WPR continuously measures horizontal wind and vertical velocity profiles in the lower and middle troposphere at high temporal and vertical resolution (Liu et al., 2020), making them particularly suitable for studying the evolution of pre-precipitation wind fields associated with mesoscale weather systems. In this study, we used data from 14 WPR deployed throughout Guangdong Province during 2016–2020 (Fig. 1): boundary-layer wind profilers (BWPR, L–C band, 0–6 km, 60 m vertical resolution, ∼ 6 min native temporal resolution) at most sites, and a tropospheric profiler (TWPR, P–A band) at Luogang (LG). The Guangdong Meteorological Administration provides two operational products: ROBS (real-time observation, raw profiles at the native resolution) and HOBS (half-hour observation, quality-controlled, 30 min effective resolution). HOBS was used for calculations of all indices, limited to 0–6 km due to reliability constraints at higher altitudes for L–C band systems. The 14 stations and their abbreviations are as follows:

Zhanjiang (ZJ), HaiLingDao (HLD), Luoding (LD), Xinhui (XH), Zhuhai (ZH), Nansha (NS), Huadu (HD), Conghua (CH), Luogang (LG), Zengcheng (ZC), Shenzhen (SZ), Huizhou Longmen (HZ), Wuhua (WH), and Chaozhou (CZ).

All WPR data have undergone strict quality control. Records marked invalid by the instrument were removed. Data exceeding 6000 m in range, horizontal wind speeds outside the range of 0–40 m s⁻¹, and vertical wind speeds outside the range of -10 to 10 m s⁻¹ were also discarded. Subsequently, a two-dimensional median filter was applied in the time–height domain: for each data point, the median value of its valid neighbors within ±2 time steps and ±2 range gates was calculated. If there were at least three valid neighbors, any values that deviated from the median by more than 12 m s⁻¹ (horizontally) or 5 m s⁻¹ (vertically) were set to missing. This process effectively removed persistent contamination from ground clutter. After quality control, the overall availability of data during the analysis period was relatively high.

The number of high-density automatic weather stations (AWS) in Guangdong Province gradually increased between 2016 and 2020, reaching over 3000 in 2020. This study used hourly precipitation data from over 3000 AWS stations across the province to identify WSHR events and calculate precipitation weighted centroids. The locations of AWS stations and WPR, as well as the spatial distribution of urban pixels, are shown in Fig. 1. Urban extent in this study is characterised using the Local Climate Zone (LCZ) classification of Stewart and Oke (2012), with LCZ 1–10 (built-type classes) treated as urban pixels. LCZ maps for Guangdong Province are obtained from Demuzere et al. (2022a). This classification is used both for the province-wide visualisation in Fig. 1 and for the 3 km urban fraction reported per profiler station in Sect. 4.5 and Fig. 16 (areal percentage of LCZ 1–10 pixels within a 3 km buffer centred on each station).

Large-scale fields are obtained from the ERA5 reanalysis (Hersbach et al., 2023a, b) at 0.25° × 0.25° horizontal resolution and 1 h temporal resolution, including geopotential height and winds at 925, 850, and 500 hPa, specific humidity for moisture transport and surface-based convective available potential energy (CAPE). Vertically integrated water vapour transport (IVT) is calculated between 1000 and 850 hPa as: 1IVTu=1g∫1000hPa850hPaqudp,2IVTv=1g∫1000hPa850hPaqvdp,3IVT=IVTu2+IVTv2, where the IVTu (IVTv) are zonal (meridional) component of the moisture transport vector in kg m⁻¹ s⁻¹, IVT is the magnitude of the vector, g is gravitational accelerations in m s⁻², u and v are zonal and meridional components of the layer-mean wind speed, respectively, q is the layer-mean specific humidity in kg kg⁻¹, and dp is pressure difference between two neighboring pressure levels. Positive values of IVTu (IVTv) indicate eastward (northward) zonal (meridional) components.

In addition, vertical moisture transport profiles are characterized by each unintegrated term in Eqs. (4)–(6) for diagnosis in horizontal and vertical components for cross-sectional analysis of moisture transport: 4VTu=1gqu,5VTv=1gqv,6VTw=1gqω, where VTu/v are horizontal components of moisture transport in each vertical layer in kg m⁻¹ s⁻¹ Pa⁻¹, VTw is vertical component of moisture transport in kg m⁻² s⁻¹, ω is vertical components of velocity in Pa s⁻¹.

WSHR events during the pre-flood season (primarily April–June) of 2016–2020 are identified in two stages. First, candidate warm-sector days are screened according to the method of Liu et al. (2019), which requires the following conditions to be met: (1) the influence of typhoons or tropical depressions is excluded; (2) the daily precipitation of three or more adjacent stations within 100 km is ≥ 50 mm, and there is a period of continuous 3 h with hourly precipitation ≥ 5 mm and a 3 h cumulative total ≥ 30 mm; (3) the warm-sector conditions are confirmed through vertical profiles of θse and temperature advection along the representative station longitude, that is: if there is no obvious frontal system on the profile, the precipitation area is dominated by southerly winds at low-to-middle levels, and the precipitation area is more than 200 km away from the nearest surface northerly wind; if there is an obvious frontal system on the profile, the precipitation area is more than 200 km away from the front. It should be noted that the detection capability of this method is limited by the density and distribution of the precipitation station network.

Secondly, individual events for each candidate day were objectively defined in three Guangdong sub-regions: western Guangdong (western GD, 21.0–22.5° N, 110.0–113.5° E), central Guangdong (central GD, 22.5–24.5° N, 112.5–114.5° E), and eastern Guangdong (eastern GD, 22.5–24.0° N, 115.0–117.0° E), as shown in Fig. 1. These regions correspond to the main areas of high precipitation observed in the province's WSHR. For each sub-region and candidate day, hourly regional average precipitation was calculated from all AWS sites within the region. The onset of an event was defined as when the average precipitation rate reached or exceeded 1.0 mm h⁻¹ for at least three consecutive hours. The event was considered to end after three consecutive hours below 1.0 mm h⁻¹ ; brief intervals of less than 3 h were considered intra-event intervals. Events with a regional average cumulative precipitation of < 5 mm were excluded. The precipitation weighted centroid was calculated as the cumulative precipitation weighted average of the station coordinates. For each event, the WPR station closest to the precipitation weighted centroid (limited to the same sub-region and ≤100 km away) was designated as the representative station for the precursor index calculation; events not meeting this condition were excluded from the analysis. Subsequent analysis (Sect. 4) focused primarily on precursor signals preceding precipitation, as precipitation particles and related contamination significantly degrade radar data quality after precipitation begins. The procedure generated a total of 226 independent WSHR events (65 in western GD, 88 in central GD, and 73 in eastern GD).

Based on the dynamic and thermodynamic conditions of precipitation, this paper selects five indices for study: LLJI, two VWS layers (0.5–1.5 and 1.5–3.0 km), ALI, and BLH. These indices represent different physical processes–LLJ intensity, vertical shear structure, convective uplift, and boundary layer depth. Their relative importance in each region will be empirically assessed in Sect. 4. The indices were calculated from quality-controlled WPR data from representative stations for each event, at native ∼ 30 min resolution and aligned with the precipitation onset time t=0. Raw index values were winsorised at the 5th and 95th percentiles across all events before composite analysis to reduce the influence of extreme outliers. Sensitivity tests using non-winsorised data (not shown) produced consistent temporal patterns, indicating that the results are insensitive to this preprocessing choice.

Composite time series are constructed by aligning all events at the beginning (t=0) and computing the median value of each index in each time step. Uncertainty in the composite median is quantified using bootstrap resampling (2000 resamples with replacement); the 2.5th and 97.5th percentiles of the bootstrap distribution define the 95 % confidence interval.

Associations between index values and subsequent rainfall intensity are assessed using the Spearman rank correlation coefficient. Pre-onset values are averaged over four non-overlapping periods (t=-6 to -4 h, -4 to -2 h, -2 to -1 h, and -1 to 0 h) before computing the correlations. The boundaries of these periods are motivated by the composite wind field analysis (Sect. 4.1, Fig. 6), which identifies t≈-2 h as the onset of the most active pre-onset wind field evolution; the innermost period (-1 to 0 h) also represents a practically relevant nowcasting lead time for WSHR. Significance is tested at the 95 % level (p < 0.05).

Regional differences in precipitation characteristics and precursor indices were tested using Kruskal–Wallis one-way ANOVA, with paired comparisons performed where appropriate. Strong and weak precipitation events within each subregion were demarcated by the median of the maximum hourly precipitation intensity recorded for each event; events above the median were classified as strong precipitation, and those below as weak precipitation. Differences between strong and weak precipitation event groups were assessed using the Mann–Whitney U test. All tests were two-tailed nonparametric tests. Sample sizes varied slightly across different figures because each analysis required continuous WPR data within different time windows; events with significant data gaps within the required windows were excluded. Exact numbers, determined independently for each analysis, are reported in the figure captions.

Figure 6 presents composite time–height cross-sections of horizontal wind and vertical velocity anomalies for the three regions. The vertical velocity anomaly is defined as the departure from the median over t=-6 to -4 h at each height level, thus removing the instrument-related background and highlighting the changes associated with approaching convection. Two features are shared in all three regions: (i) a progressive strengthening of low-level winds in the hours before onset, and (ii) a transition from near-zero to positive vertical velocity anomaly beginning 1 to 3 h before onset.

In western GD (Fig. 6a), southerly winds prevailed below 2 km before precipitation, shifting clockwise to westerly winds above 3 km, consistent with low-level warm advection. Several hours before precipitation, wind speeds gradually increased at the 2–4 km level, with significant fluctuations in the 10 m s⁻¹ contour around t=-2 h. In central GD (Fig. 6b), the low-level jet began to descend around t=-4 h, with winds exceeding 12 m s⁻¹ appearing above 3.5 km several hours before precipitation. After the onset of precipitation, the vertical ascent signal was the strongest among the three regions, with positive anomalies rapidly expanding upward. In eastern GD (Fig. 6c), the wind field is more vertically uniform: south-southwest winds extend from the surface to above 3 km with modest directional shear. The low-level jet persists steadily near 2–3 km throughout the pre-onset period. A notable feature is the negative vertical velocity anomaly at 3–5 km during t=-3 to 0 h, indicating a mesoscale suppression of ascent in the middle troposphere prior to onset, followed by a positive anomaly that develops below 3 km and intensifies through onset.

The composite VWS profiles (Fig. 7) quantify the vertical shear structures in the pre-onset period. VWS is computed between adjacent 60 m range gates and averaged over t=-3 to 0 h; this window is chosen because it encompasses the period of most active pre-onset wind field evolution identified in Fig. 6.

Three local VWS maxima are identified below 3 km in all regions. The first maximum occurs below 0.5 km and is more pronounced in the central GD (median ∼ 20 × 10⁻³ s⁻¹), reflecting the strong near-surface wind gradient over the PRD surface. The second maximum lies in the 0.5 to 1.5 km layer and is present in all three regions, corresponding to the wind speed gradient between the surface layer and the jet core. The third maximum, in the 1.5 to 3.0 km layer, is strongest in eastern GD.

These three maxima motivate the choice of analysis layers for Sect. 4.2: 0.5–1.5 km (Layer 1, capturing BLJ-related shear) and 1.5–3.0 km (Layer 2, capturing the BLJ–free-troposphere transition). The lower bound of 0.5 km is adopted because profiler measurements in the lowest few hundred meters are subject to ground clutter and large frictional variability (Liu et al., 2020). The upper bound of 1.5 km encompasses the jet core, which typically resides near 1 km over coastal South China (Du and Chen, 2018).

In summary, the composite analysis reveals coherent pre-onset signals across all three regions–progressive low-level wind intensification, a transition from suppressed to enhanced ascent, and distinct vertical shear layering–but with regional differences in timing, vertical structure, and magnitude. To quantify these signals systematically, Sect. 4.2 examines the temporal evolution of five precursor indices.

To quantify the pre-onset signals identified in Sect. 4.1, composites of precursor indices are constructed for each region (Fig. 8). Five quantities are examined: LLJI, VWS at 0.5 to 1.5 km and 1.5 to 3.0 km, ALI, and BLH. All indices are computed at 30 min resolution and aligned by onset time. In each panel, the solid line shows the composite median, and the dark shading indicates bootstrap 95 % confidence interval of the median (2000 resamples), and the light shading shows the interquartile range.

The LLJI over three regions in GD increases with development of WSHR events (Fig. 8a–c), the LLJI in western GD increasing the sharpest among three regions (median value approximately 1.0 × 10⁻² s⁻¹). The VWS (solid line in Fig. 8d–f) at 0.5–1.5 km exceeded the values at 1.5–3.0 km (dashed line) in all regions and at all times, reflecting the strong wind speed gradient between the surface layer and the jet core. The highest VWS was recorded in central GD (median value approximately 5.5–6.0 × 10⁻³ s⁻¹). Neither layer showed systematic variation in the 1–2 h preceding precipitation. The ALI (Fig. 8g–i) was intermittent, with intermittent pulses but no sustained accumulation in western and central GD; it remained low and flat in eastern GD. BLH (Fig. 8j–l) showed a weak pattern of initial rise followed by decline in all regions–most pronounced in central GD. For example, in central GD, the median value reached approximately 800–850 m near t=-3 h, then dropped to approximately 700 m at the onset of precipitation. However, the bootstrap confidence interval was wide, and the signal was indistinguishable from the sampling variability at the 95 % level. A notable feature of Fig. 8 is the large IQR relative to the temporal changes of the medians, reflecting the diversity of synoptic settings, onset times, and rainfall intensities among WSHR events. Stratified analyses addressing this heterogeneity are presented in Sect. 4.4. Whether pre-onset index values are associated with subsequent rainfall intensity is examined in Sect. 4.3.

To assess whether the pre-precipitation index value is related to the subsequent precipitation intensity, Spearman rank correlations are computed between each index and the maximum event-accumulated rainfall, defined as the highest event-total precipitation recorded across all AWS within the sub-region during the event period. The pre-onset window is divided into four periods: t=-6 to -4 h, t=-4 to -2 h, t=-2 to -1 h, and t=-1 to 0 h. For each event, the index value is averaged over each period, and the correlation with rainfall intensity is calculated for all events in a given region. This period-averaging approach reduces the number of independent tests relative to a point-by-point sliding-window analysis, while providing a more robust estimate of the index value at each lead time. Significance is assessed at 95 % level (p < 0.05).

The results are summarized in Fig. 9. In central GD (Fig. 9b), VWS at 0.5 to 1.5 km is significantly correlated with rainfall intensity in three consecutive periods (r=0.28 at t=-4 to -2 h; r=0.25 at t=-2 to -1 h; r=0.35 at t=-1 to 0 h; all p < 0.05), the correlation strengthening toward onset. This sustained positive correlation indicates that stronger low-level shear over the PRD is associated with more intense rainfall.

In eastern GD (Fig. 9c), BLH shows the strongest association with rainfall intensity among all region–index combinations. The correlation is significant in the two earliest periods (r=0.46 at t=-6 to -4 h; r=0.39 at t=-4 to -2 h; both p < 0.05) but weakens closer to onset. Since eastern WSHR events predominantly occur in the early morning (Sect. 3), when the boundary layer is typically shallow, an anomalously high BLH before onset implies enhanced pre-onset thermodynamic preconditioning of the lower troposphere. The significant correlations at longer lead times suggest that BLH acts as an observational indicator of the pre-convective thermodynamic environment: stronger energy accumulation produces both a deeper boundary layer and, subsequently, more intense rainfall.

Across all three regions, LLJI is not significantly correlated with rainfall intensity during any period. This does not imply that the low-level jet is irrelevant to rainfall: the composite analysis in Sect. 4.2 shows that LLJI increases before onset in western GD, and stratified analysis in Sect. 4.4 confirms that the LLJI is significantly higher for strong-rainfall events. The absence of a linear Spearman correlation suggests that the relationship between LLJI and rainfall intensity may be nonlinear.

In summary, time-dependent correlation analysis identified two robust associations: the 0.5–1.5 km VWS in central GD and precipitation intensity, and the BLH in eastern GD. Both signals persisted over multiple consecutive time periods. While the LLJI is the clearest temporal precursor – showing a marked increase before onset (Sect. 4.2) – it does not serve as a linear intensity predictor, as it shows no significant rank correlation with rainfall amount. This distinction suggests that LLJI may play a nonlinear, threshold-type role in convective triggering rather than scaling proportionally with rainfall intensity. The physical explanation for these regional differences is discussed in Sect. 4.4.

Section 4.2 and 4.3 have shown that the nature of the precursor signals differs in the three regions: LLJI is the clearest temporal precursor in western GD, VWS at 0.5–1.5 km is the strongest correlate of rainfall intensity in central GD, and BLH shows the strongest intensity association in eastern GD. This section investigates the physical processes underlying these regional contrasts through stratified composites, vertical structure analysis, and ERA5 diagnostics.

Stratified composites (Fig. 10) show the most significant differences in the selected index-stratified combinations; complete stratification results for all indices and the three stratification methods are shown in Figs. S1–S3 in the Supplement. Pre-monsoon events exhibit larger LLJI fluctuations before precipitation than post-monsoon events (Fig. 10a, b). The strongest contrasts emerge when stratified by precipitation intensity (Fig. 10d–f): strong precipitation events in western GD show higher LLJI starting 2 h before precipitation, strong events in central GD show persistently higher 0.5–1.5 km VWS throughout the entire pre-precipitation window, and strong events in eastern GD show higher BLH 2 h before precipitation. Stratification by onset time (Fig. 10g–i), where daytime and nighttime events are defined according to the onset time distribution identified in Sect. 3, indicates that nighttime events in central GD are accompanied by a significantly enhanced 0.5–1.5 km VWS approximately 2 h before precipitation. These patterns prompt a more in-depth examination of the vertical structure of the low-level jet and its relationship with precipitation intensity.

The composite wind speed profiles of strong and weak precipitation events (Fig. 11) reveal different characteristics of the low-level jet in the three regions. In western GD (Fig. 11a), the strong precipitation group systematically shows higher wind speeds below 1.5 km. In eastern GD (Fig. 11c), the strong precipitation group has a lower jet core height (2.40 km vs. 2.94 km for the weak events) and stronger peak wind speeds (12.8 vs. 11.2 m s⁻¹). In central GD (Fig. 11b), the two groups of strong events show slightly higher wind speeds below 0.5 km, and the precipitation intensity in this region cannot be distinguished solely by the overall structure of the jet.

A further diagnostic is provided by the change in upper-level (3–5 km) vertical velocity between the baseline period (t=-6 to -4 h) and the near-onset period (t=-2 to 0 h), denoted Δw (Fig. 12). A positive Δw indicates upward strengthening motion approaching the onset; a negative Δw indicates weakening. The correlation between Δw and rainfall intensity reveals a striking regional contrast. In western GD (Fig. 12a), Δw is positively correlated with rainfall (r=0.37, p=0.006): events with strengthening ascent before onset produce heavier rainfall. Moreover, Δw is positively correlated with ΔLLJI (r=0.34, p=0.012; Fig. S4), and this association holds after accounting for synoptic-scale geopotential height trends at 925 and 850 hPa (partial r=0.27–0.31, p < 0.05; not shown).

The vertical structure of Δw is consistent with this: Δw is close to zero in the jet core (0.8–1.5 km) and increases sharply above 1.5 km (Fig. 13), consistent with the upward motion driven by convergence above the jet core. In strong precipitation events, the larger Δw above 1.5 km is more pronounced. Combined with the fact that there is no difference in CAPE between strong and weak events in this region (Fig. 14a, p=0.66), these results support the view that the enhancement of LLJ in western GD directly amplifies the low- to mid-level upward motion, resulting in stronger precipitation.

In eastern GD (Fig. 12c), the relationship is reversed: Δw is negatively correlated with precipitation (r=-0.43, p=0.001). This decrease in Δw is not driven by changes in the low-level jet stream, as ΔLLJI and Δw are uncorrelated in this region (Fig. S4, r=0.15, p=0.26), suggesting that the suppression mechanism operates independently of local jet dynamics. The pre-precipitation CAPE of strong precipitation events is significantly higher than that of weak events (Fig. 14c), and the Δw of strong precipitation events is significantly more negative (Fig. S5d). All these indications suggest that the suppression of upward motion in the middle and upper levels prevents the premature release of convective instability, causing instability energy to build up in the lower troposphere; the significantly higher pre-precipitation CAPE and anomalously deep BLH in strong precipitation events are consistent with this accumulation process. ERA5 diagnostics in eastern GD (Fig. S5) showed no significant differences in large-scale 500 hPa vertical velocity (p=0.80) and 925 hPa relative humidity (p=0.21) between strong and weak precipitation events. Only 925 hPa temperature showed a significant difference (median 22.5 vs. 21.8 °C; p=0.004), with strong events consistently being warmer. In central GD, neither Δw nor CAPE distinguishes between strong and weak events (Fig. 12b, p=0.75; Fig. 14b, p=0.96). The relationship between ΔLLJI and ΔVWS is region-dependent (Fig. S6): in western GD they are significantly anti-correlated (r=-0.26, p=0.044), in central GD they are uncorrelated (r=0.03, p=0.76), and in eastern GD they are also uncorrelated (r=0.03, p=0.84). These contrasts indicate that central-GD low-level shear is modulated by surface friction rather than LLJ pulsing, while in western and eastern GD, LLJ pulses accelerate the boundary-layer and lower-free-troposphere winds more uniformly.

The composite IVT fields (Fig. 15) show that central GD lies at the confluence of southwesterly flow from the Indian Ocean and southeasterly flow steered by the western Pacific subtropical high (consistent with the moisture transport climatology of Ning et al., 2023). Within this convergent large-scale environment, the low-level wind shear is strongest in central GD among the three regions (Fig. 7). This shear may partly arise from frictional deceleration of the low-level jet over the PRD surface, which would enhance convergence beneath the jet core. To further characterize the spatial structure of pre-onset boundary-layer VWS enhancement and the associated divergence pattern across the PRD, a multi-event profiler analysis is presented in Sect. 4.5.

In summary, the regional differences in precursor signals reflect fundamentally different relationships between low-level jets and precipitation intensity. In western GD, low-level-jet enhancement independent of synoptic-scale forcing promotes mid-to-upper-level upward motion, with a Δw vertical structure consistent with convergence-driven ascent. In eastern GD, pre-precipitation upward motion weakens, accompanied by higher boundary-layer temperatures and associated CAPE accumulation, while large-scale dynamics and low-level moisture conditions show no significant difference between strong and weak events. In central GD, governed by frictional shear modulation, the magnitude of lower-level VWS determines precipitation intensity.

To verify that the correlation between the lower-level VWS and precipitation intensity in central GD in Fig. 9 reflects a boundary-layer signal rather than a byproduct of synoptic-scale forcing, Spearman partial correlations (Fig. S7) were calculated after controlling for the mean ERA5 geopotential height (Z850 and Z925) in central GD. The 0.5–1.5 km VWS remained significantly correlated with precipitation in the P2–P4 windows, while the 1.5–3.0 km control layer showed no significant correlation in any window, confirming that the signal is confined to the boundary layer.

To examine the spatial structure of pre-onset VWS enhancement, per-station anomaly composites at 0.5–1.5 km were computed by subtracting each profiler's onset-hour-matched background median from the 3 h pre-onset mean (Fig. 16). All seven stations showed positive median anomalies, confirming that VWS enhancement is a widespread feature across central GD. The largest median anomaly (+1.81 m s⁻¹ km⁻¹) was observed at Luogang (LG), located within the Changling National Trail System, a forested recreation area in eastern Guangzhou, with the lowest 3 km urban fraction (22 %, LCZ 1–10) among all seven stations. This is consistent with the VWS enhancement at LG arising from the frictional environment of the surrounding urban agglomeration through which the southerly inflow has passed, rather than from the local surface beneath the profiler. Three other stations located well within the PRD urban agglomeration – Huadu (HD), Conghua (CH), and Shenzhen (SZ) – also showed sizeable positive anomalies. Stratifying events by hourly peak rainfall (median split at 71 mm h⁻¹), six of seven stations showed larger anomalies during strong precipitation events, consistent with the VWS–rainfall relationship documented in Fig. 9.

The wind speed anomaly profiles along the dominant southerly inflow path (Fig. 17a; station and triangle locations shown in Fig. S9) show that WSHR events enhance the low-level flow at the coastal station NS by approximately +2.4 m s⁻¹ in the lowest 0.5 km, but this enhancement is progressively attenuated inland – to roughly +0.8 m s⁻¹ at LG and near zero at HD. At heights above ∼ 1 km, the three stations show comparable enhancement. Such spatially non-uniform deceleration within the boundary layer implies, by mass continuity, enhanced convergence where the inflow decelerates rapidly and relative divergence where it remains strong. The absolute divergence composites (Fig. S8) show that the four Bellamy triangles already possess a strong climatological gradient: the background median divergence ranges from +2.5 × 10⁻⁵ s⁻¹ (divergent) at LG–CH–ZC to -2.5 × 10⁻⁵ s⁻¹ (convergent) at LG–HD–CH, indicating that the northern triangle is climatologically convergent due to its location in the terrain-transition zone north of HD. Against this baseline, WSHR events impose an additional south-to-north divergence anomaly dipole (Fig. 17b). At 1.15 km, the four triangles trace a monotonic gradient: from +2.2×10-5 s⁻¹ (divergence enhancement) at the southern coastal triangle NS–ZC–SZ, through -1.5 × 10⁻⁵ s⁻¹ at NS–LG–ZC and -2.6 × 10⁻⁵ s⁻¹ at LG–CH–ZC, to -3.8 × 10⁻⁵ s⁻¹ (convergence enhancement) at LG–HD–CH. This dipole is vertically confined to below ∼ 1.5 km, matching the layer of spatially non-uniform deceleration in Fig. 17a, and its convergent northern branch provides a favourable dynamical environment for convective initiation during WSHR events. HD itself combines a 99 % 3 km urban fraction with an immediate transition to higher terrain to the north, and whether the climatological convergence is controlled primarily by the urban surface, the orographic discontinuity, or their interaction cannot be determined from the present profiler network. The spatially non-uniform deceleration diagnosed in Fig. 17a may arise from a combination of urban roughness, orographic forcing, and land–sea contrast; disentangling these contributions is beyond the scope of the present observational study and requires future numerical sensitivity experiments.

This study investigates the precursor signals of WSHR over Guangdong Province using WPR observations of 226 events during 2016–2020 across three subregions: western, central, and eastern GD. Five profiler-derived indices–LLJI, VWS at 0.5 to 1.5 km and 1.5 to 3.0 km, ALI, and BLH–are analysed through composite time series, period-based rank correlations, stratified composites, and ERA5-based thermodynamic diagnostics.

Among the five indices, LLJI exhibits the clearest temporal evolution before onset, roughly doubling its baseline value in western GD with the increase concentrated in the final 1–2 h before rainfall. However, this temporal trend is not statistically robust in central and eastern GD. LLJI does not show a significant rank correlation with rainfall intensity in any region; however, the stratified analysis confirms that strong-rainfall events are characterised by systematically higher overall LLJI levels than weak-rainfall events. VWS at 0.5 to 1.5 km maintains a significant positive correlation with rainfall intensity across three consecutive pre-onset periods in central GD, making it the most persistent intensity predictor in that region. In eastern GD, BLH shows the strongest association with rainfall intensity among all region–index combinations, with the correlation most pronounced in the early pre-onset window (t=-6 to -4 h); strong-rainfall events are also characterized by significantly higher pre-onset CAPE than weak-rainfall events. ALI displays episodic pulses rather than a sustained trend, and does not show consistent predictive value in any region.

The analysis of the pre-onset changes in upper-level (3–5 km) vertical velocity (Δw) reveals a fundamental contrast in the rainfall-producing mechanisms between regions. In western GD, Δw is positively correlated with rainfall intensity and with the concurrent change in LLJI, the Δw-ΔLLJI correlation holds after accounting for synoptic-scale geopotential height trends, and the Δw vertical structure is consistent with low-level convergence forcing upper-level ascent. In eastern GD, the relationship is reversed: weakening of the upper-level ascent before the onset is associated with stronger rainfall. This suppression is independent of both local LLJ changes and large-scale subsidence. Strong precipitation events exhibit significantly higher 925 hPa temperatures, and boundary-layer warming drives the CAPE accumulation and boundary-layer deepening observed in strong events. In central GD, neither Δw nor CAPE distinguishes between strong and weak events; instead, the 0.5–1.5 km VWS serves as the primary modulator of precipitation intensity. The LLJI and VWS are temporally decoupled in this region. Multi-event profiler composites show that the WSHR-enhanced southerly inflow undergoes spatially non-uniform frictional deceleration as it crosses the PRD, producing a south-to-north divergence anomaly dipole with convergence enhancement concentrated in the northern PRD (Sect. 4.5).

Among the five indices, three emerge as regionally dominant precursors: LLJI in western GD, VWS at 0.5–1.5 km in central GD, and BLH in eastern GD. These results provide an observational basis for developing region-specific nowcasting guidance: LLJI temporal changes for western GD, VWS at 0.5–1.5 km for central GD, and BLH combined with low-level temperature for eastern GD.

The three rainfall processes identified here are broadly consistent with, but extend, previous findings on WSHR mechanisms. Pulsed LLJ intensification in western GD is in agreement with BLJ–terrain interaction and orographic lifting mechanisms widely documented in coastal South China (Liang and Gao, 2021; Pu et al., 2022; Zhang et al., 2022b). The frictional deceleration effect documented across central GD events supports the findings of Gao et al. (2021) and Huang et al. (2019), who attributed convective initiation in Guangzhou to the urban heat island and enhanced convergence. Furthermore, convective maintenance in this urban agglomeration is often aided by evaporative cooling and cold pool outflows from upstream precipitation (Wang et al., 2014), though our current analysis highlights the primary triggering role of friction-induced VWS rather than thermal or cold-pool forcing alone. Regarding eastern GD, our finding of an energy-accumulation process linked to boundary layer warming rather than large-scale moisture transport or subsidence complements and qualifies the previous emphasis on large-scale moisture transport and spatial configurations. Past studies have demonstrated that robust moisture transport from the South China Sea to the Shanwei coast is crucial for triggering warm-sector rainstorms in the eastern region (Ning et al., 2023; Wang et al., 2025; Zhong and Chen, 2017).

This study uses profiler-observed vertical velocity changes to differentiate between direct dynamical forcing and energy-accumulation mechanisms. Previous studies have generally treated BLJ as a uniform triggering factor (e.g., Sun et al., 2019; Zhang and Meng, 2019), while the opposing signs of the Δw–rainfall correlation in western and eastern GD demonstrate that the role of LLJ varies qualitatively in the region. WPR, with its approximately 6 min temporal resolution and continuous vertical sampling from the surface to 5 to 6 km, is well suited to resolving these pre-onset signals that are not captured by twice-daily radiosondes or hourly reanalysis fields.

This study has several limitations. First, while the 14-station wind profiler network provides high temporal and vertical resolution, its spatial sampling is limited. Observations are lacking on the windward slopes of the Yunwu and Lianhua Mountains, in the central urban area of Guangzhou, and in the gaps between profiler stations, precluding a complete characterization of inflow evolution from ocean to inland. Second, the five-year record (2016–2020) yields limited sample sizes (65–88 events) in each sub-region, which constrains the statistical power of stratified analyses and may not capture the full range of interannual variability in large-scale forcing. Third, as discussed in Sect. 4.5, the observational approach cannot separate the individual contributions of urban roughness, orographic forcing, and land–sea contrast to frictional deceleration and convergence enhancement in central GD; the relative importance of these factors remains unquantified.

Several directions for future work emerge from these findings. High-resolution numerical sensitivity experiments – systematically removing or modifying urban land use, terrain, and coastline geometry – are needed to disentangle the respective contributions to boundary-layer convergence over the PRD. The suppression mechanism that allows CAPE to accumulate before precipitation in eastern GD requires verification with independent thermodynamic profiling to confirm the boundary-layer warming signal beyond ERA5 diagnostics. Finally, translating the regionally distinct precursor signals identified here into operational nowcasting algorithms – for example, real-time monitoring of LLJI trends in western GD, 0.5–1.5 km VWS in central GD, and BLH combined with low-level temperature in eastern GD – represents a natural next step, though the practical skill of such guidance needs to be evaluated through independent case testing and probabilistic verification.