Characterizing Short-Duration Summer Rainstorms in Nanjing, China, Using Multi-Source Remote Sensing and Explainable AI

Highlights

• Multi-source remote sensing datasets (CLDAS, FY-3G and IMERG) are combined to investigate urban summer rainstorms in Nanjing, China, during 2017–2024.
• China’s new-generation FY-3G radar is employed, for the first time, to reveal the three-dimensional structure of an illustrative Meiyu rainstorm in Nanjing.
• Integrating satellite remote sensing with explainable AI improves spatiotemporal characterization and interpretable statistical understanding of urban rainstorms.

Abstract

With global warming and rapid urbanization, short-duration summer rainstorms are becoming more intense and localized, posing growing challenges to urban flood resilience. However, their spatiotemporal characteristics, vertical structures, and environmental drivers remain poorly understood. Here, we combine multi-source remote sensing datasets and China’s new-generation satellite-borne dual-frequency precipitation radar observations to investigate summer rainstorms in Nanjing, China, during 2017–2024. Results reveal pronounced spatiotemporal heterogeneity, with higher rainfall intensities concentrated over urban and adjacent areas. During the study period, rainstorm intensity and duration increased by 7.44% and 38.63%, respectively, while the affected area decreased by 8.18%, indicating a transition toward more localized yet more intense rainfall events. Environmental analyses suggest that large-scale thermodynamic conditions and regional topographic forcing provide a favorable background for convection development, while local urban thermal effects may further modulate rainfall enhancement. Three-dimensional radar detection of an illustrative rainstorm event indicates an inverted-cone vertical structure, suggesting a mixed convective-stratiform precipitation structure involving both warm-rain and ice-phase processes. An Explainable Bayesian-Optimized XGBoost (EBOX) model further identifies near-surface air temperature and specific humidity as the primary environmental factors associated with rainstorm occurrence and development. Overall, this study highlights the value of integrating satellite remote sensing with explainable artificial intelligence to improve understanding of urban extreme rainfall and provide new insights into how climate change, topography, and urbanization jointly shape precipitation extremes in rapidly urbanizing monsoon regions.

1. Introduction

Owing to the influence of the East Asian monsoon, floods triggered by summer rainstorms have become one of the most frequent and destructive natural disasters in urban areas of China, causing huge social and economic losses [1]. In recent decades, climate warming has further increased the intensity, variability, and unpredictability of extreme rainfall events, repeatedly breaking historical records. This trend is illustrated by several high-impact events, including the 2020 Yangtze River “Super Meiyu” event [2], the “21·7” Henan flood [3], and the “23·7” Hebei flood [4], which caused severe urban waterlogging and cascading secondary hazards. These disasters have posed serious threats to local agricultural production, ecological systems, critical infrastructure, and public safety. Meanwhile, rapid urbanization has also altered the local hydroclimatic environment. The pronounced urban heat island effect has contributed to more localized, short-duration, high-intensity rainstorms, particularly in densely populated and economically developed metropolitan areas, thereby intensifying urban flood risks [5].

As a major city in the Middle-Lower Yangtze Plain, Nanjing has a subtropical monsoon climate characterized by abundant precipitation and a distinct Meiyu season [6]. Under the combined influences of global warming and rapid urbanization, summer rainstorm disasters in Nanjing have become increasingly frequent [7]. These events pose significant challenges to the city’s sustainable development, causing considerable economic losses and attracting widespread public concern [8]. Therefore, in the context of climate change and urbanization, investigating the characteristics and underlying mechanisms of summer rainstorms in Nanjing, as a representative microcosm of major cities in China, is of considerable practical importance.

Regarding summer rainstorms in Nanjing, most previous studies primarily relied on rain gauges to analyze the spatiotemporal characteristics of extreme rainfall events. For example, Sun et al. [9] analyzed minute-scale rainfall data collected over the past 25 years to investigate the precipitation patterns of short-duration summer rainstorms and found that 38.6–70.2% of rainstorm events across different gauges exhibited a single-peak structure. Du et al. [7] utilized hourly precipitation observations from seven rain gauges over the Nanjing region to investigate the spatial distribution and long-term variability of rainstorms from 1975 to 2015. The results indicated an increasing trend in maximum rainfall and revealed alternating wet and dry cycles of rainstorms in the central urban area during the past 40 years.

Beyond gauge-based investigations, satellite retrievals and radar observations have been increasingly applied to characterize urban extreme rainfall in recent studies. For instance, Li et al. [10] utilized several near-real-time satellite precipitation products to monitor a typical extreme rainfall event in Nanjing and found that satellite retrievals can effectively capture the temporal evolution of precipitation, although limitations remain in accurately representing rainfall intensity and spatial variability. Similarly, Sun et al. [11] adopted Doppler weather radar observations to examine the climatological characteristics of convective rainstorms, demonstrating pronounced spatial heterogeneity and clear diurnal variations in summer rainstorms over the Nanjing region.

In addition, some studies have employed reanalysis data and numerical modeling approaches to investigate the mechanisms underlying rainstorm occurrence in Nanjing. For example, Mao et al. [12] used reanalysis data to show that a typical Meiyu rainstorm event was driven by interactions within the large-scale circulation and enhanced moisture transport. Shen et al. [13] further applied the mesoscale Weather Research and Forecasting (WRF) model to examine the influence of urban underlying surfaces on the characteristics of summer rainstorm cloud clusters, illustrating that the urban heat island effect enhances convective activity in the atmospheric boundary layer and increases precipitation over urban areas.

In summary, existing studies have made significant progress in analyzing the temporal variability and rainfall patterns of rainstorms in Nanjing. However, due to limitations in observational data and analytical techniques, a comprehensive understanding of the high-resolution spatiotemporal variability of rainstorms across urban areas is still lacking. In particular, joint analyses based on multi-dimensional and heterogeneous datasets are still scarce, and exploratory studies on the three-dimensional internal structure of rainstorms remain limited. Moreover, systematic investigations of the local near-surface environmental drivers influencing summer rainstorms in Nanjing are still needed. These limitations hinder a more comprehensive understanding of rainstorm processes and constrain the development of refined urban disaster prevention and mitigation strategies.

With the rapid advancement of the Global Precipitation Measurement (GPM) mission, precipitation observation methods have become increasingly diverse. New-generation observation technologies, such as satellite-borne sensors and weather radars, have enabled the rapid acquisition of large areal precipitation information [14]. Numerous multi-source precipitation datasets integrating satellite, radar, and rain gauge observations have been developed both internationally and domestically, providing new data sources for investigating rainstorm spatiotemporal patterns [15,16,17,18]. In addition, advanced satellite-borne precipitation radar technologies have recently emerged, creating new opportunities for three-dimensional detection of rainstorm systems. On 16 April 2023, China successfully launched its first dedicated precipitation measurement satellite, Fengyun-3G (FY-3G). This satellite carries a domestically developed dual-frequency active phased-array precipitation measurement radar (PMR), which enables three-dimensional tomographic observations of rainstorms and offers a new approach for probing their fine-scale internal structures [19]. Moreover, in recent years, explainable artificial intelligence (AI) models have demonstrated considerable potential in precipitation research, offering new ways to identify the key factors influencing rainstorm formation and development [20].

Therefore, this study aims to analyze the spatiotemporal variability of summer rainstorms in Nanjing using multi-source precipitation datasets that integrate satellite retrievals, radar data, and rain gauge observations, thereby overcoming the limitations of previous gauge-only studies. The three-dimensional structure of an illustrative Meiyu rainstorm event in Nanjing is further examined, using new-generation FY-3G satellite observations for the first time. In addition, an explainable artificial intelligence model is employed to identify and interpret the key near-surface environmental factors influencing short-duration summer rainstorms in Nanjing. Overall, this study provides new insights into improving the scientific understanding of rainstorm characteristics in China’s large cities and offers data and technical support for addressing urban rainstorm disasters.

In the next section, we describe the data and methods used in this study. A presentation of the results and discussion follows in Section 3. Summarizing remarks and conclusions, we finalize the paper in Section 4.

2. Materials and Methods

2.1. Study Area

As the central city of the Yangtze River Delta urban agglomeration in China, Nanjing is located in the lower reaches of the Yangtze River, spanning 31°14′–32°37′N and 118°22′–119°14′E. The city covers an area of 6587.04 km² and has a typical subtropical monsoon climate characterized by abundant precipitation. During the summer months, strong convective processes frequently trigger high-intensity, short-duration rainstorms in Nanjing. For example, on 24 June 2022, a sudden downburst struck the central area near Xuanwu Lake, causing severe urban flooding (Figure 1a). The maximum rainfall intensity during this representative short-duration summer rainstorm reached 98.4 mm hr⁻¹, resulting in serious urban waterlogging and disruptions to daily life and transportation systems (Figure 1b,c).

In terms of the surface characteristics, Nanjing is dominated by hills and low mountains, which account for 64.68% of the total area, while plains, low-lying areas, and water bodies comprise the remaining 35.32% (Figure 2a). The city is traversed by the main stem of the Yangtze River and contains several major water systems, including the Qinhuai River and Xuanwu Lake. As shown in Figure 1d, highly urbanized areas are concentrated in the Gulou, Xuanwu, Jianye, Qinhuai, and Yuhuatai districts. Owing to their relatively small spatial extent and similar urban characteristics, these five districts are collectively defined as the “Urban Area” in this study (Figure 2b). The aforementioned extreme rainfall event occurred precisely within this urban area.

2.2. Data Sources

2.2.1. CMA Land Data Assimilation System (CLDAS) Datasets

The China Meteorological Administration (CMA) Land Data Assimilation System (CLDAS) provides hourly datasets for the Asian region (0°–65°N, 60°–160°E) at a spatial gridded resolution of 0.0625° × 0.0625°. The datasets include five major product categories: atmospheric driver parameters (ADP), land surface temperature, soil moisture, soil temperature, and soil relative humidity. CLDAS integrates multi-source information from satellite data (e.g., FY-2 and CMORPH), reanalysis and forecast products (e.g., ECMWF, GFS, EMSIP), automatic weather station observations, and digital elevation model (DEM) data. Data processing incorporates several techniques, including the STMAS, optimal interpolation, CDF matching, physical inversion, and topographic correction. Current evaluation results demonstrate that CLDAS provides high-quality meteorological datasets over China, with the ADP product outperforming comparable international datasets in both accuracy and spatial resolution [17]. The CLDAS near real-time (NRT) dataset, updated with a two-day delay, can offer reliable support for high-resolution land–atmosphere interaction studies.

This study employs the CLDAS-V2.0 NRT ADP product, which includes six meteorological variables: precipitation (PRCP), 2 m air temperature (TAIR), 2 m specific humidity (QAIR), 10 m wind speed (WIND), near-surface air pressure (PAIR), and shortwave radiation (SWDN). These data are used to analyze the spatiotemporal characteristics and near-surface atmospheric drivers of summer rainstorms in Nanjing from June to August during the period 2017–2024.

2.2.2. FY-3G Precipitation Measurement Radar (FY-3G PMR) Observations

The Fengyun-3 (FY-3) series represents China’s second-generation polar-orbiting meteorological satellite system, forming an integrated observation framework that combines operational polar orbiters with dedicated precipitation measurement satellites in low-inclination orbits. As the third satellite in this series, FY-3G was successfully launched on 16 April 2023. Its core payload is China’s first spaceborne dual-frequency precipitation measurement radar (FY-3G PMR), which operates at Ku- and Ka-bands (13.6 GHz and 35.55 GHz) with horizontal polarization. Using a cross-track phased-array scanning system, the PMR achieves a scanning swath of ±20.3°, enabling three-dimensional observations of internal precipitation structures within 50° latitude and at altitudes below 18 km [21].

The FY-3G PMR provides a horizontal resolution of 5 km and a vertical resolution of 250 m at the subsatellite point, enabling the capture of fine spatial details of precipitation structures. Compared with the Dual-frequency Precipitation Radar onboard the Global Precipitation Measurement mission (GPM-DPR), China’s FY-3G PMR provides a wider swath (up to 303 km vs. 245 km for the GPM-DPR) and finer vertical resolution, allowing more detailed three-dimensional detection of precipitation systems [19].

In this study, rain rate data retrieved from the Ku-band radar observations of FY-3G PMR are used to analyze the three-dimensional structure of intense precipitation systems during an illustrative extreme rainfall event.

2.2.3. Other Data Sources

The study also employs the widely used IMERG V07 product as a supplementary satellite estimate to compare with CLDAS and to examine the spatiotemporal patterns of typical rainstorm events. As a core dataset of the global precipitation observation system, IMERG integrates passive microwave observations from the GPM constellation with infrared data from geostationary satellites to generate global precipitation estimates with a spatial resolution of 0.1° and a temporal resolution of 30 min [14].

ERA5 reanalysis data provided by the European Centre for Medium-Range Weather Forecasts (ECMWF) are also used to characterize large-scale atmospheric conditions associated with rainstorm events, offering hourly global atmospheric fields at a spatial resolution of approximately 0.25° [22].

Additional datasets describing the underlying surface include the Shuttle Radar Topography Mission digital elevation model (SRTM-DEM) [23], global 10 m land cover data provided by the Environmental Systems Research Institute (ESRI Land Cover) derived from Sentinel-2 imagery [24], and a standard administrative boundary map supplied by the National Geomatics Center of China (NGCC). The main datasets used in the study are listed in

Table 1. Overview of the main datasets used in this study.

Product	Resolution	Coverage	Producer	Data Acquisition Website
CLDAS-NRT	0.0625°/1 h	Asia	PRC/CMA	https://data.cma.cn/data/cdcdetail/dataCode/NAFP_CLDAS2.0_NRT.html (accessed on 15 January 2025)
FY-3G PMR	5 km/250 m	Orbital swath	PRC/NSMC	https://satellite.nsmc.org.cn/DataPortal/cn/data/structure.html (accessed on 7 July 2025)
IMERG-Final	0.1°/30 min	Global	USA/NASA	https://gpm.nasa.gov/data/imerg (accessed on 14 July 2025)
ERA5	0.25°/1 h	Global	ECMWF	https://cds.climate.copernicus.eu/datasets (accessed on 17 June 2026)
SRTM-DEM	30 m	Near global	USA/NASA	https://www.earthdata.nasa.gov/data/instruments/srtm (accessed on 15 January 2025)
ESRI Land Cover	10 m/1 y	Global land	USA/ESRI	https://livingatlas.arcgis.com/landcover/ (accessed on 15 January 2025)

.

2.3. Methods

2.3.1. Analysis for Time Series

To obtain a preliminary understanding of summer precipitation characteristics in Nanjing prior to the rainstorm analysis, the study first conducts a time series analysis of precipitation data for the summers from 2017 to 2024. Since precipitation typically follows a long-tailed distribution, the time series is decomposed into trend, seasonal, and residual components using an additive model to reduce the influence of short-term fluctuations and highlight underlying trends [25].

The trend component, computed by a moving average, captures the long-term evolution of precipitation. Missing values at the boundaries are filled using linear extrapolation to maintain consistency with the original time series. The seasonal component, which reflects fixed periodic fluctuations, is calculated by averaging detrended values at corresponding positions within each cycle and projecting the results along the time axis. This component is subsequently smoothed using a Savitzky-Golay (Savgol) filter, which applies local polynomial fitting to reduce noise while preserving abrupt variations [26]. To further identify summer precipitation trends, the Mann–Kendall (M-K) test, a nonparametric method that is robust to outliers and distribution assumptions, is applied to the trend components at both regional and grid scales [27,28].

2.3.2. Analysis for Spatial Characteristics

The study develops an hourly summer rainfall dataset for Nanjing using high-resolution CLDAS-V2.0 gridded precipitation data for the period 2017–2024. Following previous studies that apply percentile-based thresholds to define extreme precipitation [29,30,31], short-duration rainstorms are identified using the 95th percentile threshold at each grid cell based on precipitation samples (≥0.1 mm·h⁻¹), thereby generating a corresponding hourly summer rainstorm dataset. Given the typically single-peak structure of these events, rainstorms are treated as temporally continuous at the hourly scale to avoid excessive fragmentation of intermittent precipitation occurrences [32]. Accordingly, a finite state machine is employed to detect rainstorm events and reconstruct precipitation systems using hourly sliding windows. Within this framework, consecutive rainstorm hours are merged into a single event when they are continuously flagged by the state machine, while short temporal interruptions are allowed and re-absorbed into the same event if they occur within the continuity window defined by the algorithm.

Four key characteristics of short-duration rainstorms are defined: intensity (average hourly rainfall exceeding the threshold), frequency (number of discrete rainstorm events at the hourly scale), duration (mean persistent period per event), and coverage (average proportion of the total area affected per event). At the grid scale, events are first identified independently for each grid cell, whereas regional-scale events are subsequently constructed by merging spatiotemporally overlapping grid-level events that occur within the same time window and exhibit spatial adjacency, thereby forming coherent regional precipitation systems.

This study systematically investigates short-duration rainstorms from three complementary perspectives across multiple spatial scales. First, a grid-based analysis is conducted to characterize the spatial patterns of rainstorm properties across the summer months in Nanjing. Second, a process-based rainfall event identification framework is applied to detect all rainstorm events during the summers of 2017–2024 at the regional scale. Based on this framework, the diurnal cycles, monthly differences, and interannual variability in rainstorm intensity, frequency, duration, and spatial coverage are quantitatively evaluated. Third, multi-year averages and interannual variations of rainstorm characteristics are further examined at both grid and regional scales across seven administrative districts (Pukou, Luhe, Qixia, Jiangning, Lishui, Gaochun, and the Urban Area) to identify spatial heterogeneity and regional contrasts. Based on these analyses, an illustrative rainstorm event is selected, and FY-3G satellite precipitation radar observations are used to detect its three-dimensional structural characteristics.

2.3.3. Explainable Bayesian-Optimized XGBoost (EBOX) Model

To examine how near-surface meteorological factors influence urban summer rainstorms, we propose an Explainable Bayesian-Optimized XGBoost (EBOX) model. EBOX integrates XGBoost, Bayesian Optimization, and SHAP into a unified framework for multi-source data training, model optimization, and result interpretation. XGBoost is an efficient gradient boosting algorithm that employs additive training and regularization to improve model performance. It enhances computational efficiency without compromising accuracy through feature importance ranking and parallel processing [33]. Bayesian Optimization is adopted for hyperparameter tuning, as it provides a sequential global optimization method based on probabilistic models and is particularly suitable for black-box problems with computationally expensive or analytically intractable objective functions. This method iteratively approaches the global optimum by alternately updating a surrogate model and optimizing an acquisition function [34]. Finally, SHAP (SHapley Additive exPlanations) provides a game-theoretic interpretability framework that quantifies the contribution of each feature to the model predictions. By satisfying the properties of additivity and consistency, SHAP offers a rigorous and intuitive approach for interpreting complex black-box models [35]. In this study, SHAP is used primarily for ranking feature importance and is further employed to examine nonlinear response patterns through dependence analyses.

3. Results and Discussion

3.1. Spatiotemporal Variation Analysis of Rainstorms

First, we analyzed the interannual variability of summer precipitation in Nanjing. During the study period (2017–2024), the mean total precipitation across the three summer months reached 528 mm. The time series exhibited a clear quasi-periodic fluctuation with an approximately three-year cycle (Figure 3a). Additionally, statistical analysis indicated that the mean rain rate increased significantly by 33.18% over the study period (z = 44.60; Figure 3b). Specifically, the seasonal component of precipitation displayed a distinct periodic pattern, characterized by higher rainfall in June and July and lower rainfall in August (Figure 3c). This pattern is likely associated with the dominant Meiyu rainfall period, which typically begins in mid-June and ends in early July. Furthermore, analysis of the residual distribution revealed that although negative residuals were more frequent, positive residuals exhibited larger absolute magnitudes. This suggests that extreme precipitation events, when they occur, tend to be more intense and contribute disproportionately to the overall variation (Figure 3d).

Table 2. Interannual variation rates of rainstorm characteristics in Nanjing during the study period.

	Intensity	Frequency	Duration	Coverage
June	12.19%	46.43%	59.07%	4.04%
July	11.11%	29.49%	56.81%	−17.47%
August	5.41%	−49.42%	18.97%	8.64%
Average	7.44%	0.16%	38.63%	−8.18%

summarizes the interannual variation rates of summer rainstorm characteristics (including intensity, frequency, duration, and coverage) in Nanjing during the study period. Both rainstorm intensity and duration exhibited increasing trends, rising by 7.44% and 38.63%, respectively. The increase in rainstorm intensity was mainly concentrated in June (12.19%) and July (11.11%), corresponding to Figure 4e–g, which may be associated with intensified Meiyu activity and more frequent typhoon influences [6,36,37]. In contrast, rainstorm frequency increased in June (46.43%) and July (29.49%) but decreased substantially in August (−49.42%), resulting in no significant overall trend. Unlike the other characteristics, the overall spatial coverage of rainstorms decreased by 8.18%, suggesting that summer rainstorms have become increasingly localized. The most pronounced reduction occurred in July, with a coverage decrease of 17.47%.

Figure 5 shows the spatial distribution of monthly summer rainstorm characteristics in Nanjing. Rainstorm intensity peaks in June and July before declining in August (Figure 5a–c) were consistent with the cumulative rainfall patterns shown in Figure 4a–c. Although the total accumulated rainfall is relatively higher in July, the most intense short-duration events are primarily concentrated in late June. This pattern is mainly associated with the Meiyu season, which typically occurs from late June to early July under the influence of the western Pacific subtropical high [6,38,39]. The Meiyu front represents the dominant source of intense summer precipitation over the Middle-Lower Yangtze Plain, including Nanjing, contributing approximately 37% of the total summer rainfall on average (Figure 3a,c). Furthermore, typhoon-induced precipitation, which mainly occurs in July, often overlaps with the Meiyu period and further increases total rainfall [40].

Spatially, rainstorms are mainly concentrated in the Urban Area and its adjacent districts such as Qixia, Jiangning, and Pukou (Figure 5d), aligning with the spatial pattern of average summer rainfall (Figure 4d). This concentration is likely associated with the urban heat island effect. Elevated urban temperatures may enhance heat fluxes into the near-surface atmosphere, thereby strengthening local convection and intensifying precipitation, which could contribute to the formation of an urban rain island [41,42,43,44].

In terms of frequency, rainstorms occur most frequently in July, followed by June, while August shows the lowest occurrence (Figure 5e–h). In July, rainstorms are mainly concentrated in the hilly regions of Gaochun, Lishui, and Luhe, whereas in June and August they occur more frequently in the Urban Area (Figure 5e,f).

Regarding duration, long-lasting rainstorm events are primarily observed near major water bodies and are most prominent in August (Figure 5k,l). Enhanced evaporation from water surfaces increases atmospheric moisture availability, which may prolong rainfall duration [45,46]. Moreover, in June and July, rainstorm duration also shows a potential relationship with local topographic conditions (Figure 5i,j).

Corresponding to the spatial distribution in Figure 5,

Table 3. Summary of rainstorm characteristics across the administrative districts of Nanjing.

Month	District	Intensity (mm hr−1)	Frequency (times)	Duration (min)
June	Luhe	12.96	23.73	94.37
	Pukou	15.16	25.61	109.95
	Qixia	13.56	27.92	101.73
	Jiangning	15.88	28.67	105.96
	Lishui	14.17	28.59	94.82
	Gaochun	12.78	28.78	95.22
	Urban Area	15.60	32.70	104.12
July	Luhe	13.91	31.57	89.83
	Pukou	15.18	30.65	97.81
	Qixia	15.12	33.62	103.62
	Jiangning	14.35	29.20	98.92
	Lishui	13.84	32.69	106.39
	Gaochun	13.03	34.75	92.93
	Urban Area	15.20	30.70	94.36
August	Luhe	13.75	19.55	96.26
	Pukou	13.93	21.97	97.61
	Qixia	14.91	23.23	113.97
	Jiangning	13.10	26.16	92.91
	Lishui	13.16	23.97	89.85
	Gaochun	13.19	23.16	84.90
	Urban Area	13.89	27.40	104.48
Average	Luhe	13.56	24.95	92.74
	Pukou	14.83	26.08	101.16
	Qixia	14.55	28.26	105.63
	Jiangning	14.54	28.01	99.09
	Lishui	13.78	28.42	97.74
	Gaochun	12.99	28.90	91.32
	Urban Area	14.93	30.27	100.90

Notation: Statistical analyses were based on the average values of all grid cells within each administrative district. Maximum values for rainstorm characteristics are highlighted in red, while minimum values are shown in blue.

further presents the regional averages of summer rainstorm characteristics across the administrative districts of Nanjing. The highest rainstorm intensities in June, July, and August are observed in Jiangning (15.88 mm hr⁻¹), Urban Area (15.20 mm hr⁻¹), and Qixia (14.91 mm hr⁻¹), respectively. Overall, Urban Area exhibits the highest summer intensity (14.93 mm hr⁻¹), whereas Gaochun records the lowest, approximately 2 mm hr⁻¹ lower than that of the Urban Area and more than 3 mm hr⁻¹ lower than the June maximum.

On the other hand, the Urban Area also shows the highest rainstorm frequency, with approximately thirty short-duration rainfall events per month during summer at the grid scale. Although Luhe exhibits the lowest frequency, its rainstorm occurrence exceeds that of the Urban Area in July. Regarding duration, long-lasting rainstorm events are more frequently observed in Pukou, Lishui, and Qixia. This pattern may be associated with their proximity to major water bodies, which could enhance atmospheric moisture availability, as well as complex terrain that may help sustain precipitation systems.

To further examine the spatial heterogeneity of summer rainstorm variations in Nanjing, Figure 6 illustrates interannual changes in rainstorm intensity across the city’s administrative districts. The results reveal a clear spatial contrast. Districts located south of the Yangtze River exhibit increasing trends in rainstorm intensity (Figure 6b,c,e–g), whereas the northern districts of Pukou and Luhe display no significant trends (Figure 6a,d). Among all districts, Urban Area and Gaochun show the largest increases in rainstorm intensity, with increases of 10.52% and 12.95%, respectively, while Pukou exhibits an overall decrease of 3.55%.

These spatial patterns suggest that the Yangtze River, as a major river traversing the Nanjing region, plays an important role in shaping local microclimates. The southern districts, located at lower latitudes, represent key areas where Meiyu rainfall frequently overlaps with typhoon-related precipitation [47,48,49]. Although suburban districts, such as Lishui and Gaochun, which are less affected by the urban heat island effect, generally experience lower rainstorm intensities, recent years have witnessed increases in extreme Meiyu events and a northward shift of summer typhoon activity under global warming [50,51,52]. These changes have likely strengthened the combined influence of Meiyu and typhoon precipitation in the southernmost districts, where rainstorm intensity increased by 8.40% and 12.95%, respectively, while duration lengthened by nearly 25% (

Table 4. Statistical results of the average ISR (impervious surface ratio), ELE (elevation), TAIR (air temperature), and QAIR (specific humidity) for each district in Nanjing.

District	ISR (%)	ELE (m)	TAIR (°C)	QAIR (g·kg−1)
Gaochun	16.49	12.31	28.34	19.37
Lishui	20.84	28.35	27.95	18.95
Luhe	22.77	28.44	27.79	18.76
Pukou	30.28	21.62	28.21	18.89
Jiangning	35.58	36.21	28.15	18.85
Qixia	42.42	28.93	28.11	18.66
Urban Area	72.01	14.40	28.38	18.77

Notation: Statistical analyses were based on the average values of all grid cells within each administrative district. Maximum values are highlighted in red, while minimum values are shown in blue.

). In addition, the large water bodies south of the Yangtze River, such as Xuanwu Lake and Shijiu Lake, may provide additional moisture to the lower atmosphere, potentially contributing to local rainfall enhancement.

The Urban Area, located south of the river, is characterized by the highest impervious surface ratio (ISR; 72.01%) and limited vegetation and water coverage (Figure 2b), which tends to favor greater surface heating and less evaporative cooling [53]. Notably, although the Urban Area is neither the lowest in latitude nor the lowest in elevation, it still exhibits the highest summer mean air temperature (28.38 °C) among all districts (Table 4). This pattern is not strictly consistent with variations in elevation or latitude and may reflect the combined influence of spatial variability in specific humidity and urban thermal conditions. Combined with intense human activities and anthropogenic heat emissions, these conditions may contribute to pronounced urban heat island effects [54]. Global warming may further exacerbate these effects by intensifying summer heat anomalies and convective activity [55]. Consequently, rainstorm frequency has increased more markedly in the Urban Area and its surrounding districts than in suburban areas such as Luhe, Lishui, and Gaochun (

Table 5. Interannual variation rates of summer rainstorm characteristics by administrative district in Nanjing during the study period.

	Intensity	Frequency	Duration	Coverage
Pukou	−3.55%	29.84%	3.24%	4.13%
Luhe	−0.27%	9.03%	13.86%	−17.83%
Qixia	1.42%	35.38%	14.81%	4.97%
Jiangning	4.18%	35.93%	24.13%	−7.44%
Lishui	8.40%	9.78%	24.34%	4.12%
Gaochun	12.95%	11.19%	24.57%	2.97%
Urban Area	10.52%	81.01%	−5.52%	−7.95%

).

However, this enhancement should be interpreted with caution. Analysis of boundary layer stability indicates that rainstorm conditions are generally characterized by increased CAPE and markedly reduced CIN, both of which favor the initiation and development of convective activity (Figure A1a,c,d,f). In contrast, BLH is typically lower than that observed under non-rainstorm conditions. This difference likely reflects the combined influence of thermodynamic and precipitation processes and warrants further investigation (Figure A1b,e).

Our analyses additionally suggest that large-scale thermodynamic and dynamical conditions provide an important background for convection development, while local urban thermal effects may play a modulating role rather than acting as an isolated driver.

ERA5 reanalysis results further reveal that rainstorms in the Nanjing region are characterized by a transition from subsidence to pronounced upward motion (Figure 7a,d), accompanied by enhanced low-level humidity and substantially strengthened moisture transport at 850 hPa, while the prevailing circulation shifts from southerly to southwesterly flow (Figure 7b,c,e,f). These changes suggest that large-scale dynamical conditions, latitudinal gradients, and regional topography collectively influence moisture transport and convective development.

From a topographic perspective, the low-elevation corridor in the southern part of the region may facilitate the northward transport of warm, moist air into the urban area, whereas the gradually rising terrain to the north may enhance low-level lifting and favor convective development, thereby providing favorable conditions for convection initiation and intensification.

Further analysis incorporating both impervious surface ratio and elevation demonstrates that, under comparable topographic conditions, higher rainstorm intensity is generally associated with higher levels of urbanization. Conversely, within specific urbanization ranges, rainfall intensity also exhibits an increasing trend with elevation. These findings suggest that rainstorm characteristics in Nanjing are governed by the combined influences of urbanization and topographic forcing, with neither factor acting in isolation, highlighting the synergistic roles of land-surface modification and terrain effects in shaping precipitation patterns (Figure A2).

During the extreme summer drought of 2022 (Figure 3a,b), a downburst occurred in central Nanjing (Figure 1a), characterized by high intensity, short duration, and limited spatial coverage. This event was likely associated with a brief outbreak of severe localized convection. Urban thermal environments may have contributed to its development, but the available evidence does not allow a robust attribution of this event to urban heat island effects [56].

Given the hourly temporal resolution of the CLDAS dataset, the diurnal variation of summer rainstorm characteristics in Nanjing is further analyzed (Figure 8). The mean intensity, frequency, duration, and area ratio of rainstorms are calculated for each hour during the summer season. The results show that the regional mean rainstorm intensity exhibits relatively weak diurnal variation (Figure 8a).

However, rainstorm frequency exhibits a clear bimodal pattern. The first peak occurs at 14:00, corresponding to the period when afternoon thermal convection is most active. The second peak appears around 18:00, suggesting that convective systems initiated in the afternoon often continue to develop and persist into the early evening (Figure 8b).

In contrast, rainstorm duration shows a pattern opposite to that of frequency. From midnight to early morning, rainstorms tend to persist for longer periods, whereas daytime rainstorms are generally shorter in duration (Figure 8c). This difference likely arises because nighttime rainfall is often associated with mesoscale convective systems or stratiform precipitation processes, which are more organized and propagate slowly, thereby producing longer-lasting rainfall [57,58]. By contrast, daytime rainstorms are primarily triggered by local thermal convection, which develops rapidly but persists for shorter periods [59,60].

The spatial coverage of individual rainstorms is generally limited, with mean coverage not exceeding 40% of the total area of Nanjing. Its diurnal variation is similar to that of rainstorm duration, with larger coverage at night and smaller coverage during the daytime (Figure 8d). This pattern may be explained by the fact that nighttime precipitation systems typically have larger spatial scales and stronger organization, whereas daytime convective rainfall is usually more localized and scattered [61,62].

3.2. Three-Dimensional Structure of an Illustrative Rainstorm

According to the spatiotemporal overlap between the scanning of FY-3G satellite orbits and the occurrence of heavy rain, we selected an illustrative Meiyu rainfall event that occurred on 12 July 2024. The representativeness of this case is further supported by its rainfall characteristics, with daily cumulative rainfall and maximum rain rate reaching the 91.1st and 81.5th percentiles, respectively, among all summer rainstorm events from 2017 to 2024.

The illustrative rainstorm event mainly affected Pukou, Urban Area, Jiangning, Lishui, and Gaochun (Figure 9a,d), with rainfall peaks observed in Pukou and Gaochun (Figure 9b,e). The rainstorm center migrated southward from Pukou at 3:00 a.m. and reached Gaochun by 9:00 a.m. (Figure 9c,f). Combined with its high percentile ranking in rainfall intensity, these characteristics support the selection of this event as a representative Meiyu-period rainstorm case for detailed analysis.

Analysis of continuous rainfall during 9–14 July shows that both CLDAS and IMERG consistently capture rainfall occurrence. However, IMERG systematically underestimates extreme rainfall intensities, which may lead to a smoothing of peak magnitudes and a conservative representation of rainstorm severity (Figure 10). This discrepancy implies that IMERG may underestimate the magnitude of peak rainfall intensity, particularly during peak phases. Although both datasets consistently capture the spatial evolution and overall movement of the rainfall system, IMERG’s underestimation may slightly weaken the apparent spatial gradients of rainfall intensity and thus influence the characterization of rainfall cores. In addition, the underestimation of peak intensity may affect the identification of peak rainfall characteristics, particularly during periods of rapid intensity fluctuations.

FY-3G PMR observations reveal a well-organized three-dimensional precipitation structure (Figure 11). The maximum precipitation rate occurred near the rainstorm center, where the scan line passed directly over the urban area of Nanjing (Figure 11b). A vertical profile of precipitation rate with altitude was obtained by constructing a cross-section perpendicular to the scan line (Figure 11c).

The vertical structure reveals an inverted-cone morphology. Intense precipitation is mainly concentrated below the ~5 km freezing layer, where localized precipitation cores indicate strong updrafts within the storm system. These intense updrafts likely arise from strong low-level convergence of warm, moist air, forming a narrow yet vigorous convective column [63]. Such a compact convective system can efficiently generate heavy precipitation within a short period [64]. The strong low-level reflectivity and rapid intensification below the freezing level may indicate contributions from warm-rain processes. However, the vertical structure also clearly suggests a deep convective system with a well-developed stratiform region, implying that ice-phase microphysical processes are also likely involved, particularly above the freezing level and within the stratiform regions. Meanwhile, the coexistence of convective cores and upper-level precipitation signatures suggests a mixed-phase microphysical structure involving both warm-rain and ice-phase processes.

Horizontally, near-surface rainfall exhibits a distinct north–south rainband (Figure 11d), suggesting that the rainstorm was governed by a mesoscale convective line [65,66]. This structure was likely maintained by a southward-propagating cold-pool outflow boundary that continuously triggered convection along its leading edge [67,68].

Figure 12 further illustrates the relationships among the bright band, precipitation type, precipitation rate, and storm top height, following the analysis framework of [69]. As shown in Figure 12a, the bright band spatially coincides with regions where precipitation rates are lower than 4 mm hr⁻¹ from Figure 11d, indicating a strong association with stratiform precipitation [70,71]. This rainfall type normally results from gradual and sustained uplift of warm, moist air and is characterized by relatively low storm top heights (Figure 12b).

In contrast, regions without a bright band either exhibit no precipitation or are dominated by convective rainfall [72,73]. Convective precipitation, driven by strong convection, is generally characterized by higher precipitation rates and larger storm top heights (Figure 12d). Combined with the spatial pattern shown in Figure 11d, these results indicate that convective precipitation dominates high-intensity rainfall regions, whereas stratiform precipitation prevails in areas with lower rainfall intensity (Figure 12c). The spatial organization highlights the strongly mixed convective-stratiform nature of this rainstorm and is consistent with the vertical structural characteristics shown in Figure 11c.

3.3. Near-Surface Environmental Drivers of Rainstorm Development

Finally, we developed an Explainable Bayesian-Optimized XGBoost (EBOX) model to quantitatively identify the near-surface environmental drivers of summer rainstorms in Nanjing. The overall EBOX modeling framework is illustrated in Figure 13. Rainfall samples (≥0.1 mm hr⁻¹) were first extracted from the hourly precipitation field (PRCP) of the CLDAS dataset. Samples exceeding the 95th percentile threshold were defined as positive cases (heavy rainfall), while the remaining samples were treated as negative cases (non-heavy rainfall). Five near-surface environmental variables derived from the CLDAS atmospheric driving product (ADP), including air temperature (TAIR), specific humidity (QAIR), downward shortwave radiation (SWDN), surface pressure (PAIR), and wind speed (WIND), were then selected as input features. The dataset was subsequently divided into training and testing subsets using stratified sampling with a ratio of 3:1.

To better assess model performance for heavy rainfall events, the precision–recall area under the curve (PR-AUC) was adopted as the primary optimization objective [74]. Model hyperparameters were optimized using Bayesian optimization with ten-fold cross-validation. A surrogate probabilistic model iteratively approximates the mapping between hyperparameter configurations and model performance, while an acquisition function efficiently guides the search toward promising regions of the parameter space. This process was integrated with the XGBoost algorithm, in which gradient-boosted decision trees are sequentially constructed to minimize the loss function, enabling adaptive learning of complex nonlinear relationships.

Class imbalance was addressed by assigning different weights to positive and negative samples [33], although the extreme imbalance between heavy rainfall and non-heavy rainfall samples inevitably limits classification precision and leads to elevated false-alarm rates in hourly-scale rainstorm identification. Model performance in classifying hourly heavy rainfall events was evaluated using standard binary classification metrics (

Table A1. Primary classification metrics used in this study.

Metric	Abbreviation	Formula	Best Value
Precision	$P$	$P=TP/(TP+FP)$	1
Recall (True Positive Rate)	$R \left(TPR\right)$	$R=TPR=TP/(TP+FN)$	1
F1-Score	$F1$	$F1=2\times P\times R/(P+R)$	1
Accuracy	$A$	$A=(TP+TN)/(M+N)$	1
False Positive Rate	$FPR$	$FPR=FP/(FP+TN)$	0
Receiver Operating Characteristic Area Under Curve	$ROC\text{-}AUC$	$ROC\text{-}AUC={\int }_0^{1}R\left(FPR\right)d\left(FPR\right)$	1
Precision-Recall Area Under Curve	$PR\text{-}AUC$	$PR\text{-}AUC={\int }_0^{1}P\left(TPR\right)d\left(TPR\right)$	1

Notation: $TP$ represents the number of samples correctly predicted as positive; $FN$ represents the number of samples that are actually positive but predicted as negative; $FP$ represents the number of samples that are actually negative but predicted as positive; $TN$ represents the number of samples correctly predicted as negative; $M$ represents the total number of positive samples; and $N$ represents the total number of negative samples.

), with results presented in Figure A3 and

Table A2. Classification report on the EBOX model.

Label	Precision	Recall	F1-Score	Support
0	0.97	0.81	0.88	87,021
1	0.21	0.67	0.32	6808
Macro Average	0.59	0.74	0.60	93,829
Weighted Average	0.91	0.80	0.84	93,829
Accuracy	-	-	0.80	93,829

. Notably, Figure A3 shows that the model exhibits stable convergence, with PR-AUC increasing on both the training and validation sets as the number of iterations progresses. However, a clear gap between training and validation PR-AUC also indicates unavoidable generalization limitations under noisy high-resolution precipitation inputs. Finally, a tree model interpreter incorporating SHAP explanations was employed to quantify the relative importance of near-surface environmental factors on summer rainstorm development in Nanjing.

Importantly, the EBOX model was not designed for operational forecasting or causal attribution but rather for identifying statistical relationships in the data. Accordingly, SHAP-based interpretation was conducted primarily on the training set, where PR-AUC approaches ~0.7 (Figure A3), providing a sufficiently stable basis for subsequent interpretability analysis.

Figure 14 presents the SHAP-based analysis of near-surface environmental drivers influencing summer rainstorms in Nanjing. Among the five environmental variables analyzed, air temperature (TAIR) and specific humidity (QAIR) exert the strongest influence on rainstorms, with average contribution rates of 37.09% and 21.45%, respectively (Figure 14a). The SHAP feature density distribution further indicates that relatively lower near-surface air temperature, higher specific humidity, and reduced downward shortwave radiation favor the occurrence and development of rainstorms (Figure 14b). Excessively high air temperature, however, is generally unfavorable for rainstorm development. Strong surface heating is typically associated with clear-sky conditions or the pre-convective stage prior to convective initiation. Such conditions tend to enhance atmospheric stability and suppress convective development, thereby limiting the occurrence of rainstorms [75,76]. Overall, these relationships likely reflect the combined thermodynamic environment associated with rainstorm occurrence at the regional scale, as opposed to localized surface effects.

The feature dependence plots illustrate the distribution of SHAP values for each variable, indicating that the TAIR and QAIR values associated with the highest rainstorm probability are approximately 22.13 °C and 20.67 g kg⁻¹, respectively (Figure 14c,d). These results suggest nonlinear response patterns and potential threshold-like behavior in key meteorological variables associated with increased rainstorm probability. A near-surface air temperature of around 22 °C generally reflects thermodynamically favorable warm and humid conditions that support convective systems and promote latent heat release through water vapor condensation, thereby intensifying precipitation processes [77,78]. Meanwhile, elevated specific humidity indicates abundant low-level moisture supply, which favors strong convection and facilitates the initiation and persistence of rainstorms [79,80,81]. Given that the analysis integrates observations from both urban and rural environments, the identified temperature optimum is more appropriately interpreted as a regional-scale statistical characteristic of rainstorm-favorable atmospheric conditions.

4. Conclusions

This study presents a systematic characterization of short-duration summer rainstorms in Nanjing by using multi-source precipitation datasets that integrate satellite retrievals, radar data, and rain gauge observations. In particular, observations from China’s new-generation satellite-borne, dual-frequency precipitation radar are employed, for the first time, to reveal the three-dimensional structure of an illustrative Meiyu rainstorm in Nanjing. An explainable artificial intelligence model (EBOX) was developed to quantify the near-surface environmental drivers of rainstorm occurrence and development. The principal findings of this study are summarized as follows:

Short-duration summer rainstorms in Nanjing exhibit pronounced spatiotemporal heterogeneity regulated by the combined effects of urban land-surface modification, regional topographic forcing, and large-scale atmospheric processes. Spatially, higher rainfall intensity is primarily concentrated over the urban area and surrounding districts, such as Qixia, Jiangning, and Pukou. Interannual variability further reveals an intensification of rainstorms south of the Yangtze River, particularly in the Urban Area and Gaochun, suggesting combined influences from Meiyu fronts, typhoon systems, and their interactions. Temporally, both intensity and duration of summer rainstorms increased by 7.44% and 38.63%, respectively, while the affected area decreased by 8.18%, indicating a transition toward more localized yet intense precipitation events. Monthly variations show peak intensity in June and July, reflecting the dominant influence of the Meiyu season. In terms of the diurnal cycle, rainstorm frequency exhibits a bimodal distribution, which is associated with afternoon convective triggering and nighttime system persistence.

Observations from the dual-frequency precipitation measurement radar onboard the FY-3G satellite (FY-3G PMR) reveal the three-dimensional structural characteristics of an illustrative Meiyu rainstorm. The illustrative event exhibits a distinct inverted-cone vertical structure, with intense precipitation mainly confined below the ~5 km freezing layer, indicating strong low-level moisture convergence and vigorous updraft development in warm-rain processes. However, the vertical structure also indicates that ice-phase microphysical processes and mixed-phase interactions are likely active above the freezing level, especially within the stratiform precipitation region. Horizontally, the rainfall displays a north–south-oriented banded structure associated with a mesoscale convective line, where cold-pool outflow boundaries continuously trigger new convection. The coexistence of convective and stratiform precipitation further demonstrates the complex mixed convective-stratiform nature of this Meiyu rainstorm system.

The proposed explainable EBOX model identifies near-surface environmental conditions as key environmental factors on summer rainstorms in Nanjing. Among the considered variables, near-surface air temperature (TAIR) and specific humidity (QAIR) exert the strongest influences, contributing 37.09% and 21.45% to model predictions, respectively. Results indicate that moderately warm temperatures combined with high atmospheric humidity are associated with enhanced probabilities of deep convection and increased latent heat release processes. These thermodynamic conditions reflect the broader environmental background over the Nanjing domain, rather than being confined to urban areas.

Generally, this study establishes an integrated framework for investigating urban summer rainstorms by combining multi-source satellite observations with an explainable machine learning approach. The results advance the understanding of the spatiotemporal variability, three-dimensional internal structure, and near-surface environmental drivers of short-duration summer rainstorms in monsoon-influenced megacities. These findings highlight the value of integrating satellite remote sensing with explainable artificial intelligence to improve statistical understanding for urban extreme rainfall under climate change and rapid urbanization. Future work will focus on physical mechanism verification using additional observational evidence and numerical simulations to further test and refine the proposed interpretations.