Improving the Data Consistency Between GPM and Weather Radar with Advection Correction

Highlights

What are the main findings?

• Advection correction can effectively address the temporal mismatch in multi-source data.
• The performances among the various advection correction methods are similar. Overall, the LK method performs slightly better than AD, followed by VET.

What are the implications of the main findings?

• When studying fast-moving convective storms, temporal mismatch among multi-source instruments should not be ignored.
• The choice of advection correction method has little impact on the performance of temporal matching.

Abstract

Multi-instrument synergistic observation is vital for studying cloud and precipitation physics. However, using the nearest scan time for matching inevitably introduces temporal mismatches. Here we employ three advection correction methods for temporal matching in weather radar and spaceborne radar observations: Lucas–Kanade (LK), Variational Echo Tracking (VET), and Anisotropic Diffusion (AD). These methods calculate the movement speed of the storms using optical flow methods, and then determine their positions based on the elapsed time between instruments. Next, we conducted a quantitative assessment of the performance of these three methods based on the consistency of storm morphology and rainfall rates. Our results demonstrate that all three advection correction methods effectively reduce the discrepancies in morphology and rainfall rate among multi-source data. Without correction, the Coincidence Rate (CR) and Structural Similarity (SSIM) were 30.96% and 0.689 in the US and 29.44% and 0.670 in China, respectively. In comparison, applying the LK, VET, and AD methods increased those indices to 32.94%, 32.72%, 32.85% and 0.718, 0.715, 0.716 in the US, and 31.34%, 31.17%, 31.24% and 0.696, 0.694, 0.693 in China, respectively. The rainfall rate inconsistencies were also effectively reduced after advection correction. The performances among the three methods were similar. Overall, the LK method performed slightly better than AD, followed by VET.

1. Introduction

Weather radar observations play a crucial role in meteorological monitoring and early warning, due to their spatially continuous observation of clouds and rain, especially in plateau and mountainous regions where rain gauges are sparse. Starting from the 1990s, the US gradually upgraded its operational weather radars from single-polarization to dual-polarization. Dual-polarization radars enable one to infer the size, shape, orientation, and composition of cloud and precipitation particles [1,2], significantly improving the accuracy of hydrometeor classification [3,4,5] and Quantitative Precipitation Estimation (QPE) [6,7]. The entire Next Generation Weather Radar (NEXRAD) network had completed the dual-polarization upgrade by the early 2000s, covering the continental United States, Alaska, Hawaii, and several overseas regions, comprising over 150 dual-polarization weather radars. In comparison, China started its dual-polarization upgrade later, roughly beginning in the 2010s. In recent years, about 100 China New Generation Weather Radars (CINRADs) have been upgraded, covering the major heavy rainfall areas in eastern, southern, and central China, with upgrades for the remaining radars underway. CINRAD and NEXRAD have become the two largest S-band dual-polarization weather radar networks in the world.

Despite their good performance in near-surface rainfall estimation [8,9,10], dual-polarization weather radars still have some inherent shortcomings, such as data voids below the lowest elevation angle and above the highest elevation angle [11]. Furthermore, the vertical resolution decreases sharply at long distances, due to the limitation of the number of elevation angles [11]. Therefore, discontinuities may appear when there is strong convection or a cloud top. To alleviate the limited elevation angles of ground-based radars, a multi-instrument synergy approach may be a viable solution, specifically by integrating spaceborne payloads or complementary ground-based observations to fill the gap in the three-dimensional (3D) observations [11].

The Global Precipitation Measurement (GPM) is a mission that provides global observations of rain and snow by constructing a satellite constellation. Its core satellite was launched on 27 February 2014, and the Dual-Frequency Precipitation Radar (DPR) onboard has already provided over ten years of 3D precipitation products [12]. Their vertical resolution can reach 250–500 m, enabling detailed profiling for rainfall storm studies [13,14]. By combining the vertical observations from the GPM DPR, the insufficient elevation angles of weather radars may be effectively mitigated.

However, since the scanning time of weather radars is 3 to 10 min, selecting any scan close to the GPM observation time will inevitably result in a certain time mismatch. In previous studies involving the matching of spaceborne and ground-based instruments, such as calibrating ground weather radars using spaceborne radar data [15,16] or conducting consistency analysis of rainfall rates between the two radar estimates [17,18], the closest weather radar scan in time is typically utilized. Consequently, the temporal mismatch between observations is often overlooked, potentially introducing significant spatiotemporal discrepancies, especially for fast-moving convective storms.

Advection correction may offer a unique perspective for addressing temporal mismatches [19]. This method assumes a constant storm movement speed between two scans and determines the storm’s position at the GPM scan by calculating the storm’s displacement. It is widely employed in storm tracking, precipitation nowcasting, or spatial smoothing within Quantitative Precipitation Estimation (QPE). However, its potential in temporal matching for rapidly moving convective storms remains unclear.

In this study, we temporally collocated long-term near-surface heavy rainfall observations from CINRAD and NEXRAD with GPM DPR. It aims to quantitatively evaluate the performance of different advection correction algorithms in temporal matching between ground-based and spaceborne radars, with consideration of the morphological consistency of rainfall storms across the two data sources. Then, we conduct a consistency analysis between weather radar rainfall estimates derived from different advection correction algorithms and spaceborne radar retrieval results. By comprehensively considering the storm morphology and rainfall rate matching, we determined the applicable conditions of different advection correction algorithms.

2. Data and Methods

2.1. Polarimetric Rainfall Estimation

The parameters of the typical dual-polarization CINRAD and NEXRAD are listed in

Table 1. The basic information of typical dual-polarization CINRADs and NEXRADs.

Radar Network	CINRAD	NEXRAD
Band	S-band	S-band
Peak power	650 to 750 kW	650 to 750 kW
Range resolution	250 m	250 m
Azimuth	0 to 360 deg	0 to 360 deg
Azimuth resolution	1 deg	1 deg
Elevation	Typical 0.5 to 19.5 deg	Typical 0.5 to 19.5 deg
Positioning error	±0.05 deg	±0.22 deg
Time resolution	6 min	3 to 10 min

. The two networks have similar configurations, e.g., the S-band with a frequency range of 2700–3000 MHz, a range resolution of 0.25 km and an azimuthal resolution of 1°. The NEXRAD network employs a more diverse range of dual-polarization radar models, resulting in greater variability in temporal resolution, ranging from 3 to 10 min. In comparison, dual-polarization CINRAD typically has a temporal resolution of around 6 min.

Compared to traditional single-polarization radars, dual-polarization ones can transmit and receive horizontally and vertically polarized waves. By analyzing the differences between these two polarizations, this upgrade introduces new polarimetric radar variables, including differential reflectivity ($Z_{DR}$), co-polar cross-correlation coefficient (${\rho }_{hv}$), differential phase (${\mathsf{\Phi }}_{DP}$), and its derivative specific differential ($K_{DP}$) [20,21]. Current studies have shown that $K_{DP}$-based QPE offers distinct advantages over the equivalent radar reflectivity factor ($Z_e$-based) approaches during high-rain-rate events, due to minimal sensitivity with drop size distributions (DSDs) and immunity to radar miscalibration, signal attenuation or partial beam blockage [22,23]. However, in weaker rainfall events (<10 mm·h⁻¹), $K_{DP}$ also exhibits a dependency on DSDs and is susceptible to clutter contamination, thus providing no distinct advantage over $Z_e$-based methods [23].

We first derive $K_{DP}$ from ${\mathsf{\Phi }}_{DP}$ measurements using a least-squares fitting method [24], which is commonly employed in operational systems. Cases where $Z_e$ exceeds 60 dBZ are excluded to mitigate potential hail contamination [25,26]. The $K_{DP}$ and other radar variables are interpolated to a 1 km resolution using PY-ART [27]. To mitigate clutter and ensure near-surface observations, we utilize $K_{DP}$ estimates from multiple elevation angles. Specifically, the elevation angle from the 2.2°–2.7° is used within 30 km of the radar site, the one from the 1.2°–1.7° is used for distances between 30 and 50 km, and the one from the lowest elevation angle (~0.5°) is used for distances between 50 and 200 km. Then, surface rainfall rate (R) was estimated with $R=51K_{DP}^{0.86}$ [9,28], a parameterization scheme proven reliable during an extreme rainfall event. We applied the same rainfall estimation algorithm to both CINRAD and NEXRAD to ensure a relatively constant potential estimation error. As weak $K_{DP}$ is susceptible to interference from clutter [6], this study focuses on rainfall events exceeding 20 mm·h⁻¹, conforming to the threshold of heavy hourly rainfall (20 mm·h⁻¹) defined by the China Meteorological Service Association (2019) [29]. We also checked radar beam blockage and clutter contamination to further guarantee data quality [22].

2.2. GPM DPR

Dual-Frequency Precipitation Radar (DPR) onboard the Global Precipitation Measurement mission Core Observatory (GPM-CO) satellite is the second spaceborne precipitation radar following Tropical Rainfall Measuring Mission (TRMM) Precipitation Radar (PR). Compared to TRMM PR, GPM DPR extends the observation range to more overland regions with its higher-inclination orbit (65°). Furthermore, GPM DPR operates at two distinct frequency channels: Ku-band (13.6 GHz) and Ka-band (35.5 GHz). Thanks to the improved sensitivity of the Ka-band channel, DPR expands its measurement capability for light rain and snowfall [12,30,31].

The basic information of GPM DPR is presented in

Table 2. The main parameters of GPM DPR.

Band	Ku-Band	Ka-Band
Frequency	13.6 GHz	35.547 GHz
Orbit height	407 km	407 km
Beam number	49 normal scans	25 matched and 24 interlaced scans
Spatial resolution	5 km	5 km
Swath width	245 km	120/245 km
Observable range	Surface to 20 km height	Surface to 20 km height
Range resolution	250 m	250/500 m
Sensitivity	0.5 mm·h−1	0.2 mm·h−1

. The orbit height of GPM-CO is 407 km, with a period of ~1.5 h. A whole horizontal DPR scan consists of 49 beams, with a distance of about 5 km between each beam. The beams are divided into 172 heights from surface to about 20 km, with a vertical resolution of 250∼500 m. The swath width of the Ku-band is about 245 km, and the width of the Ka-band is 120 km at launch and is adjusted to 245 km to match the Ku-band observation in May 2018 [32].

We recorded the retrieval variable “precipRateESurface”, or long name “Precipitation rate for the estimated surface”, to conduct a consistency analysis with ground-based radar polarimetric rainfall estimates, aiming to assess the applicability of different advection correction algorithms.

2.3. Advection Correction

Advection correction methods are widely used in the temporal matching process of ground-based radars [19]. They assume that the subjects in consecutive radar images move at a constant velocity, and most of the points in the previous image frame can be found in the subsequent frame. For radar scanning intervals of 3 to 10 min, the morphological changes in rainfall storms are generally insignificant, and their movement speed can be considered continuous; hence, these assumptions are typically satisfied.

The core idea of advection correction is to calculate the movement speed of subjects (V_LK, V_VET and V_AD in Figure 1) within the radar images by using optical flow methods [33], and then determine their corresponding positions based on the elapsed time (R(KDP) at T_DPR in Figure 1). Optical flow algorithms can generally be categorized into two types: dense optical flow and sparse optical flow [33]. Dense optical flow tracks every point in the image and optimizes by minimizing a cost function between the whole displaced image and the reference image; conversely, sparse optical flow tracks only a subset of feature points within the image, making a high computational efficiency.

Here, three advection correction methods compiled in Py-STEPS are considered:

(a) Lucas–Kanade method (LK) [34]: This algorithm assumes that the appearance of target image pixels remains unchanged as they move from frame to frame. It starts tracking from local image features and gradually extends to finer details.

(b) Variational Echo Tracking (VET) [35,36]: A McGill Algorithm for Prediction by Lagrangian Extrapolation (MAPLE) described in Germann and Zawadzki (2002) [35]. This method is essentially a global optimization routine designed to minimize the cost function between the two images.

(c) Anisotropic Diffusion method (AD) [37]: An optical flow technique which employs the notion of inconsistency during the solution of the optical flow equations, proposed by Proesmans et al. (1994) [37].

2.4. Evaluation Methodology

The evaluation of the advection correction algorithm is divided into two parts. First, the morphological consistency of rainfall storms is retrieved from ground-based and spaceborne radars. For spatial matching, the spaceborne radar data and weather radar rainfall images are interpolated into a 5 km equal-latitude/longitude grid using the nearest-neighbor interpolation. Subsequently, two morphological features are considered:

(1) Rainfall Center Coincidence Rate (CR): The rainfall images from weather radar and GPM DPR are divided into relatively weaker (RW) and relatively stronger (RS) portions based on a threshold of 10 mm·h⁻¹. CR is calculated by the following equation:

$$CR={\displaystyle \frac{N\left(R{S}_{KDP\&DPR}\right)}{N\left(R{S}_{KDPorDPR}\right)}}$$

(1)

Here, $R{S}_{KDP\&DPR}$ is defined as RS type at identical locations for both weather radar and GPM DPR, while $R{S}_{KDPorDPR}$ is RS type for weather radar or GPM DPR. “N” is the total number of ${RS}_{KDP\&DPR}$ or $R{S}_{KDPorDPR}$ pixels. A larger CR value means a better consistency.

(2) Structural Similarity Index Measure (SSIM) [38]: A Human Visual System (HVS)-based image quality assessment method used to measure the local similarity between two images in terms of luminance, contrast, and structural information. The pixel values of the rainfall images are clipped to the range of 0–255 to conform to standard grayscale image specifications. Values exceeding 255 mm·h⁻¹ are capped at 255 mm·h⁻¹. Then, the SSIM index is calculated by the following function:

$$\mathrm{SSIM}\left(x,y\right)={\displaystyle \frac{\left(2{\mu }_x{\mu }_y+C_1\right)\left(2{\sigma }_{xy}+C_2\right)}{\left({\mu }_x^2+{\mu }_y^2+C_1\right)\left({\sigma }_x^2+{\sigma }_y^2+C_2\right)}}$$

(2)

$${\sigma }_{xy}={\displaystyle \frac{1}{N}}\sum _{i=1}^N\left(x_i-{\mu }_x\right)\left(y_i-{\mu }_y\right)$$

(3)

${\mu }_x$ and ${\mu }_y$ denote the local means of images $x$ and $y$ within a sliding window. ${\sigma }_x^2$ and ${\sigma }_y^2$ are the corresponding variances; ${\sigma }_{xy}$ is the covariance between $x$ and $y$. $C_1$ and $C_2$ are small positive constants introduced to stabilize the division when the denominator is close to zero. The overall SSIM score for an entire image is obtained by averaging the SSIM values computed over all local windows, with an index range from $\left[-1,1\right]$. A larger SSIM value means a better consistency.

In addition to storm morphology consistency, the rainfall rate matching between the two databases was also evaluated. To minimize the impact of interpolation, the coarse-resolution GPM data were left unaltered, while the weather radar rainfall maps were interpolated into the GPM DPR grid using the inverse distance weighting of the four nearest points. Subsequently, a scatter plot between the two datasets was generated, and the Correlation Coefficient (CC), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE) were calculated as metrics to assess the consistency of rainfall rates:

$$\mathrm{CC}\left(x,y\right)={\displaystyle \frac{{\sum }_{i=1}^N\left(x_i-\overline{x}\right)\left(y_i-\overline{y}\right)}{\sqrt{{\sum }_{i=1}^N{\left(x_i-\overline{x}\right)}^2}\sqrt{{\sum }_{i=1}^N{\left(y_i-\overline{y}\right)}^2}}}$$

(4)

$$\mathrm{RMSE}\left(x,y\right)=\sqrt{{\displaystyle \frac{1}{N}}\sum _{i=1}^N{\left(x_i-y_i\right)}^2}$$

(5)

$$\mathrm{MAE}\left(x,y\right)={\displaystyle \frac{1}{N}}\sum _{i=1}^N\left|x_i-y_i\right|$$

(6)

$x_i,$ $y_i$ are corresponding data points and $\overline{x}$, $\overline{y}$ are the mean values of $x$ and $y$, respectively. $N$ is the total number of points in the rainfall image.

3. Results

3.1. Advection Correction for Temporal Match

We based our research on 6-year (2019–2024) CINRAD, 11-year (2014–2024) NEXRAD observations and 11-year record (2014–2024) of GPM DPR L2A V07. GPM DPR rainfall retrievals were used to roughly locate the heavy rainfall events, with a threshold of 20 mm·h⁻¹ (R(DPR) > 20 mm·h⁻¹). Then, CINRAD and NEXRAD observations were used to construct surface rainfall maps. With a weather radar rainfall estimate exceeding 20 mm·h⁻¹, we ultimately obtained extreme rainfall datasets of 901 and 6087 events in China and the US, respectively. The data flowchart can be found in Figure A1. A typical temporal matching process between spaceborne and ground-based radars is shown in Figure 2.

On 21 July 2023, an extreme rainfall event hit Ozark County, Missouri, a sparsely populated and scenic region in central United States. This event was captured by the GPM DPR at 14:53:15 UTC (09:53:15 CDT), which reported an instantaneous rainfall intensity exceeding 200 mm·h⁻¹ (Figure 2a₃). At the same time, approximately 84 km from the rainfall center, a WSR-88D dual-polarization weather radar (designated KSGF) also detected the event. The two closest radar scans occurred at 14:50:34 UTC (Figure 2a₁) and 14:55:55 UTC (Figure 2a₂), respectively, with maximum rainfall rates of 148.9 mm·h⁻¹ and 171.3 mm·h⁻¹.

Then, three optical flow methods were employed to these two scans to match DPR observations: the Lucas–Kanade method (LK), Variational Echo Tracking (VET), and the Anisotropic Diffusion method (AD). The storm motions estimated by these three methods show good consistency in direction, all moving roughly from west to east (Figure 2b₁–b₃ arrows). However, they exhibit distinct characteristics in magnitude (Figure 2b₁–b₃). The LK algorithm focuses more on the overall motion of the storm, resulting in relatively uniform speeds across different parts in the storm, mostly ranging between 50 and 60 km·h⁻¹ (Figure 2b₁). VET adopts a dense optical flow algorithm, calculating movement speed through global optimization, which produces a comparatively smooth speed field (Figure 2b₂). The maximum speed estimated by VET is located in the southern part of the rainfall storm, ranging from 80 to 100·km h⁻¹. In comparison, AD is capable of better preserving motion details and can distinguish speed variations at smaller local scales (e.g., the areas near 36.4° N, 93.3° W and 36.1° N, 93.4° W). The AD-calculated high speed is also situated in the southern area of the rainfall storm, mostly between 70 and 80 km·h⁻¹, with some places exceeding 80 km·h⁻¹.

Based on the calculated storm movement speed and the temporal interval between the ground and spaceborne radar observations, R(KDP) at the GPM DPR scan time can be derived by advecting the storm (Figure 2c₁–c₃). Although the estimated storm motions by the LK, VET, and ADF exhibit distinct characteristics, which may lead to positional shifts in the rainfall storm, the patterns of high rainfall rates still show good consistency between the three methods, with max rainfall intensities of 152.6, 149.7 and 150.2 mm·h⁻¹, respectively. In-depth quantitative analysis can be found in the following sections. Meanwhile, all three methods introduce certain spurious low-rain-rate areas (gray shaded areas in Figure 2c₁–c₃), so that studies utilizing advection correction for temporal matching should primarily focus on heavy rainfall events.

Given that a storm movement speed at 80 km·h⁻¹ can travel 7.3 km during the 5.5 min scan cycle of a ground-based radar, this distance (7.3 km) exceeds 45% of the GPM DPR grid resolution (5 km), highlighting the non-negligible effects of temporal mismatch in fast-moving storm research. Those advection correction approaches may also be applied to other spatially continuous storm observation means, such as spaceborne imagers [39,40].

3.2. Morphological Assessment of Rainfall Storms

The consistency of storm morphologies from spaceborne and ground-based radar perspectives can reflect whether advection correction methods mitigate the time mismatch. Here, we divide the experiments into four groups: one without advection correction and three using the LK, VET, and AD advection correction methods to construct surface rainfall images. Specifically, the “without advection correction” group refers to using the closest weather radar scan (CWRS) in time to the spaceborne radar observation. The results of these four experiments are shown in Figure 3.

When advection correction is not applied, both the CR and SSIM metrics are consistently inferior to those obtained after temporal correction (Figure 3), a trend observed in both China and the US. Specifically, the CWRS method yields CR and SSIM values of 30.96% and 0.689 in the US and 29.44% and 0.670 in China, respectively. In comparison, applying the LK, VET, and AD methods increased those indices to 32.94%, 32.72%, 32.85% and 0.718, 0.715, 0.716 in the US, and 31.34%, 31.17%, 31.24% and 0.696, 0.694, 0.693 in China, respectively. This demonstrates the necessity of employing an advection correction method for temporal matching. On the other hand, the choice of method has a relatively minor impact on morphological consistency, with LK performing slightly better than AD, followed by VET. Interestingly, both CR and SSIM values are lower in China than those in the US, revealing a geographical dependence in the consistency between spaceborne and ground-based radar observations.

To further investigate the impact of varying rainfall intensities on the three advection correction methods, we categorized four storm types based on the maximum R(KDP) derived from the AD method: 20–30 mm·h⁻¹, 30–50 mm·h⁻¹, 50–80 mm·h⁻¹, and >80 mm·h⁻¹. As shown in Figure 4, as the maximum rainfall intensity increases from 20–30 mm·h⁻¹ to >80 mm·h⁻¹, the SSIM index of the CWRS method drops from approximately 0.74 to 0.63 in China and from 0.73 to 0.65 in the US, while the three advection correction methods decrease from roughly 0.76 to 0.65 in China and from 0.75 to 0.68 in the US. Overall, the differences among the three advection correction methods are minimal across varying rainfall intensities, and all outperform the “closest weather radar scan” (CWRS) approach, which demonstrates the stability under high rainfall rate conditions.

Considering the idealized assumptions inherent in advection correction, such as storms moving at a constant speed between consecutive radar scans and minimal physical variability in rain rates over 3–7 min timescales, we here evaluated the morphological consistency across different radar scan cycles. The results show that the performance of all three advection correction methods degrades as the time interval increases (Figure 5). Specifically, when the radar scan cycle extends from 200 s to 420 s, the SSIM index decreases from approximately 0.779 to 0.688. Therefore, storm morphology undergoes significant changes over time, causing the idealized assumptions of advection correction to become less valid. For longer time intervals (e.g., >10 min), greater caution is required when applying advection correction for temporal matching.

3.3. Consistency Between Polarimetric Rainfall Estimates and DPR Retrievals

Having established that advection correction can improve the consistency of storm morphology across multi-source observations, we then evaluate the consistency of rainfall rate measurements between spaceborne and ground-based radars.

Consistent with the morphological results, using only the closest weather radar scan in time without advection correction yields the poorest overall consistency between the two observations (

Table 3. The rainfall rate consistency for temporal matching methods.

Region	Index	CWRS	LK	VET	AD
The US	CC	0.243	0.251	0.249	0.250
	RMSE	35.71	35.52	35.69	35.51
	MAE	18.79	18.31	18.42	18.34
China	CC	0.296	0.309	0.303	0.307
	RMSE	39.51	39.45	39.68	39.38
	MAE	21.45	21.22	21.36	21.22

). Specifically, the Correlation Coefficients (CCs) are 0.243 in the US and 0.296 in China. In contrast, after applying the LK, VET, and AD advection correction methods, the CC values improved to 0.251, 0.249, and 0.250 in the U.S. and to 0.309, 0.303, and 0.307 in China, respectively. The consistency indices derived from the three advection correction methods are generally similar. In terms of the correlation coefficient, LK performed slightly better than AD, followed by VET. On the other hand, the AD method demonstrates superior performance regarding the RMSE metric, achieving the smallest RMSE values.

Similarly, after classifying storms into four types based on the maximum R(KDP), the variation in MAE with rainfall intensity is illustrated in Figure 6. In the weaker types (20–30 mm·h⁻¹), the MAE of the three advection correction methods is similar to that of CWRS. However, when the rainfall rate exceeds 80 mm/h, the MAE of the advection correction methods is significantly lower than that of CWRS, highlighting the necessity of advection correction for temporal matching in high-rain-rate events. Furthermore, compared to the CWRS method, the three advection correction methods effectively mitigated the issue where R(DPR) was lower than R(KDP) across different rainfall intensities (

Table A1. The mean bias between R(DPR) and R(KDP). A negative value indicates that R(DPR) is less than R(KDP).

Region	Rain Rate	CWRS	LK	VET	AD
China	20–30 mm/h	−0.073	0.431	0.046	0.474
	30–50 mm/h	−6.155	−4.754	−4.506	−4.960
	50–80 mm/h	−19.297	−14.198	−14.772	−14.881
	>80 mm/h	−43.231	−37.262	−39.648	−36.374
The US	20–30 mm/h	−5.266	−3.412	−3.562	−3.680
	30–50 mm/h	−11.648	−10.424	−10.408	−10.490
	50–80 mm/h	−27.204	−25.055	−25.258	−25.202
	>80 mm/h	−51.391	−44.970	−45.865	−45.445

), resulting in a reduction in the absolute bias values. However, in the 20–30 mm·h⁻¹ bin over China, the differences among the various methods were marginal, which may be attributed to the relatively slower storm motion speeds.

The dependency of advection correction performance on the weather radar scan cycle is also reflected in the consistency of rainfall rates (Figure 7). As the scanning interval increases, the MAE between ground-based and spaceborne radar rainfall estimates increases accordingly. Specifically, when the radar scanning cycle is extended from 200 to 420 s, the MAE rises from approximately 16 to 20 mm·h⁻¹, which also indicates that the ideal assumptions underlying advection correction may become less reliable over longer time intervals.

4. Conclusions and Discussion

In this study, we quantitatively evaluated the effectiveness of advection correction in temporal matching between spaceborne and ground-based radar observations. The temporal offset between the two observations was corrected using LK, VET, and AD advection correction methods, based on the two weather radar scans nearest in time to the spaceborne observation. Those methods assume a constant storm motion speed between the two scans and determine the storm’s position at the GPM scan time by calculating its displacement. Simultaneously, a control experiment for comparison was conducted using only the single nearest weather radar scan without advection correction. With this dataset, we investigated the consistency in both storm morphology and rainfall rates. Our key findings are conceptualized as below:

1. Compared to using the temporally nearest radar scan, the three advection correction methods—LK, VET, and AD—all demonstrated certain advantages in the consistency of storm morphology and rainfall rates between spaceborne and ground-based radar observations, highlighting the importance of temporal matching for fast-moving convective rainfall storms.

2. The performances of the three advection correction methods are similar. Overall, LK is slightly better than AD, followed by VET. AD performs best in terms of RMSE for rainfall rate consistency.

3. The SSIM structural indices produced by the three advection correction methods deteriorate as the radar scan interval increases, accompanied by a rise in MAE for rain rate, which suggests that the ideal assumptions underlying advection correction may become unreliable over longer time intervals.

4. Compared to weaker rainfall events, advection correction can more effectively reduce the MAE of rainfall rates between ground-based and spaceborne radars in high-rain-rate events.

While advection correction consistently mitigates temporal mismatches in both China and the US, certain regional characteristics emerge upon comparison. The CR, SSIM, and CC values are generally lower in China than in the US, which may be related to geographical dependence, such as dataset size differences, geographical/climatological influences or potential radar system differences. However, these characteristics do not affect our main conclusions.

By properly addressing temporal matching, a synergistic analysis method combining ground-based radar rainfall estimates with the spaceborne radar three-dimensional observations will be established, which is expected to provide new insights into the three-dimensional structure and microphysical processes of extreme rainfall storms.

It should be acknowledged that this advection correction method for temporal matching similarly relies on the assumption of short-term uniform motion. Consequently, for data sources with longer time intervals (e.g., 15, 30 min, or more), these assumptions become less reliable, necessitating a cautious evaluation of the method’s performance. Furthermore, due to the long revisit time of the spaceborne radar, it is difficult to analyze the performance of advection correction across different storm life cycle stages, which requires more in-depth investigation in future studies.