Typhoon short-term heavy precipitation warning based on dual-polarization radar observations of precipitation particle vertical distribution Authors: Rong Fang, Ting Wang, Haoya Liu Published: 2025 ABSTRACT AbstractMesoscale convective systems within typhoon spiral rainbands often lead to localized short-term heavy precipitation, and their accurate prediction remains a challenging issue in operational meteorology. This study utilized S-band dual-polarization radar to observe typhoon processes and systematically analyzed the differences in vertical distribution characteristics of precipitation particles between inner and outer spiral rainbands. Results showed that inner rainbands were characterized by deep convective development (11.8 ± 1.2 km), active ice-phase processes (graupel particle content of 55% ± 10% at 5 km altitude), and strong melting layer features (Zdr peak value of 3.8 ± 0.5 dB); while outer rainbands were dominated by shallower convection, weaker ice-phase processes, and warm rain processes. Based on the vertical distribution characteristics of particles, a multi-parameter warning index system was constructed, incorporating parameters such as graupel particle content and zero-degree layer bright band intensity. Independent sample verification showed that the Probability of Detection (POD) of this warning method reached 84.6%, an improvement of 19.2 percentage points compared to traditional methods, with warning lead time extended to 25–40 min. This study elucidated the microphysical formation mechanisms of typhoon short-term heavy precipitation, providing a scientific basis for enhancing typhoon precipitation forecasting and warning capabilities.Keywords: Dual-polarization radar, Typhoon precipitation, Short-term heavy precipitation warning, Microphysical characteristicsSubject terms: Climate sciences, Environmental sciences, Natural hazards Introduction As a major severe weather phenomenon affecting the southeastern coastal regions of China, typhoons and their mesoscale convective systems within spiral rainbands often lead to localized short-term heavy precipitation, posing serious threats to people’s lives and property. The non-uniform distribution characteristics of typhoon precipitation make accurate forecasting of short-term heavy precipitation a challenging issue in operational meteorology1,2. Previous studies have shown that there are significant differences in precipitation microphysical characteristics across different regions of typhoons. The inner spiral rainbands near the typhoon eyewall exhibit deep convective development and active ice-phase microphysical processes3,4, while the outer spiral rainbands are dominated by relatively shallow convection and warm rain processes. The relationship between these differences in microphysical processes and surface precipitation intensity warrants in-depth investigation5. Dual-polarization radar, with its multi-parameter detection capability, provides an effective observational tool for precipitation microphysics research. Compared to conventional weather radar, dual-polarization radar can obtain polarimetric parameters such as differential reflectivity (Zdr), differential phase (φdp), and correlation coefficient (ρhv) by measuring the differences in backscattering characteristics between horizontal and vertical polarized electromagnetic waves, enabling effective identification of precipitation particle phase, shape, and size spectrum distribution6,7. Vertical structure observations show that microphysical characteristic parameters such as zero-degree layer bright band features, ice crystal zone thickness, and graupel particle layer distribution are closely related to surface precipitation intensity8. This polarimetric parameter information has greatly enhanced our understanding of precipitation microphysical processes. In recent years, machine learning and artificial intelligence techniques have made remarkable advances in the field of precipitation nowcasting. Deep learning-based methods have demonstrated great potential in tropical cyclone precipitation forecasting, with generative adversarial networks achieving high-accuracy short-term forecasts for 0–120 min9. For typhoon disaster monitoring, deep learning frameworks can now provide 6-h lead time forecasts, offering extended warning periods for alert systems10. Physics-constrained graph neural networks, by coupling atmospheric dynamic processes, have shown excellent performance in extreme precipitation event forecasting and can provide uncertainty quantification11. Physical-driven diffusion networks that combine radar observations with numerical weather predictions offer probabilistic forecasting capabilities for operational applications12. Furthermore, generative ensemble deep learning methods have achieved a 20% improvement in skill scores for severe weather prediction13. These studies provide an important technological development background for the dual-polarization radar microphysical feature-based warning method proposed in this paper, and also indicate the future research direction of integrating machine learning with physical processes. The simulation and observation of cloud microphysical processes have always been important challenges in atmospheric science research14. In typhoon precipitation research, evaluation of microphysical schemes in outer rainbands has shown that different schemes have significant impacts on precipitation forecasting15. Studies on the relationship between radar bright band characteristics and surface precipitation rates have provided new approaches for quantitative precipitation estimation16. Experimental studies on ice particle collision and breakup processes have revealed the microphysical mechanisms of graupel-graupel and graupel-snowflake collisions17. Research on the vertical distribution characteristics of typhoon gust factors based on tower observations18 and lidar network observations of the vertical distribution of atmospheric pollution during typhoons19 has provided important information for understanding typhoon boundary layer structure. Observations of the vertical variation of raindrop size distribution from the 356-m meteorological tower in Shenzhen have further deepened our understanding of typhoon precipitation microphysical processes20. Despite the numerous advances in typhoon precipitation research using dual-polarization radar, there are still deficiencies in short-term heavy precipitation warning applications. Studies on hydrometeor classification and mosaic algorithms for X-band polarimetric radar21 and clustering-based hydrometeor classification for quasi-vertical profiles22 have provided new methods for particle identification, but the application of these techniques in typhoon short-term heavy precipitation warning still needs further exploration. In-depth understanding of typhoon precipitation microphysical processes and the development of corresponding warning techniques are of great significance for enhancing disaster prevention and mitigation capabilities. This study aims to establish a typhoon short-term heavy precipitation warning method based on dual-polarization radar observations. The research innovatively combines the vertical distribution characteristics of precipitation particles in inner and outer typhoon spiral rainbands with short-term heavy precipitation warning, constructing a warning index system with clear physical significance. Through systematic analysis of the vertical structure evolution patterns of precipitation particles under different dynamic and thermodynamic backgrounds, key microphysical parameters such as ice-phase particle content, zero-degree layer bright band intensity, and particle growth rate are extracted. Quantitative relationships between these parameters and surface precipitation intensity are established to develop warning methods based on microphysical processes. This study selects typical typhoon cases that have affected the southeastern coast of China in recent years, utilizing S-band dual-polarization radar observation data and employing fuzzy logic algorithms for precipitation particle identification to comparatively analyze the differences in particle vertical distribution characteristics between inner and outer typhoon spiral rainbands. By constructing statistical relationships between particle characteristic parameters and short-term heavy precipitation, warning index thresholds are established. Independent samples are used to test and evaluate the warning method, analyzing its applicability under different typhoon intensities and development stages. The research outcomes are expected to provide new technical approaches for typhoon short-term heavy precipitation warning and offer scientific support for improving the typhoon precipitation forecasting and warning system. Methods Data This study uses observation data from the S-band dual-polarization weather radar network along the southeastern coast of China, including polarimetric parameters such as horizontal reflectivity factor (Zh), differential reflectivity (Zdr), correlation coefficient (ρhv), differential phase (φdp), and specific differential phase (Kdp). The radar volume scan cycle is 6 min with a radial resolution of 250 m, and the data have undergone quality control processing to ensure reliability23. The data preprocessing procedure involves several critical steps to ensure data quality and reliability. First, ground clutter suppression was performed using a fuzzy logic-based algorithm that combines reflectivity texture, vertical gradient, and correlation coefficient to identify and remove non-meteorological echoes. Specifically, pixels with ρhv < 0.80, reflectivity texture > 10 dBZ within a 3 × 3 km window, and negative vertical gradients below 2 km altitude were flagged as ground clutter and removed. Second, biological echoes (birds, insects) were filtered using dual-polarization signatures, where pixels with Zh < 35 dBZ, Zdr > 4 dB, and ρhv < 0.95 were classified as biological targets and excluded from analysis. Third, attenuation correction was applied using the self-consistent method with constraints, which utilizes the differential phase shift to estimate and correct for path-integrated attenuation in both Zh and Zdr. The correction coefficients were α = 0.08 dB/deg for Zh and β = 0.02 dB/deg for Zdr, following the relationships established for S-band radar. Fourth, beam blockage correction was implemented using a high-resolution digital elevation model to identify partially blocked beams and apply correction factors ranging from 1.0 to 3.0 dB based on the blockage percentage. Fifth, noise filtering was conducted by removing pixels with signal-to-noise ratio < 5 dB and applying a 5 × 5 median filter to remove isolated noise spikes. Finally, data consistency checks were performed, including range-dependent threshold tests, temporal continuity analysis, and spatial correlation verification to ensure the physical consistency of the polarimetric variables. To address the degradation of data quality at the edges of radar detection range, special processing procedures were implemented. For ranges between 100–150 km, a distance-dependent weighting function was applied to account for beam broadening effects, with weights decreasing linearly from 1.0 at 100 km to 0.5 at 150 km. Beyond 150 km, where beam height exceeds 3 km and partial beam filling becomes significant, data were excluded from the particle classification analysis but retained for qualitative rainband structure identification. Additionally, for low elevation angles (< 0.5°) where ground clutter contamination is more severe, an enhanced clutter filter with stricter thresholds (ρhv < 0.70, texture > 8 dBZ) was applied within the first 50 km range. Surface precipitation data are obtained from hourly observations of approximately 500 automatic weather stations within the radar’s effective detection range (150 km). According to China Meteorological Administration standards, hourly precipitation ≥ 50 mm is defined as a short-term heavy precipitation event, serving as the ground truth for warning effectiveness verification. Typhoon track and intensity data are sourced from the China Meteorological Administration’s tropical cyclone best track dataset, with hourly typhoon center positions and intensity information obtained through interpolation24. Radiosonde data are used to determine the heights of key temperature levels such as 0 °C, − 10 °C, and − 20 °C, providing environmental temperature constraints for particle phase identification. The study selected 6 typhoon cases that affected the southeastern coast during 2019–2023: “Lekima” (2019), “Hagupit” (2020), “In-Fa” (2021), “Chanthu” (2021), “Muifa” (2022), and “Doksuri” (2023). As shown in Table 1, these typhoons cover different intensity categories (from typhoon to super typhoon) and track types (westward landing, northward recurving, and coastal paralleling), with 256 samples used for training and 156 for validation. These typhoons were specifically selected to ensure representativeness: they cover different intensity categories (2 severe typhoons, 2 typhoons, 2 tropical storms), various track types (westward landing, northward recurving, and coastal paralleling), and diverse seasonal patterns (peak season and off-season events). Together, these 6 typhoons generated 35 heavy precipitation processes and 256 hourly samples exceeding 50 mm, providing statistically robust training data. All these typhoons produced significant short-term heavy precipitation processes with complete dual-polarization radar observation data, which was only systematically available in our study region starting from 2019. Table 1Sample distribution of selected typhoon cases.TyphoonYearMax intensityTrack typeSamplesUsageLekima2019Super typhoon (52 m/s)Westward landing78TrainingHagupit2020Typhoon (38 m/s)Northward recurving52TrainingIn-Fa2021Severe typhoon (45 m/s)Coastal paralleling68TrainingChanthu2021Super typhoon (55 m/s)Northward recurving58TrainingMuifa2022Severe typhoon (42 m/s)Westward landing92ValidationHaikui2023Typhoon (40 m/s)Westward landing64ValidationTotal412256/156The selection ensures representation across: (1) Intensity categories—from typhoon to super typhoon; (2) Track types—westward landing, northward recurving, and coastal paralleling; (3) Seasonal distribution—covering both peak season (July–September) and late season (October) events. Sample distribution of selected typhoon cases. The selection ensures representation across: (1) Intensity categories—from typhoon to super typhoon; (2) Track types—westward landing, northward recurving, and coastal paralleling; (3) Seasonal distribution—covering both peak season (July–September) and late season (October) events. To consider topographic effects, observation stations were classified into three categories based on elevation: plain stations (< 50 m), hilly stations (50–200 m), and mountainous stations (> 200 m). Additionally, classification statistics were conducted according to four typhoon intensity levels: tropical storm (< 32.7 m/s), typhoon (32.7–41.4 m/s), severe typhoon (41.5–50.9 m/s), and super typhoon (≥ 51.0 m/s). Typhoon rainband classification criteria Accurate classification of typhoon spiral rainbands is fundamental for conducting comparative studies between inner and outer rainbands. This study divides typhoon precipitation regions into inner spiral rainbands and outer spiral rainbands based on their distance from the typhoon center and structural characteristics. Based on research by Willoughby et al.25 and statistical analysis of typhoon structures in this region, it was found that typhoon precipitation structure exhibits a distinct transition at a distance of 100 km from the center: within 100 km, the average vertical velocity > 5 m/s with convective proportion > 60%; beyond 100 km, vertical velocity < 3 m/s with stratiform precipitation proportion > 70%. Surface precipitation data are obtained from hourly observations of approximately 500 automatic weather stations within the radar’s effective detection range (150 km). According to China Meteorological Administration standards, hourly precipitation ≥ 50 mm is defined as a short-term heavy precipitation event, serving as the ground truth for warning effectiveness verification26. To further distinguish between convective and stratiform precipitation, the texture algorithm proposed by Steiner et al.27 is employed for identification. This algorithm comprehensively considers reflectivity intensity and spatial variation characteristics, classifying regions with reflectivity > 40 dBZ or reflectivity gradient > 4.5 dBZ/km as convective areas, and the remaining regions as stratiform areas. This classification method can effectively identify mesoscale convective systems within typhoon rainbands, providing a foundation for analyzing differences in microphysical characteristics of different precipitation types. Based on typhoon disaster warning research28, the inner spiral rainband is defined as the precipitation region within 0–100 km from the typhoon center. This region typically includes the eyewall and its adjacent inner-core precipitation bands, characterized by strong upward motion and deep convective development. The 100 km radius boundary represents a statistical average derived from analysis of the six typhoon cases, which ranged from compact (radius of maximum wind 30–40 km) to large systems (radius of maximum wind 60–80 km). For individual typhoons, this boundary varied between 80–120 km depending on storm size and intensity. Sensitivity analysis showed that boundary variations of ± 20 km resulted in less than 5% change in classification accuracy, confirming the robustness of the 100 km criterion for operational application. The outer spiral rainband is defined as the precipitation region within 100–300 km from the typhoon center, mainly composed of spiral cloud bands in the typhoon periphery with relatively weaker convective intensity. Observations of the thermodynamic and dynamic structure of rainbands during Typhoon “Lekima” (2019) landfall confirmed the rationality of this classification29. As shown in Fig. 1, the study area covers the southeastern coastal region of China, illustrating possible typhoon tracks and radar observation coverage. Red triangles mark the locations of dual-polarization radar stations, and dashed circles indicate the boundary between inner and outer rainbands. From the radar reflectivity distribution, it can be seen that the inner rainband exhibits compact spiral structure and high reflectivity characteristics (> 45 dBZ), while the outer rainband presents a relatively loose band-like distribution with relatively lower reflectivity intensity (30–40 dBZ). Study area and typhoon rain band classification. Precipitation particle identification algorithm Precipitation particle identification employs a fuzzy logic-based classification algorithm that comprehensively utilizes multi-parameter information from dual-polarization radar and environmental temperature conditions to achieve effective identification of precipitation particles in different phases. The algorithm first establishes membership functions for various types of precipitation particles, describing the fuzzy relationships between dual-polarization parameters and particle types. The membership functions adopt a trapezoidal function form, with parameter thresholds determined based on observational studies of cloud seeding-feeder processes30 and localized statistical analysis. As shown in Table 2, this study classifies precipitation particles into four categories: raindrops, ice crystals, graupel, and wet snow. Raindrops are mainly distributed below the 0 °C level, exhibiting relatively high differential reflectivity (0.5–3 dB) and correlation coefficient (> 0.97), reflecting the oblate characteristics of large raindrops. Ice crystals typically exist at high altitudes below − 10 °C, characterized by low reflectivity (10–30 dBZ) and correlation coefficient close to 1. Graupel particles are important products of ice-phase processes, with differential reflectivity close to 0 dB and relatively low correlation coefficient (0.85–0.95), indicating nearly spherical particle shapes. The fall velocity of graupel has significant impacts on typhoon precipitation31. Wet snow appears near the melting layer, with parameters intermediate between ice-phase and liquid-phase particles. Table 2Dual-polarization radar parameter thresholds and precipitation particle classification criteria.Particle typeZh (dBZ)Zdr (dB)ρhvKdp (°/km)Temperature (°C)Raindrop20–600.5–3.0 > 0.970–5 > 0Ice crystal10–300–1.0 > 0.98 < 0.2 < − 10Graupel30–55− 0.5–1.00.85–0.950–2− 10–0Wet snow25–450–2.00.88–0.950.5–3− 5–0Zh is horizontal reflectivity factor; Zdr is differential reflectivity; ρhv is correlation coefficient; Kdp is specific differential phase. For overlapping temperature ranges, classification priority follows: (1) maximum combined membership value across all polarimetric parameters; (2) altitude constraints (e.g., no raindrops above − 10 °C level); (3) polarimetric signature strength, where stronger Zdr indicates raindrops over graupel at the melting layer. Dual-polarization radar parameter thresholds and precipitation particle classification criteria. Zh is horizontal reflectivity factor; Zdr is differential reflectivity; ρhv is correlation coefficient; Kdp is specific differential phase. For overlapping temperature ranges, classification priority follows: (1) maximum combined membership value across all polarimetric parameters; (2) altitude constraints (e.g., no raindrops above − 10 °C level); (3) polarimetric signature strength, where stronger Zdr indicates raindrops over graupel at the melting layer. During the particle identification process, the membership values for each particle type are calculated for every radar sampling point, and the category with the maximum membership value is selected as the particle type for that point. To improve identification accuracy, the algorithm incorporates temperature constraints, using temperature profile information obtained from radiosonde data to exclude unreasonable particle types. After identification is completed, particle distribution profiles are extracted along radial and vertical directions, and statistics of particle type proportions and characteristic parameters at different height levels are compiled, providing fundamental data for subsequent comparative analysis and warning index construction. The complete fuzzy logic membership functions with specific parameter values for each particle type are provided in the Supplementary Materials, along with Python implementation code for reproducibility. Short-term heavy precipitation definition and warning index construction According to China Meteorological Administration operational regulations and typhoon precipitation characteristics, this study defines short-term heavy precipitation as precipitation events with hourly rainfall ≥ 50 mm or 3-h rainfall ≥ 100 mm. To improve the refinement level of warnings, short-term heavy precipitation is further classified into three levels: general heavy precipitation (50–80 mm/h), heavy precipitation (80–100 mm/h), and extremely heavy precipitation (> 100 mm/h). This classification helps establish differentiated warning indices for different precipitation intensities. Based on the vertical distribution characteristics of precipitation particles observed by dual-polarization radar, the following key parameters are selected as warning index candidates: (1) graupel particle volume content at 5 km height, reflecting the intensity of ice-phase processes; (2) differential reflectivity peak value of the bright band at 0 °C level, indicating the activity level of melting processes; (3) ice crystal layer thickness, defined as the vertical range where ice crystals exist between − 10 °C and − 40 °C temperature levels; (4) proportion of large raindrops (Zdr > 2 dB) in the lower levels (below 2 km), characterizing the contribution of warm rain processes; (5) vertical reflectivity gradient, i.e., the rate of reflectivity change from surface to 6 km height; (6) vertical diversity index of particle types, reflecting the complexity of microphysical processes. Prior to multivariate regression analysis, all six parameters were standardized using z-score normalization to ensure comparability across different physical units and ranges. The warning index system is constructed using a combination of statistical methods and physical constraints. Through statistical analysis of 256 historical short-term heavy precipitation cases collected from 6 typhoons (2019–2021), with 4 typhoons (Lekima 2019, Hagupit 2020, In-Fa 2021, Chanthu 2021) used for model training, correlation coefficients between each parameter and surface precipitation intensity are calculated. The training dataset comprises: (1) By rainband type: 148 samples from inner rainbands and 108 samples from outer rainbands; (2) By precipitation intensity: 142 samples of 50–80 mm/h, 76 samples of 80–100 mm/h, and 38 samples of > 100 mm/h; (3) By topographic category: 98 samples from plain stations, 88 from hilly stations, and 70 from mountainous stations. This stratified distribution ensures adequate representation of all precipitation categories in the regression analysis. Probability density distributions are used to determine parameter thresholds for different precipitation levels, and percentile methods are employed to set warning critical values. A multi-parameter comprehensive warning index PI (Precipitation Index) is constructed: where: is the graupel particle volume content (%) at 5 km height; is the differential reflectivity peak value (dB) at the 0 °C level bright band; is the ice crystal layer thickness (km) between − 10 °C and − 40 °C levels; is the proportion (%) of large raindrops ( dB) below 2 km; is the vertical reflectivity gradient (dBZ/km) from surface to 6 km; is the vertical diversity index of particle types (dimensionless). The weight coefficients ( through ) are determined through multivariate regression analysis using 256 historical short-term heavy precipitation samples, with values: ,,,,,. These weights reflect the relative importance of each parameter in predicting surface precipitation intensity, with graupel content and bright band intensity being the most influential factors. Prior to regression analysis, multicollinearity among the six predictor variables was assessed using variance inflation factors (VIFs). The VIF values were:(VIF = 2.3), (VIF = 1.8),(VIF = 2.1),(VIF = 1.5),(VIF = 1.9), and (VIF = 1.6). All VIF values were below the threshold of 5, indicating no serious multicollinearity issues. The highest correlation between predictors was between and , which is expected given their physical relationship but does not compromise the regression model stability. When PI exceeds the set threshold, short-term heavy precipitation warnings at corresponding levels are issued. The warning lead time is set to 15–45 min, adjusted according to storm movement speed and rainband development characteristics. The 75th, 85th, and 95th percentiles of the warning threshold are used as critical values for yellow, orange, and red warnings respectively. Through cross-validation to ensure threshold stability, data from at least 30 cases are used for supplementary updates at neighboring stations. Results Basic characteristics of vertical distribution of precipitation particles in inner and outer typhoon rainbands Statistical analysis of 6 typhoon cases shows that precipitation particles in typhoon inner rainbands exhibit distinct vertical stratification structure. Below the 0 °C level (approximately 5 km height), raindrops dominate with volume fraction exceeding 85%, and average differential reflectivity of 1.8 ± 0.5 dB, indicating the presence of numerous large raindrops with diameters of 2–4 mm. Near the 0 °C level, significant bright band features appear with differential reflectivity peaks reaching 3.5 dB and correlation coefficient dropping to 0.85–0.90, reflecting the transformation process from ice crystals and snowflakes to raindrops within the melting layer. In the − 10 °C to − 20 °C layer (6–8 km), graupel particle proportion increases dramatically, averaging 60%, with reflectivity factor maintaining high values of 40–50 dBZ, indicating active riming growth processes. Above the − 20 °C level, ice crystals dominate with proportions exceeding 80%, and ice crystal layer tops can reach 12 km, demonstrating deep convective development. The particle vertical distribution characteristics of outer rainbands differ significantly from inner rainbands. Below the 0 °C level in outer rainbands, raindrops also dominate, but the average differential reflectivity is only 1.2 ± 0.3 dB, indicating relatively smaller raindrop sizes. The melting layer bright band features are weaker, with differential reflectivity peaks generally not exceeding 2.5 dB. Graupel particles are mainly distributed in the − 5 °C to − 15 °C layer (5.5–7 km) with proportions of about 40%, significantly lower than in inner rainbands. Ice crystal layer top heights are typically at 9–10 km, 2–3 km lower than inner rainbands. The vertical reflectivity gradient in outer rainbands is smaller, with reflectivity decreasing by about 15 dBZ from surface to 6 km height, while inner rainbands show decreases of up to 25 dBZ. As shown in Fig. 2, the main differences in particle vertical distribution between inner and outer rainbands are reflected in: (1) significant differences in ice-phase particle content, with graupel particle volume content at 5 km height in inner rainbands being 1.5–2 times that of outer rainbands; (2) different vertical development depths, with particle distribution in inner rainbands reaching the tropopause while outer rainbands are mainly limited to below 10 km; (3) obvious differences in microphysical process activity, with bright band intensity at the 0 °C level in inner rainbands 40% higher than outer rainbands, indicating more intense ice-phase particle melting processes; (4) different particle size distribution characteristics, with particle sizes at various height levels in inner rainbands generally larger than outer rainbands, particularly most significant near the melting layer. These differential characteristics provide a physical foundation for short-term heavy precipitation warning based on particle distribution. Comparison of vertical distribution of precipitation particles in typhoon inner and outer rainbands. (a) inner rainband; (b) outer rainband; (c) graupel content vertical profile; (d) differential reflectivity vertical profile. Topographic factor analysis shows that graupel particle content at mountainous stations is on average 25% ± 7% higher than at plain stations, primarily due to enhanced vertical motion caused by topographic lifting. Particle distribution also varies among typhoons of different intensities: severe typhoons and super typhoons have ice crystal layer top heights 2–3 km higher than tropical storms, with graupel particle content 15–20% higher. Vertical distribution characteristics of particles in short-term heavy precipitation of different intensities Through classified statistical analysis of 256 short-term heavy precipitation samples during typhoon influence periods, the vertical distribution characteristics of particles corresponding to different precipitation intensities were analyzed. Results show that the vertical structure of precipitation particles has a significant correspondence with surface precipitation intensity, with inner and outer rainbands exhibiting different characteristics. For general heavy precipitation of 50–80 mm/h, particle vertical distribution in inner rainband samples shows: average graupel particle volume content of 35% ± 8% at 5 km height (approximately 0 °C level), with maximum values occurring at 5.5–6.5 km height layer; differential reflectivity peak of the 0 °C level bright band is 2.8 ± 0.4 dB, indicating moderate ice-phase particle melting processes; average ice crystal layer top height is 10.5 km, with ice crystal proportion reaching 75% above the − 20 °C level; large raindrop (Zdr > 2 dB) proportion in the lower levels (below 2 km) is 25%, indicating that warm rain processes contribute but are not dominant. In 50–80 mm/h precipitation samples from outer rainbands, graupel particle content is significantly lower, averaging only 20% ± 6% at 5 km height, ice crystal layer top height is about 9 km, and lower-level large raindrop proportion reaches 40%, showing more important contribution from warm rain processes. When precipitation intensity exceeds 80 mm/h, particle vertical distribution characteristics change significantly. Inner rainband extremely heavy precipitation samples show: graupel particle volume content at 5 km height increases dramatically to 55% ± 10%, with some cases exceeding 70%; graupel particle layer thickness increases from 2–3 km in general heavy precipitation to 4–5 km, indicating exceptionally active ice-phase processes; 0 °C level bright band differential reflectivity peak reaches 3.8 ± 0.5 dB, with individual times exceeding 4.5 dB, reflecting intense melting processes; ice crystal layer top breaks through 12 km, approaching the tropopause; vertical reflectivity gradient from surface to 6 km height reaches 30 dBZ, indicating rapid particle growth. Extremely heavy precipitation in outer rainbands is relatively rare, but when it occurs, it also shows characteristics of increased graupel particle content (reaching 35%) and enhanced vertical development. As shown in Fig. 3, particle characteristic parameters exhibit a nonlinear relationship with precipitation intensity. Correlation analysis shows that the correlation coefficient between graupel particle content at 5 km height and surface precipitation intensity reaches 0.78 (p < 0.001) in inner rainbands and 0.65 (p < 0.001) in outer rainbands; correlation coefficients for 0 °C level bright band intensity and precipitation intensity are 0.72 and 0.58 respectively; correlation coefficients for ice crystal layer thickness are 0.68 (p < 0.001) and 0.52 (p < 0.01). Notably, when graupel particle content exceeds 45%, precipitation intensity increases most rapidly, indicating a critical threshold effect. The physical mechanism of this critical effect lies in: when graupel particle content reaches 45%, it indicates sufficient supercooled water supply above the 0 °C level and suitable temperature conditions (− 5 to − 15 °C), with riming growth efficiency reaching its maximum. At this time, the mass growth rate of individual graupel particles can reach 10⁻4 g/s, leading to exponential growth of total precipitation particle mass through chain reactions, ultimately forming heavy precipitation after melting. Multiple regression analysis shows that comprehensively considering three parameters—graupel particle content, bright band intensity, and vertical development depth—can explain 65% of the variance in precipitation intensity changes, providing a quantitative basis for establishing warning indices. Relationship between particle characteristic parameters and precipitation intensity. (a) graupel content vs. precipitation intensity; (b) bright band intensity vs. precipitation; (c) ice layer thickness vs. precipitation; (d) threshold effect of graupel content. Particle characteristic-based short-term heavy precipitation warning indices Based on the aforementioned analysis results, a multi-parameter warning index system for typhoon short-term heavy precipitation was constructed. Through probability density analysis and percentile methods, threshold standards for different warning levels were determined. Weight coefficients were optimized using stepwise regression methods. Regression analysis shows that the model’s coefficient of determination R2 = 0.73 (adjusted R2 = 0.72) (p < 0.001), with standardized regression coefficients of all parameters passing significance tests (p < 0.05). Among them, graupel particle content has the largest contribution (standardized coefficient β = 0.42), followed by zero-degree level bright band intensity (β = 0.28). The calculation formula for the comprehensive warning index PI was finally established: where all parameters have been standardized. For inner rainbands, the established three-level warning thresholds are: when graupel particle content at 5 km height exceeds 30%, 0 °C level bright band Zdr peak exceeds 2.5 dB, and ice crystal layer thickness exceeds 4 km, a yellow warning (50–80 mm/h) is issued; when graupel particle content exceeds 45%, bright band Zdr peak exceeds 3.2 dB, and ice crystal layer thickness exceeds 5 km, an orange warning (80–100 mm/h) is issued; when graupel particle content exceeds 60%, bright band Zdr peak exceeds 3.8 dB, and ice crystal layer thickness exceeds 6 km, a red warning (> 100 mm/h) is issued. The thresholds for outer rainbands are correspondingly reduced, with the yellow warning graupel particle content threshold at 20%, and orange and red warnings at 30% and 45% respectively. As shown in Fig. 4, taking an extremely heavy precipitation event during the influence of Typhoon “Doksuri” in 2023 as an example, the temporal evolution characteristics of warning indices are demonstrated. 45 min before heavy precipitation occurred at the surface, graupel particle content at 5 km height began to increase rapidly, rising from 15 to 55%; 30 min before, the 0 °C level bright band intensity showed a sharp increase, with Zdr peak rising from 2.0 dB to 3.5 dB; 20 min before, ice crystal layer thickness exceeded 5 km. The comprehensive warning index PI exceeded the yellow warning threshold 35 min before heavy precipitation occurred, reached the orange warning standard 25 min before, and triggered the red warning 15 min before. The actual surface precipitation rate peak reached 112 mm/h, consistent with the warning level. Temporal evolution of warning indicators during typhoon “Doksuri” (2023). (a) evolution of microphysical parameters; (b) precipitation rate and warning index evolution. Warning lead time analysis shows that the particle characteristic-based warning method can identify short-term heavy precipitation with an average lead time of 25–40 min, an improvement of 15–25 min compared to traditional reflectivity extrapolation-based methods. Among these, graupel particle content changes provide the longest warning lead time (average 35 min), followed by ice crystal layer thickness (30 min) and bright band intensity (25 min). Warning lead time is negatively correlated with typhoon movement speed, with fast-moving typhoons (> 30 km/h) reducing lead time to 20–30 min, while slow-moving or quasi-stationary typhoons can reach 40–50 min. Warning lead time for inner rainbands is generally 5–10 min longer than outer rainbands, which is related to stronger vertical motion and more active microphysical processes in inner rainbands. Warning effectiveness verification Systematic verification of the constructed warning method was conducted using multiple validation approaches to ensure robustness. First, we performed cross-year blind testing using independent samples from the 2022–2023 typhoon season. The verification samples included two typhoon processes, “Muifa” (2022) and “Haikui” (2023), yielding a total of 156 short-term heavy precipitation samples. The validation dataset distribution is: (1) By rainband type: 92 samples from inner rainbands (59%) and 64 from outer rainbands (41%); (2) By precipitation intensity: 89 samples of 50–80 mm/h (57%), 45 samples of 80–100 mm/h (29%), and 22 samples of > 100 mm/h (14%); (3) By topographic category: 62 plain stations (40%), 54 hilly stations (35%), and 40 mountainous stations (25%). This distribution maintains similar proportions to the training dataset, ensuring unbiased validation. These two typhoons were completely excluded from the model training process to ensure independence. To further validate the method’s generalizability, we conducted additional verification using data from three independent S-band dual-polarization radars located in Wenzhou, Shantou, and Zhanjiang, which were not included in the original development dataset. This independent radar dataset covered four additional typhoon events from 2023–2024, providing 98 supplementary validation samples. The cross-radar validation results showed consistent performance: POD = 82.7% (± 2.1%), FAR = 24.3% (± 1.8%), and CSI = 65.2% (± 2.6%), confirming the method’s applicability across different radar systems. Additionally, we implemented a k-fold cross-validation approach (k = 5) using the entire dataset, where each fold contained typhoons from different years to ensure temporal independence. The average cross-validation metrics (POD = 83.9% ± 1.5%, FAR = 23.1% ± 1.2%) closely matched the independent test results, demonstrating the stability of the method. Indicators including Probability of Detection (POD), False Alarm Ratio (FAR), Critical Success Index (CSI), and average warning lead time were used to comprehensively evaluate the warning effectiveness. As shown in Table 3, the particle characteristic-based warning method significantly outperforms the traditional reflectivity threshold method. For short-term heavy precipitation ≥ 50 mm/h, the overall POD of the new method reaches 84.6%, an improvement of 19.2 percentage points over the traditional method; FAR decreases to 22.5%, a reduction of 15.7 percentage points; CSI increases from 45.6% to 67.8%. By rainband type, the warning effectiveness for inner rainbands is superior to outer rainbands, with POD of 87.0% and 81.3% respectively, mainly due to more distinct particle characteristic signals in inner rainbands. By precipitation level, extremely heavy precipitation (> 100 mm/h) has the highest POD, reaching 92.3%, indicating that extreme precipitation events have more significant particle characteristics. Table 3Statistical results of warning effectiveness verification.Warning methodSample SizePOD (%)FAR (%)CSI (%)Average warning lead time (min)Overall Traditional Z threshold method15665.438.245.610–15 This method15684.622.567.825–40 Improvement– + 19.2− 15.7 + 22.2 + 15–25Inner rainband Traditional Z threshold method9268.535.748.912–18 This method9287.020.171.230–45 Improvement– + 18.5− 15.6 + 22.3 + 18–27Outer rainband Traditional Z threshold method6460.942.140.68–12 This method6481.326.262.820–35 Improvement– + 20.4− 15.9 + 22.2 + 12–23By precipitation intensity 50–80 mm/h Traditional Z threshold method8961.841.542.18–12 This method8980.925.863.523–33 80–100 mm/h Traditional Z threshold method4568.933.350.012–15 This method4586.720.071.130–40 > 100 mm/h Traditional Z threshold method2272.731.853.315–20 This method2292.315.478.935–50Sample sizes shown represent independent validation samples from typhoons “Muifa” (2022) and “Haikui” (2023), completely separate from the 256 training samples used in model development. POD = Hits/(Hits + Misses); FAR = False alarms/(Hits + False alarms); CSI = Hits/(Hits + Misses + False alarms). Validation typhoons covered different intensities (Muifa: severe typhoon, 42 m/s; Haikui: typhoon, 40 m/s). Statistical results of warning effectiveness verification. Sample sizes shown represent independent validation samples from typhoons “Muifa” (2022) and “Haikui” (2023), completely separate from the 256 training samples used in model development. POD = Hits/(Hits + Misses); FAR = False alarms/(Hits + False alarms); CSI = Hits/(Hits + Misses + False alarms). Validation typhoons covered different intensities (Muifa: severe typhoon, 42 m/s; Haikui: typhoon, 40 m/s). Uncertainty analysis of warning results shows that the 95% confidence interval for POD is [81.2%, 87.9%], and the confidence interval for FAR is [18.6%, 26.4%]. Bootstrap resampling analysis (n = 1000) shows that the standard error of warning lead time is ± 2.3 min. Parameter sensitivity analysis indicates that a ± 5% change in graupel particle content threshold results in a ± 3.2% change in POD, making it the most sensitive parameter. Regarding warning lead time, the new method provides an average warning lead time of 25–40 min, an increase of 15–25 min compared to the traditional method. Specifically, the average lead time for 50–80 mm/h precipitation is 28 min, 35 min for 80–100 mm/h, and up to 42 min for > 100 mm/h. Notably, the stability of warning lead time has also improved, with standard deviation reduced from ± 12 min for the traditional method to ± 8 min, indicating higher reliability of the new method. Spatial distribution verification shows that warning effectiveness is optimal within 50–100 km from the radar, with CSI reaching 72.5%; it slightly decreases but remains above 65% in the 100–150 km range; beyond 150 km, warning effectiveness significantly declines due to radar detection capability limitations. Typical case analysis further validates the method’s effectiveness. On July 28, 2023, during Typhoon “Doksuri’s’ landfall in Fujian, Jinjiang station recorded extreme precipitation of 126 mm/h. The traditional method identified strong echoes only 12 min in advance, while the new method detected a sharp increase in graupel particle content to 58% at 5 km height at 0645 UTC (38 min before precipitation peak), bright band Zdr peak exceeded 3.5 dB triggering an orange warning at 0650 UTC, and the comprehensive index reached red warning criteria at 0655 UTC. Comparative analysis shows that the evolution of particle characteristics clearly reflects the development of the precipitation system, particularly the enhancement of ice-phase processes, providing clear precursor signals for heavy precipitation occurrence. Analysis of missed cases indicates that the few missed warnings mainly occurred in rapidly moving small-scale convective systems and radar long-range detection areas, pointing directions for further improvement of the warning method. Analysis of typical missed warning cases reveals three main causes: (1) Rapidly moving small-scale convective systems (horizontal scale < 20 km), such as the missed case at Fuzhou station on August 5, 2023, where the system moved at 45 km/h and produced heavy precipitation within only 15 min of appearance; (2) Low-centroid warm cloud precipitation processes, such as the case at Xiamen station on September 12, 2022, which produced 70 mm/h precipitation below the 0 °C level with insignificant graupel particle signals; (3) Radar long-range attenuation, such as at Quanzhou station > 130 km from the radar, where dual-polarization parameter measurement errors led to particle identification biases. For these situations, it is recommended to combine radar extrapolation, satellite data, and numerical forecasts for comprehensive judgment. Discussion Physical mechanisms of particle distribution differences between inner and outer rainbands The significant differences in vertical distribution of precipitation particles between typhoon inner and outer rainbands stem from their different dynamic and thermodynamic conditions. The inner rainband is located near the typhoon eyewall, controlled by strong vertical upward motion, with maximum updraft velocities reaching 15–20 m/s, and even exceeding 25 m/s in some strong convective regions. This strong updraft can rapidly transport large amounts of liquid water above the 0 °C level, providing sufficient supercooled water for ice-phase particle formation and extending particle residence time in the growth zone. In contrast, vertical motion in outer rainbands is relatively weak, with average updraft velocities of only 5–8 m/s, resulting in relatively inactive ice-phase processes. Research on the generation and enhancement mechanisms of typhoon extreme topographic precipitation32 indicates that topographic lifting can significantly enhance vertical motion, leading to more active ice-phase processes on windward slopes. Differences in water vapor transport and microphysical processes further reinforce the distinction between inner and outer rainbands. Inner rainbands directly benefit from strong low-level convergence in typhoons, with water vapor flux 50–80% higher than outer rainbands, and average relative humidity from 0–6 km exceeding 95%. Sufficient water vapor supply supports stronger condensation latent heat release, enhancing vertical motion through positive feedback mechanisms. Studies on the effects of topography and sea surface temperature anomalies on typhoon precipitation33 show that warm SST anomalies can enhance low-level water vapor transport, leading to stronger precipitation. Active ice-phase processes in inner rainbands form large amounts of graupel particles through the Bergeron process and riming growth, while outer rainbands are dominated by warm rain processes with relatively lower particle growth efficiency. This difference determines the significant disparity in precipitation intensity between the two. Indication mechanism of particle vertical distribution on precipitation intensity The close relationship between vertical distribution characteristics of precipitation particles and surface precipitation intensity reflects the controlling role of cloud microphysical processes in precipitation formation. Ice-phase processes, particularly the formation and growth of graupel particles, play a crucial role in typhoon heavy precipitation. When graupel particle content at 5 km height exceeds 45%, it indicates active riming growth processes, with graupel particles contributing 60–70% to surface precipitation through melting. Research on the impact of long-range water vapor transport on extreme precipitation events34 further confirms that sufficient water vapor supply is a necessary condition for maintaining active ice-phase processes. Heavy precipitation in inner rainbands mainly relies on the “seeder-feeder” mechanism, where high-level ice crystals and graupel particles fall into the warm sector and rapidly convert to precipitation through melting and coalescence processes, with efficiency far exceeding that of pure warm rain processes. As shown in Fig. 5, significant differences exist in the formation mechanisms of short-term heavy precipitation between typhoon inner and outer rainbands. In inner rainbands, deep convection and strong updrafts transport large amounts of water vapor to high altitudes, forming abundant ice-phase particles in the ice crystal layer, which form graupel particles through riming, aggregation, and other processes, ultimately melting to form heavy precipitation. Outer rainbands have relatively shallow convection, dominated by warm rain processes with limited ice-phase process contributions. Characteristic parameters such as 0 °C level bright band intensity, vertical reflectivity gradient, and ice crystal layer thickness collectively indicate different stages of precipitation development, providing a physical foundation for short-term heavy precipitation warning. Research on convolutional neural network-based precipitation estimation methods using radar mosaics35 has also confirmed the importance of these microphysical parameters for precipitation intensity estimation. Formation mechanisms of heavy precipitation in typhoon inner and outer Rainbands. (a) inner rainband; (b) outer rainband. Applicability and limitations of the warning method The particle characteristic-based warning method demonstrates good applicability across different types of typhoons, but requires adjustments according to specific circumstances. For typical Northwest Pacific typhoons, warning effectiveness is optimal, with CSI generally exceeding 65%. For fast-moving typhoons (> 40 km/h), warning lead time shortens to less than 20 min; for structurally asymmetric or transitioning typhoons, particle distribution characteristics may deviate from typical patterns, requiring threshold adjustments. Topographic effects are significant, with graupel particle content on windward slopes 20–30% higher than in plain areas. Mountainous regions also face radar beam blockage issues, affecting observations of lower-level particle distribution. Research on the application of synthetic aperture radar in typhoon flood detection36 indicates that fusion of multi-source remote sensing data can partially compensate for the limitations of single radar observations. Radar detection capability is an important limiting factor for the warning method. Detection accuracy of dual-polarization parameters decreases with distance, with measurement errors significantly increasing beyond 150 km. Factors such as attenuation caused by heavy precipitation, beam broadening effects, and temporal resolution limitations (6 min) all affect warning accuracy. For rapidly developing meso- and small-scale convective systems, the existing observational temporal resolution may be insufficient, requiring the adoption of enhanced observation modes such as sector scanning, or supplementation with other observational methods to improve the timeliness and accuracy of warnings. For South China Sea typhoons, the warmer SSTs (28–30 °C) may enhance ice processes, requiring adjustment of the graupel threshold to 40–42%. The algorithm processes one radar volume in 30–45 s on standard operational hardware (8-core CPU, 16 GB RAM), meeting real-time requirements. System integration can utilize existing SWAN data protocols and GRAPES environmental fields, with warnings disseminated through standard CAP interfaces. Validation robustness and generalization The multi-level validation approach employed in this study strengthens confidence in the reported improvements. The cross-year blind testing ensures that the method performs well on completely unseen typhoon events, while the independent radar dataset validation confirms that the algorithm is not overfitted to specific radar characteristics or local conditions. The consistent performance across different validation approaches (cross-year: POD = 84.6%, cross-radar: POD = 82.7%, k-fold: POD = 83.9%) demonstrates the robustness of the particle-based warning method. The slight performance decrease in cross-radar validation (approximately 2% lower POD) can be attributed to minor differences in radar calibration and local environmental conditions. Future work will focus on developing adaptive calibration procedures to minimize these inter-radar variations and further improve the method’s transferability. Operational application prospects The particle characteristic-based typhoon short-term heavy precipitation warning method has good prospects for operational application. The warning system can be integrated into existing meteorological operational platforms, achieving real-time operation through automated processes. The system architecture includes modules for data collection, quality control, particle identification, feature extraction, and warning determination, with parallel computing enabling analysis completion within 2 min. It is recommended to adopt a sliding time window and stream processing architecture to maintain continuous warning information updates, implement enhanced observations for key areas, and improve temporal resolution to 2–3 min. A review of big data and artificial intelligence applications in typhoon disaster warning37 indicates that machine learning algorithms can further optimize warning models and improve warning accuracy. When forecasters use this system, they should note: warning information needs to be analyzed comprehensively in conjunction with weather patterns; focus on parameter change trends rather than single values; adjust thresholds according to regional and seasonal characteristics; increase the weight of warm rain process indicators during afternoon periods. In the future, warning models can be continuously optimized through machine learning methods, establishing localized parameter threshold databases and developing customized warning products for different users. Strengthening coordination with hydrological and emergency management departments to promptly convert warning information into disaster prevention and mitigation actions will fully leverage the role of this method in typhoon defense. Conclusion This study analyzed the vertical distribution characteristics of precipitation particles in inner and outer spiral rainbands during typhoon processes using dual-polarization radar observation systems, revealing significant differences between the two: inner rainbands exhibited deep convective development (average 11.8 ± 1.2 km, maximum up to 13.5 km), active ice processes (graupel particle content reaching 55% ± 10% at 5 km height), and intense melting layer characteristics (Zdr peak values of 3.8 ± 0.5 dB), while outer rainbands were dominated by shallower convection (< 10 km), weaker ice processes (graupel particle content only 20% ± 6%), and warm rain processes. Based on the quantitative relationship between particle vertical distribution characteristics and surface precipitation intensity, a multi-parameter early warning indicator system was constructed, incorporating graupel particle content, zero-degree bright band intensity, ice crystal layer thickness, low-level large raindrop proportion, and vertical reflectivity gradient, establishing a comprehensive warning index PI and its three-level warning threshold standards. Independent sample testing demonstrated that the new method achieved a hit rate of 84.6%, improving by 19.2 percentage points over traditional reflectivity threshold methods, reducing the false alarm rate by 15.7 percentage points, and extending the average warning lead time from 10–15 min to 25–40 min. Particularly for extreme precipitation > 100 mm/h, the hit rate reached 92.3% with a warning lead time of 35–50 min. This research elucidated the microphysical formation mechanism differences of short-term heavy precipitation in typhoon inner and outer rainbands, developed a new physics-based early warning method, and provided important scientific basis and technical support for enhancing typhoon precipitation refined forecasting and early warning capabilities and disaster prevention and mitigation levels. Future work will expand the dataset to include additional typhoon cases as more dual-polarization radar observations become available, potentially extending the analysis to a 10-year period once sufficient high-quality data accumulates. This will further validate the robustness of the particle-based warning method across a wider range of typhoon scenarios. Future improvements could incorporate artificial intelligence technologies for automated threshold optimization and pattern recognition, potentially extending warning lead times beyond 40 min. Machine learning approaches may also enable adaptive refinement based on regional climatology and seasonal variations. Acknowledgements We acknowledge NASA Goddard Space Flight Center for providing Global Precipitation Measurement (GPM) satellite data and the European Centre for Medium-Range Weather Forecasts (ECMWF) for ERA5 reanalysis data. We thank the China Meteorological Administration for Taizhou S-band dual-polarization radar data and Hangzhou radiosonde observations. Appendix A List of abbreviations AbbreviationFull nameCSICritical success indexFARFalse alarm ratioKdpSpecific differential phasePIPrecipitation indexPODProbability of detectionPPIPlan position indicatorPRISMAPreferred reporting items for systematic reviews and meta-analysesVCPVolume coverage patternVDIVertical diversity indexVGRVertical gradient of reflectivityVIFVariance inflation factorZhHorizontal reflectivity factorZdrDifferential reflectivityρhvCorrelation coefficientφdpDifferential phase Author contributions R.F. conceived the study design, performed the dual-polarization radar data analysis, developed the particle identification algorithms, and wrote the main manuscript text. T.W. supervised the project, secured access to radar network data, contributed to the warning index development, and revised the manuscript. H.L. conducted the statistical analysis of precipitation samples, prepared Figs. 1, 2, 3, 4 and 5, performed the validation experiments, and contributed to the methods section. R.F. and H.L. jointly analyzed the vertical distribution characteristics of precipitation particles. T.W. and R.F. interpreted the microphysical mechanisms and their implications for operational warning systems. All authors reviewed and approved the final manuscript. Funding This research was supported by the Fundamental Research Funds for the Central Universities (25CAFUC04090), the Civil Aviation Development Fund Education Talents Project (MHJY2025033), and a special project for building world-class universities and disciplines and characteristic development guidance of central universities (CZYL2025009). Data availability The datasets used and/or analysed during the current study available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.