied to identify graupel in the hydrometeor identification method (Park et al., 2009). The rain water content above the melting level (0 °C isotherm height; sounding data from the Qingyuan meteorological observatory are used to obtain the environmental temperature) is defined as the supercooled raindrop mass.

ZDR columns are associated with low concentration of large raindrops (> 2 mm); thus, the median volume diameter D0 of raindrops is retrieved via the method described by Hu and Ryzhkov (2022) to provide supporting evidence for identifying ZDR columns. Notably, D0 is not used to identify ZDR columns directly but rather to ensure the threshold value of ZDR, which is utilized to identify ZDR columns directly. The fractional standard deviation of the D0 estimation is approximately 10 % (Hu and Ryzhkov, 2022).

To further explore the characteristics of the microphysics related to the ZDR/KDP column and lightning within these thunderstorms, the hydrometeor identification method involving the fuzzy-logic algorithm (as in Zhao et al., 2021b), and the microphysical fingerprint (following Kumjian et al., 2022) are conducted. Identifying polarimetric radar “fingerprints” of ongoing microphysical processes was introduced by Kumjian (2012); these fingerprints are defined as vertical changes in two (e.g. ZH, ZDR) or more of the dual-polarization radar variables (Kumjian et al., 2022).

As suggested by Kumjian et al. (2022), the co-polar correlation coefficient (CC) is neglected in most of the fingerprints discussed but it is important to indicate the melting process; thus, we added the changes in the CC towards the ground to identify the melting process. The changes in the polarimetric radar variables towards the ground for riming and aggregation are the same (Kumjian et al., 2022); however, the riming process is valuable for studying lightning activity. Thus, we followed the method for identifying the aggregation and graupel particles in Park et al. (2009), that is, we utilized the discriminated convective and stratiform echoes to determine where riming or aggregation processes occur. If the echo is classified as convective, then the aggregation process is not allowed within a whole vertical column; conversely, the riming process is excluded in the stratiform case.

In addition, the 0 °C isotherm height is used to discriminate warm-rain processes and mixed-phase processes; notably, this rule introduces potential errors when it is used where the ZDR/KDP column is used. Specifically, the polarimetric characteristics of collision–coalescence and size sorting (or evaporation) processes above the 0 °C isotherm height are regarded as vapour deposition and refreezing processes, respectively. In this study, we utilize the identified and quantified ZDR/KDP columns to correct this possible error. A summary of the changes in the polarimetric radar variables towards the ground and additional conditions for different microphysical processes is listed in Table 2. To characterize the microphysical fingerprints in this study, the changes in the polarimetric radar variables towards the ground are computed between two adjacent grids in the vertical direction (for example, ΔZH (x,y,z1)=ZH(x,y,z2)-ZH(x,y,z1), x, y, and z are the three dimensions in the Cartesian coordinate system; z1 is 500 m in height, and z2 is 1000 m in height). The minimum thresholds of ΔZH> 0.002 dB km⁻¹ and ΔZDR> 0.0001 dB km⁻¹ are applied to avoid false classifications based on noise present in the data, as in Kumjian et al. (2022).

Currently, a few methods are available to automatically identify and quantify ZDR/KDP columns (e.g. Woodard et al., 2012; Krause and Klaus, 2024; Sharma et al., 2024; Snyder et al., 2015; van Lier-Walqui et al., 2016). These methods are constructed on the basis of the morphology of the ZDR/KDP columns and/or the high values (e.g. ZDR > 1 dB; KDP> 1° km⁻¹) above the melting level.

In the stage of cloud formation, high ZDR values above the melting level are simply used to identify ZDR columns; however, these high values are not always associated with the ZDR columns. Notably, three-body scatter signatures (Zrnić, 1987), depolarization streaks associated with the canting behaviour of ice in regions of strong electrification (Kumjian, 2013c), and oblate ice particles can lead to enhanced positive ZDR values (Kumjian, 2013a). Thus, additional requirements were imposed to avoid identification errors, e.g. the reflectivity threshold value (ZH≥40 dBZ) (Woodard et al., 2012) and the height should be below the homogeneous melting level, and the maximum height of the ZDR column should be associated with the height at which ZDR displays a negative vertical gradient (van Lier-Walqui et al., 2016).

Recently, Krause and Klaus (2024) utilized the hotspot technique to identify the base of the ZDR column on the basis of constant altitude plan projection indicators. Although their results indicated an improvement in the plan region identified on the basis of the ZDR column approach over the results of two different existing algorithms (Thunderstorm Risk Estimation from Nowcasting Development via Size Sorting algorithm and the algorithm introduced by Snyder et al. (2015)), the depth and volume information for ZDR columns was lost, which was not beneficial for forecasting lightning activity. In addition, the reflectivity threshold value (ZH > 25 dBZ) was required in this algorithm. Sharma et al. (2024) used an algorithm in the “scikit-image” Python package to identify ZDR/KDP columns from Cartesian grid data via threshold values (ZDR≥1 dB; KDP≥1° km⁻¹); they restricted the ZH values to exceed 20 dBZ, and any instances of obvious data contamination or unrealistic values during the gridding process were manually removed prior to analysis. They focused on three supercells during lightning activity, and violent convection in such case will ignore the formation of ZDR/KDP columns at weak echo intensities, i.e. in the initiation stage of convective clouds.

Our objective in this study is to explore the performance of microphysics retrieved via radar for forecasting lightning activity within isolated thunderstorm cells during the cloud life cycle. As depicted in Fig. 1, our radar observations indicate that before the initiation of lightning, the echo intensity is weak during the early stage of a cloud. Moreover, we seek to determine whether ZDR columns form in the early stages of cloud formation. In addition, KDP columns are representative of small drops with high concentration shed from large ice particles (e.g. hailstones) growing in a wet regime (Hubbert et al., 1998; Loney et al., 2002).

Figure 2 shows the first appearance of the ZDR column at 17:24 (China Standard Time, CST), which is approximately 36 min earlier than the first lightning occurrence and only lags behind the first radar echo by 6 min. The high values of ZDR extend to the mixed-phase region from the cloud bottom, and the height of the ZDR column is approximately 1 km (Fig. 2c). The corresponding ZH values are smaller than 30 dBZ, but with the values of the CC are relatively high (Fig. 2a, d). These characteristics indicate the presence of large raindrops with low concentrations; the low KDP values and large size of raindrops (exceeding 2 mm) support that (Fig. 2e, f). We illustrate the microphysical structure corresponding to high ZDR values in Fig. 2b. The threshold value for identifying the ZDR column in this study is 1.5 dB, considering that the size of raindrops should exceed 2 mm within the ZDR column during the initial phase of a storm (e.g. Kumjian et al., 2014). In addition, the KDP column is absent in the initial phase of this storm, not appearing until 18:30 CST.

In this study, we improved a method that only depends on the ZDR parameter for identifying and easily quantifying the ZDR column (e.g. height and volume) during the whole life cycle of a storm, specifically, including the initial phase of a convective cloud; the morphology of the ZDR column resulting from size sorting and the high ZDR values (≥ 1.5 dB) in Cartesian grid data are combined as the basis of this method. A flow chart of this method is depicted in Fig. 3.

First, we establish logic matrices with the same specifications as the ZDR matrices and identify the layer that corresponds to the 0 °C isotherm height (with sounding data used to determine the environmental temperature). Second, from the 1 km height below the melting level to the mixed-phase region, if ZDR≥1.5 dB, the corresponding logic values in the logic matrix are equal to 1. Notably, we utilize the lower portion of the logic matrix to upwardly restrict the possible errors associated with including locations outside the successive column of high ZDR values (≥ 1.5 dB). This restriction condition is based on the column morphology (columnar shape), which results from drop size sorting. Specifically, the region of the ZDR column contracts upwards from the 0 °C isotherm height below the mixed-phase region (e.g. as shown in Amiot et al., 2019, Hubbert et al., 2018, Kumjian et al., 2014, Snyder et al., 2015 and Tuttle et al., 1989). In the Cartesian grid data, this phenomenon is particularly obvious (Fig. 2c).

Third, the size distribution of raindrops within the ZDR column is determined by size sorting; thus, we utilize the negative vertical gradient of ZDR on the basis of every grid between two adjacent layers from the 0 °C isotherm height to the uppermost limit height to further ensure that the grids are associated with the ZDR column. Finally, a 3D mapping of the ZDR column is constructed. We can use the grid information to compute the height and volume of the ZDR column. Although the formation mechanism of the KDP column is different from that of the ZDR column, the morphology of the KDP column is highly consistent with that of the ZDR column; thus, this method can be applied to map the KDP column with a 3D grid based on the threshold value for identifying the KDP column (e.g. ≥ 1° km⁻¹).

Snapshots of the identified ZDR columns via the “3D mapping columns” method at four moments (the KDP column only existed at one moment during the life cycle of this thunderstorm) are shown in Fig. 4. The identified regions of the ZDR/KDP columns are represented by white dots. We verify that the 3D mapping columns method performs well in identifying the ZDR/KDP columns via manual inspection.

The first ZDR column in this isolated thunderstorm occurred at 17:24 CST (Figs. 4a, 5a). The second ZDR column subsequently occurred at 17:42 CST after 18 min (Figs. 4b, 5a). At 18:06 CST, the lightning activity reached the first peak (Fig. 6; to compare the radar and lightning data, the lightning flash frequency was counted every 6 min), and the ZDR column was the largest (in volume) at this time (Fig. 5a). However, the overlap region between the ZDR column and reflectivity core at 17:24 and 17:42 CST disappeared at 18:06 CST (Fig. 4c); i.e. the ZDR column within the reflectivity core began to collapse because of the falling of large-sized (represented by ZH values exceeding 40 dBZ) ice particles.

Although large ice particles form and fall, the KDP column is absent. However, a KDP core with high values (≥ 1° km⁻¹) occurs near the location where large ice particles (approximately 50 dBZ) melt, and a shedding process may occur, resulting in a KDP core (Fig. 4e). At 18:30 CST, the lightning activity reaches the second peak (Fig. 6), but the ZDR column almost disappears; interestingly, the KDP column forms, with high values expanding downwards to the bottom of the cloud (Fig. 4d, f). In addition, the KDP column only occurs at 18:30 CST, which corresponds to the second peak of lightning activity (Fig. 5b).

To study the vertical thunderstorm structure related to lightning activity, we explore the vertical structures of polarimetric radar variables and microphysics, in combination with 3D lightning location data. Figure 7 displays the cross sections of polarimetric radar variables (ZH, ZDR, and KDP) and microphysics (hydrometeor types and microphysical fingerprints) from the Cartesian grid of the studied isolated thunderstorm. Figure 7a1–e1 shows the polarimetric structure prior to initiation of lightning. The ZDR column and reflectivity core (≥ 25 dBZ) begin to separate, having previously been overlapping during the initial development stage of the thunderstorm (Fig. 4a, b). Riming and graupel are present; specifically, the locations of graupel particles are associated with low ZDR values. At 18:00 CST (Fig. 7a2–e2), the lightning activity begins, and the locations of the flash sources are high and correspond mainly to graupel particles. Riming occurrence surrounds the flash sources. The ZDR column and reflectivity core (≥ 40 dBZ) are almost separated. Then, at 18:06 CST (Fig.  7a3–e3), riming has increased, the echoes strengthen (≥ 55 dBZ), and the heights of the strong echoes are lifted. The lightning activity reached the first peak, where the locations of the flash sources mainly corresponded to graupel and ice particles. This finding indicates that the convective strength is obviously increased and that the cold cloud processes are heavy. Moreover, the ZDR column is located at the periphery of the reflectivity core, and high KDP values occur and correspond to heavy rain particles, which are associated with large ice particles (e.g. hailstones) melting, raindrop coalescence and/or break. This phenomenon is consistent with that the KDP tends to be directly proportional to the rain mixing ratio (Snyder et al., 2017).

Subsequently, the lightning activity weakened at 18:12 and 18:18 CST. During this stage (Fig. 7a4–e4, a5–e5), the reflectivity core is landing and large ice particles above the melting level decrease, corresponding to heavy melting and indicating increasing downdraughts. Although ZDR columns are present, they can only indicate updraughts around the reflectivity core. However, the reflectivity core was lifted again at 18:24 CST (Fig. 7a6). The contents of rain and hail mixtures and graupel clearly increased (Fig. 7d6). This indicates that the convective strength is increased. Notably, the ZDR column and reflectivity core overlap again, just as occurred during the initial development of the thunderstorm (Figs. 4a, b; 7b6). Although a few high KDP values occurred above the melting level, a KDP column formed during the next 6 min (Fig. 7c6, c7). At 18:30 CST (Fig. 7a7–e7), the lightning activity reaches the second peak, and the riming process surrounds these flash sources. The ZDR column within the reflectivity core quickly collapses with the occurrence of abundant graupel particles.

In total, this thunderstorm shows two impulses in convective strength, which correspond to two lightning activity peaks. When the first impulse event initially develops, the ZDR column is obvious and overlaps or partly overlaps with the reflectivity core (Figs. 4a, b, and 7b1); however, the region of the ZDR column within the reflectivity core collapses, with abundant graupel particles forming by riming or freezing, stimulating updraughts and intensified lightning. When large ice particles (e.g. graupel or hailstone) subsequently decrease, indicating the end of the first impulse event, melting and shedding processes occur, resulting in more raindrops (many moderate-to-large and small raindrops) contributing to high ZDR/KDP values. These raindrops could recirculate into the updraughts and be lifted to the mixed-phase region, forming the ZDR column first, and raindrops could transfer to abundant graupel and even hailstones, promoting convection, i.e. increasing lightning activity (indicating the second impulse event). However, the ZDR column within the reflectivity core will collapse with increasing amounts of graupel and/or hailstone particles, but the KDP column will occur; this can be explained by the increased KDP values at the column top being associated with an increasing number of small-to-moderate hailstones with significant water fraction (Snyder et al., 2017). The lightning activity also reaches a peak value.

Thus, the ZDR column within the reflectivity core is probably an indicator of imminent ice microphysics, and then, the formation of abundant graupel particles promotes lightning activity via non-inductive charging; the KDP column is highly related to cold cloud processes, replacing the ZDR column to indicate updraughts within the reflectivity core when obvious graupels and hailstones occur.

Figure 8 shows the normalized distributions of the polarimetric and microphysical characteristics within the series of ZDR columns during the life cycle of the studied thunderstorm. The ZH values within the ZDR columns range from approximately 10 to 55 dBZ; specifically, weak reflectivity is present in the initial phase of the thundercloud (Fig. 8a). This suggests the use of the ZH threshold value (e.g. 25, 40, or 35–50 dBZ) to help select ZDR columns results in the loss of information, especially for the initial phase of clouds. The increase in reflectivity intensity within the ZDR columns can be used to predict lightning activity, and the peaks in both reflectivity intensity and lightning activity are consistent. A strong reflectivity intensity peak is associated with a high number of lightning flashes (Fig. 8a).

The ZDR values within the ZDR columns range from approximately 1.5 to 4 dB, and the increasing trend of the ZDR values is consistent with the first peak of lightning activity but has a low correlation with the second peak of lightning activity (Fig. 8b). The pattern of D0 within the ZDR columns is similar to that of the ZDR values, depending on the strong linear relationship between D0 and ZDR (Fig. 8c). The liquid water content within the ZDR columns ranges from approximately 0.1 to 4 gm⁻³, and the peaks correspond to the two peaks of lightning activity (Fig. 8d).

In addition, the percentage of hydrometeors within the ZDR columns is investigated on the basis of the retrieved contents of ice (including graupel) or raindrops, as described in Sect. 2.2. The results of hydrometeor identification are dominated by large-size particles. Thus, we count the grids of ice (graupel) or raindrops via the results of the ZDP method to investigate the percentage of hydrometeors within the ZDR columns, avoiding neglecting the grid that possesses both ice (graupel) particles and liquid drops simultaneously. The obvious phenomenon is that the percentage of graupel within the ZDR columns suddenly peaks before the first peak of lightning activity, but the second peak of lightning activity is not related to the presence of graupel within the ZDR columns; the hydrometeor type within the ZDR column at 18:30 CST is raindrops (Fig. 9). Notably, the results neglect raindrops smaller than 1 mm; although these small raindrops are a minority within the ZDR column. This indicate the collapse of the ZDR column within the reflectivity core at 18:30 CST, which is consistent with the results shown in Fig. 4d; the large-sized ice particles (i.e. graupel, corresponding to the reflectivity core) fall from aloft, and the smaller ZDR values of such ice particles result in this collapse phenomenon. Thus, the retained ZDR column corresponds to the periphery of the reflectivity core, indicating relatively weak updraughts and raindrops. Notably, the absence of graupel within the ZDR columns does not mean that the graupel particles are eliminated within the clouds. On the other hand, the sudden increase in graupel within the ZDR column at 18:00 CST may support the presence of a coalescence-freezing mechanism that led to graupel formation in the warm-based clouds. The hypothesis about the coalescence–freezing mechanism was proposed in previous studies (e.g. Braham, 1986; Bringi et al., 1997; Carey and Rutledge, 2000; Herzegh and Jameson, 1992).

To determine the relationship between lightning activity and the quantified ZDR columns, the height and volume of the ZDR column are calculated via the 3D mapping column method; the volume is based on the accumulation of all grids within the ZDR column, and the volume of a single grid is 0.03125 km³, with 0.25 km horizontal and 500 m vertical resolutions. The height of the ZDR column is determined by counting the grid number (n) from the melting level to the highest grid within the ZDR column; if n is determined, the ZDR column height is n×0.5 km.

The variations in the ZDR/KDP column height and volume with the life cycle of the remaining 14 cases are displayed in Fig. 10 (cases #2 to #15), as are the variations in the percentages of hydrometeor types and microphysical fingerprints. The grid is assigned to specific particle type based on the results of hydrometeor identification, and the percentage of grids for each hydrometeor type is calculated. Similarly, this process is applied to determine the percentage of grids associated with microphysical fingerprints. Each of them has a ZDR column (Fig. 10a1–a14); however, the absence of a KDP column is possible (Fig. 10b1–b14). The results of our study, from 15 isolated thunderstorms over South China, indicate lightning is not observed in the absence of a ZDR column, and that a KDP column is not observed without a ZDR column, which is consistent with the observations of Bruning et al. (2024). Moreover, although the highest lightning flash frequency (in case #11) is observed when the ZDR and KDP columns are copresent, the total flashes in cases (e.g. #6, #10, and #15) with only the ZDR column are not lower than those in cases (e.g. #2, #3, and #5) where both ZDR and KDP columns are present. In addition, our results suggest that the signal in the KDP column within these small, isolated, subtropical thunderstorms over South China is not as steady as that in the ZDR column during the life cycle.

The results show that the percentages of identified graupel particles and riming process are closely related to lightning activities (Fig. 10c1–c14, d1–d14); consistent with that shown in Fig. 7. The cross-correlation approach can be used to examine the correlation considering the time lag, which is important for verifying whether a parameter is appropriate for forecasting another parameter. To further determine the correlation between lightning activity and the polarimetric structure. The cross correlations between the lightning activity and polarimetric structure during the life cycles in all the cases are examined, and the results are displayed in Fig. 11.

Figure 11a shows that the variation in the graupel or rain water content above the melting level within the cloud can predict the lightning activity (total flashes) after 6 min well, and the correlation coefficient is approximately 0.8. However, other parameters (e.g. ZDR column volume, ice content above the melting level, and graupel volume) also exhibit good performance in forecasting lightning activity, and the correlation coefficient can reach approximately 0.7. The graupel volume is calculated based on the identification results of hydrometeors. Although the variation in the graupel or rain water content above the melting level within the cloud can also forecast the lightning activity (CG flashes) after 6 min, the correlation coefficient decreases to approximately 0.56 (Fig. 11b). Notably, the trend of the ZDR column volume implies that it may perform well with a longer warning time (e.g. 12 min) for lightning activity. In addition, the autocorrelation of each variable in time series would overestimate the strength of the relationship by approximately 0.15.