The primary data utilized in this study were acquired from the FY-3G precipitation satellite. Operating in a unique low-inclination, non-sun-synchronous, inclined orbit at a nominal altitude of 407 km and an inclination of 50° ± 1°, the satellite covers the global midlatitude to low-latitude regions (Fig. a). It is equipped with two important precipitation observation payloads, which are the Ka/Ku dual-frequency Precipitation Measurement Radar (PMR) and the Microwave Radiance Imager-Rainfall Measurement (MWRI-RM). These instruments are crucial for monitoring catastrophic weather systems like typhoons and heavy rainfall.

Serving as the core payload of FY-3G, which provides 3D structural data on precipitation systems, the PMR is a single-polarization one-dimensional (1D) phased-array radar with a cross-track scanning mechanism, featuring a scan interval of 0.7° . It covers a ground swath width of 303 km and offers a horizontal resolution of 5 km at the nadir and a range resolution of 250 m with a sampling interval of 50 m (Fig. b, c).

As a significant payload of the FY-3G satellite, the MWRI-RM has 17 frequency points ranging from 10.65 to 183 GHz, which includes nine dual-polarized channels in the 10–89 GHz spectrum and 12 oxygen absorption sounding channels around 50 and 118 GHz, totaling 26 channels with spatial resolutions ranging from 5 to 25 km. Detailed channel information for the MWRI RM can be found in Table . Employing a conical scanning regime with imaging channels and sounding channels having incidence angles of 53° ± 1° and 50° ± 1°, respectively, MWRI-RM has an effective observation swath of 800 km, as illustrated in Fig. b. This instrument captures passive microwave radiation from the Earth's surface, yielding information on precipitation, atmospheric water vapor, cloud liquid content, path-integrated liquid water thickness, melting layer height and thickness, sea surface wind speed, and more .

In this study, we also utilized the dual-polarization radar from the China Next Generation Weather Radar (CINRAD/SA) network for comparison with the land precipitation reflectivity reconstruction results. The CINRAD/SA products have a radial distance resolution of 250 m and an azimuthal resolution of 1°. The Volume Coverage Pattern 21 (VCP21) scan mode was selected, which sweeps nine elevation angles of 0.5, 1.5, 2.4, 3.4, 4.3, 6.0, 9.9, 14.6, and 19.5° in 6 min . Based on the interpolation method used in , we transformed the radar data from spherical coordinates into a unified Cartesian coordinate system, creating a 3D grid. The radar data were then stitched and organized to generate the constant-altitude plan position indicator (CAPPI) data used in this study.

The Level 1 products of PMR and MWRI-RM can be downloaded from the Fengyun Satellite Data Center website (https://satellite.nsmc.org.cn/DataPortal/en/home/index.html, last access: 4 September 2025).

During the data preprocessing phase, the spatial positions of MWRI-RM data were first unified. Due to the angular discrepancy in the observational geometry of the MWRI-RM's two arrays of feedhorns (S1 and S2), we applied the nearest-neighbor method to adjust the spatial positioning of these two sets of data for consistency. Polarization difference (PD) is introduced as a key analytical variable, calculated from the difference between the brightness temperatures (Tbs) of vertically and horizontally polarized channels (PD = Tb_V - Tb_H). This parameter aids in revealing the shape and size distribution of hydrometeors, providing unique perspectives on precipitation processes . For the MWRI-RM, polarization differences were derived from nine channel pairs across frequencies of 10.65, 18.7, 23.8, 36.5, 50.30, 52.61, 53.24, 53.75, and 89 GHz. Each frequency channel offers distinctive insights into atmospheric conditions and hydrometeor characteristics.

In processing the PMR radar reflectivity profiles, special attention was given to mitigating the influence of ground clutter, with a PMR scan interval of 0.7°, which increases with incidence angle. Reflectivity profiles with local zenith angles less than 2° were selectively analyzed to minimize the interference from ground clutter, which affects altitudes up to approximately 1.1 km. Based on the 2° incidence angle criterion, five cross-track scan points per scan line meet this requirement. This approach is based on the observation that ground clutter's detrimental effects on radar measurements intensify with greater zenith angles, thus necessitating an angular threshold to uphold data quality . To balance the number of clear-sky and cloudy pixels, we only selected radar bins below 8 km, considering the cloud tops of most precipitation clouds are below this altitude . This helps model capture the essential data while minimizing noise from higher altitudes that might not be as relevant for precipitation analysis. The final reflectivity profile dimension encompasses 138 bins ranging from 1.1 to 8 km. Bins impacted by side-lobe and ground clutter were marked as invalid (NaN) to prevent their potential negative influence on the reconstruction outcome. Radar bins affected by noise were assigned a baseline value of 10 dBZ, supported by statistical evidence indicating that precipitation echoes typically exceed 12 dBZ. This selected baseline value helps distinguish between non-precipitating and precipitating conditions without significantly altering the overall distribution of reflectivity values, thereby maintaining the integrity of the reconstruction process. The filtering of noise, side-lobe clutter, and ground clutter was performed using the quality flag data (flagEcho) provided in Level 1 products of PMR.

The subsequent step in data preprocessing involved aligning the MWRI-RM data with PMR observations. This alignment was achieved by matching each MWRI-RM observation point to the nearest PMR footprint in space, ensuring a unique correspondence. Additionally, only matches with a scan time difference of less than 80 s were considered . Considering the different scanning modes of the PMR (cross-track scanning) and MWRI-RM (conical scanning), there are inherent differences in the spatial coverage of their observations (Fig. b). Although we performed nearest-neighbor matching for the ground footprints, the actual atmospheric volumes observed by each instrument are not identical. To best match the atmospheric conditions observed by both instruments, a patch extraction technique was employed, centered on the matched MWRI-RM footprint (Fig. c). This ensures that the atmospheric column observed by the PMR is always intersected by multiple MWRI-RM scanning beams. Consequently, the MWRI-RM Tb patch encompasses the atmospheric state information of the column observed by the PMR. This approach aligns the data from the two sensors more effectively, enhancing the model's ability to accurately interpret and utilize these observations. Our preliminary studies have indicated that a patch size of 15 × 15 offers the most optimal reconstruction performance . This choice allows for the inclusion of sufficient relevant atmospheric information and ensures effective observation matching. Consequently, this study exclusively utilizes patches of this specified size for analysis.

Given the natural scarcity of precipitation events compared to non-precipitation conditions, an undersampling of non-precipitation samples was performed. This procedure aimed to balance the quantity of precipitation and non-precipitation samples, providing a more equitable dataset for model training and evaluation. Here, the precipitation flag (flagPrecip) in the PMR data was used to distinguish between precipitation and non-precipitation samples. The operationally calibrated MWRI-RM Level 1 data and PMR Level 1 data have been available since 23 October 2023. The data from 23 October to 30 November 2023 were used for model training. After preprocessing, a total of 838 591 sample pairs were generated, including 421 090 non-precipitation samples and 417 501 precipitation samples, with 741 270 ocean samples, 27 233 land samples, and 70 088 coastal samples. The distinction between “ocean”, “land”, and “coastal” samples was made using the “LandSeaMask” variable provided in the Level 1 products of MWRI-RM. Coastal samples represent an intermediate scenario where the passive microwave brightness temperatures, derived from a 15 × 15 field of view (FOV) across multiple channels, encompass both oceanic and terrestrial footprints within the same retrieval domain. These samples occur when the passive microwave FOV includes both land and ocean components, regardless of whether the corresponding radar observations are positioned over land or ocean. As a result, certain pixels or subregions within the FOV exhibit mixed surface characteristics, leading to “coastal” conditions where parts of the passive channels are influenced by land. These sample pairs were divided into training and validation subsets, with 80 % of the samples allocated for training and the remaining 20 % reserved for validation. Additionally, data from 1 to 11 December 2023 were used for model testing, yielding a total of 288 892 samples, including 144 478 non-precipitation samples and 144 414 precipitation samples, with 274 429 ocean samples, 9123 land samples, and 24364 coastal samples. All inputs were standardized according to the following equation: 1x′=x-μσ, where x represents the original input value, and x′ is the standardized value. μ is the arithmetic mean, and σ is the standard deviation. This normalization is applied independently to each channel using the mean and standard deviation. Additionally, since the reflectivity profiles exhibited a skewed distribution, a logarithmic transformation was applied. This preprocessing step is essential for reducing the impact of outliers and ensuring that the model learns from a more uniform dataset, thereby enhancing the reliability and accuracy of the predictions.

In our investigation, we adopt a deep learning architecture that diverges from the mixed model approach reported by . Unlike their study, which explored the benefits of a combined CNN-MLP approach for integrating Tb patches from multiple channels and ancillary temperature profile data, our model emphasizes the potency of CNNs in discerning the vertical structure of atmospheric temperature directly from the advanced sounding channels, particularly those near the oxygen absorption bands. These sounding channels are sensitive to the temperature profile, thus potentially reducing the necessity for separate profile information processing.

To critically evaluate the influence of various channel configurations and feature inputs on the model's ability to reconstruct radar reflectivity, we orchestrated a series of controlled experiments. The baseline experiment (Ex14) excludes the oxygen absorption channels, using 14 input channels (out of the total 26, excluding 12 oxygen absorption channels) to assess the model's basic capability without temperature information. The full channel experiment (Ex26) incorporates all 26 input channels, including the oxygen absorption bands, providing additional temperature profiles to improve precipitation detection and the delineation of melting layers. Building on this, the polarization difference enhanced experiment (Ex35) adds Tb polarization difference data, increasing the input to 35 channels (26 channels plus 9 polarization difference sets), to enhance the model's ability to distinguish precipitation types and capture finer structural details. With a consistent architecture across experiments, this setup enables a rigorous evaluation of how extended MWRI-RM channel information influences the model's performance in precipitation reconstruction.

The detailed specifications of the CNN model configuration are delineated in Table . In Table , C denotes the number of input channels. For experiments Ex14, Ex26, and Ex35, C is 14, 26, and 35, respectively. The architecture of the CNN employs a consistent structure across experiments, and the term B.R.D represents a sequence of three layers: batch normalization, rectified linear unit (ReLU) activation layer, and dropout. For all experiments, dropout is applied with a probability of p=0.2. Additionally, a max pooling layer with a 2 × 2 kernel is applied after the first convolutional block, reducing the spatial dimensions of the feature maps from 15 × 15 to 7 × 7 while preserving the number of channels.

Model training was conducted using eight A800 GPUs, with a batch size of 512, 100 epochs, and the Adam optimizer. The learning rate was initialized at 1×10-3 and decayed using the InverseTimeDecay schedule. Early stopping was employed with a patience of 20 epochs to prevent overfitting, and the best weights were restored. The loss function used was the sum of squared errors (SSEs), defined as the sum of squared differences between true and predicted values. This training protocol was applied consistently across all experiments to ensure comparability and reproducibility of results. To account for the influence of random processes, three independent training runs were conducted for each model, each initialized with a different random seed. Figure  illustrates the training and validation loss curves for the best-performing run of Ex14, Ex26, and Ex35, selected based on the lowest average validation loss over the final 10 epochs. Shaded areas represent the inter-model spread (±1 standard deviation) across three random initializations for training and validation losses.

To quantify the reconstruction accuracy of the model, we employed three statistical indices: mean bias error (MBE), standard deviation (SD) of the predictive error, and root mean square error (RMSE). Each metric offers a distinct dimension of the model's performance.

The MBE is indicative of the average bias in the model's predictions, where a value of zero represents a perfect bias-free model. It is computed as follows: 2MBE=1n∑i=1n(Ri-Oi), where Ri is the reconstructed value, Oi is the observed value, and n is the number of data points. This metric reveals whether the model tends to overpredict or underpredict the observed values, providing insight into systematic deviations.

The SD of the predictive error measures the spread of these errors around the mean error, indicating the variability within the dataset: 3SD=1n-1∑i=1nRi-Oi-MBE2. Here, the term Ri-Oi-MBE represents the deviation of each individual reconstruction error from the MBE. The SD elucidates the volatility in reconstruction performance and identifies the degree to which individual reconstructions deviate from the average bias.

Finally, the RMSE assesses the magnitude of the error, penalizing larger discrepancies more severely, thus highlighting the model's precision: 4RMSE=1n∑i=1nRi-Oi2.

The RMSE is especially useful as it relates directly to the data's scale, offering an unambiguous interpretation of the model's predictive capability.

By utilizing these indices, we can deliver a rigorous and comprehensive evaluation of our model's predictive accuracy, providing a balanced overview of both central tendency and variability in the model's performance. This multifaceted approach ensures a robust assessment, crucial for validating the model's applicability to operational forecasting scenarios.

In this section, we conduct a comprehensive evaluation of the performance of three experiments – Ex14, Ex26, and Ex35 – using 288 892 test samples. Figure  illustrates the variations in MBE, RMSE, and SD across different height levels for four scenarios: oceanic precipitation, land precipitation, coastal precipitation, and non-precipitation under the conditions of the three experiments. The RMSE profiles in Fig.  are augmented with shaded areas representing ±1 standard deviation spread across three independent training runs, explicitly quantifying the variability induced by random initialization. Additionally, Table  provides detailed numerical values for evaluation metrics under different scenarios for each of the three experiments, including the RMSE spread range calculated from the training trials, offering a thorough quantitative analysis of model performance.

A general trend observed is the reduction of errors with increasing altitude. However, the errors at approximately 4–5 km, a range where the melting layer often occurs , remain significant. This highlights a crucial area for model refinement, particularly in adjusting to phase changes in precipitation. The comparison between oceanic and land precipitation scenarios reveals greater reconstruction errors over land. This discrepancy is largely attributed to the interference of land surface emissivity at low-frequency channel bands, which compromises the channels' ability to convey accurate precipitation information . Such limitations underscore the challenges faced by the model in land precipitation scenarios due to the reduced effectiveness of low-frequency channel observations in these environments. In coastal precipitation scenarios, errors are higher than those over pure ocean but lower than those over land, reflecting a gradual decline in reconstruction accuracy as land influence increases within the 15 × 15 channel domain.

With the integration of dual oxygen absorption sounding channels, there is a notable improvement in the accuracy of precipitation reflectivity reconstruction, especially under land and coastal precipitation scenarios, with RMSEs reduced by 5.43 % ± 1.56 % and 5.47 % ± 1.17 %. This enhancement is also evident near the melting layer, underscoring the sounding channels' sensitivity to atmospheric temperature structures. This feature is instrumental in refining the model's ability to accurately map precipitation structures by leveraging the temperature profiles provided by the oxygen absorption sounding channels. However, incorporating Tb polarization difference (Ex35) yields minimal additional improvement (0 %–1.6 % RMSE reduction), with RMSE curves largely overlapping Ex26's, and this reduction is comparable to the RMSE spread, indicating that Tb polarization difference has no significant effect. In non-precipitation conditions, the model shows minimal error, with RMSE values generally less than 1 dBZ. The slight difference between the background value and the minimal precipitation threshold (2 dBZ) reinforces the model's capability to accurately detect the absence of precipitation.

Despite these advancements, all models underestimate reflectivity, evidenced by negative mean bias errors (MBE). Ex35 shows the smallest underestimation, offering a slight advantage in mitigating bias. This systematic low bias suggests challenges in capturing intense precipitation, warranting further investigation.

In this part of the paper, we focus on evaluating the model's reconstruction capabilities during two extreme precipitation events: Typhoon Khanun in 2023 and the extreme rainfall in the Beijing–Tianjin–Hebei (BTH) region in North China from 29 to 31 July 2023 . These events caused significant impacts on human life and activities, providing challenging yet informative scenarios for testing the model's precision and robustness. The challenges arise from the extreme nature of these events, which involve complex and dynamic atmospheric conditions that introduce significant variability and nonlinear interactions. These factors make it difficult for the model to generalize, especially when such events exceed the range of conditions encountered during training. Although the data used for these case studies were obtained before the official release in October 2023, we were granted access to PMR observations from July 2023, including data from these two extreme events, which are valuable for evaluating the model's generalization ability in extreme conditions.

Figures  and provide a comparative analysis of the three-dimensional reflectivity structures of the two extreme precipitation events across three different experimental models. The figure contrasts the observed reflectivity values (panels a–c of Figs.  and ) with the reconstructed data obtained from three experiments: Ex14 (panels d–f), Ex26 (panels g–i), and Ex35 (panels j–l). Each set of panels is organized to include three key perspectives: the 3D reflectivity structure (left column), the horizontal distribution at 5 km altitude (middle column), and the vertical cross-section along points A and B (right column). The reconstructed results presented here represent the best-performing outcomes from each experiment to highlight the models' optimal capabilities in capturing extreme precipitation features.

As shown in Fig. , the 3D reflectivity profile from PMR data revealed decreasing reflectivity values with altitude, typical of frozen precipitation forms such as snow, graupel, and frozen drops commonly found in the cores of active thunderstorms. Ground clutter limited the observation range at the lowest altitudes, but the model, trained on reflectivity profiles with incidence angles less than 2°, reconstructed a more complete reflectivity profile, minimizing ground clutter effects. This demonstrates the model's capability to reconstruct reflectivity impacted by clutter, an essential feature for accurate storm characterization. For a detailed horizontal analysis of the precipitation system, we chose a reflectivity cross-section at 5 km above ground, where the influence of ground clutter was minimal, revealing a more comprehensive view of the precipitation system. PMR data showed the highest reflectivity values located mainly to the south and east of Khanun's center, with moderate to heavy rainfall surrounding the eye and a prominent rainband to the northwest. Additionally, a vertical cross-section along a line from point A to point B (both directly beneath the satellite track) was analyzed to assess the vertical structure of the precipitation system. In this radar cross-section, a prominent bright band (BB) was observed, characterized by a peak in reflectivity. This feature typically signifies the presence of a melting layer, as it occurs where falling snow begins to transition into rain, thereby enhancing the radar return signal .

Analysis of the reconstruction results from different experiments showed that while all models accurately reconstructed the overall structure of the precipitation system, there were noticeable differences in detail. The Ex14 experiment exhibited systematic underestimation of high reflectivity values, particularly in the typhoon eyewall region and the northwestern rainbands (26° N, 130° E) (Fig. d, e). Quantitative evaluation showed Ex14 had a mean bias error (MBE) of 0.81 and root mean square error (RMSE) of 4.91. The Ex26 experiment demonstrated improved performance with reduced biases (MBE = 0.61, RMSE = 4.87) yet maintained a tendency to underestimate reflectivity in the northwestern rainbands while overestimating values east of Khanun's center (23° N, 133° E). Notably, Ex35 achieved superior reconstruction fidelity for the northwestern rainbands, aligning closely with observations. However, it retained a slight reflectivity overestimation east of the typhoon center (23° N, 132° E), while maintaining the best quantitative metrics (MBE = 0.58, RMSE = 4.86). However, as with the Typhoon Khanun event, all experiments showed some discrepancies in the position of the melting layer, typically positioned lower than observed. This is likely due to the inherent ambiguity of the radiative signals, which affects the precision and vertical detail of the reconstructions.

To provide a more quantitative and objective assessment, Fig.  presents scatter density plots that compare reconstructed reflectivity against the observed values for Typhoon Khanun. These plots visualize the joint distributions, with the diagonal line indicating perfect agreement. All experiments exhibit robustness in moderate reflectivity but show systematically increasing underestimation at higher reflectivity levels (> 30 dBZ).

An extreme rainfall event occurred in the Beijing–Tianjin–Hebei (BTH) region in North China from 29 July to 1 August 2023, commonly referred to as the “07.2023” BTH extreme rainfall event. According to observations, the 72 h accumulated precipitation from 00:00 UTC on 29 July to 00:00 UTC on 1 August 2023 ranged from 350 to 550 mm across southwestern Beijing, central Hebei, and Tianjin, with some localities receiving over 750 mm . The FY3G satellite flew over the BTH region during this extreme precipitation event, as illustrated in Fig. . At 5 km altitude, PMR observations showed high reflectivity predominantly in the southern parts of Beijing (BJ, 39.5° N, 115.8° E) and central Hebei (HB, 39.1° N, 116.2° E), although the edge of the satellite track was affected by terrain, leading to missing reflectivity data in these mountainous regions. In the vertical structure analysis along a cross-section from point A to point B in southern Beijing, the bright band (BB) in this melting layer occurred around 5 km. Ex14 exhibited systematic underestimation of reflectivity values, particularly failing to capture high-reflectivity regions in southern Beijing and central Hebei. This experiment also showed notable deficiencies in reconstructing the melting layer signature, suggesting that the absence of temperature information from oxygen absorption channels critically impairs vertical precipitation characterization over land. This observation aligns with previous model evaluation findings. Quantitative analysis showed Ex14 had the largest biases, with a mean bias error (MBE) of 2.65 and root mean square error (RMSE) of 5.81. While Ex26 still generally underestimated reflectivity values, it demonstrated marked improvement in capturing key high-reflectivity features and produced a more realistic representation of melting layer structure compared to observations. This enhanced performance is reflected in its reduced errors (MBE = 1.72, RMSE = 5.23). The Ex35 configuration achieved the most accurate reconstruction, successfully retrieving high-reflectivity features in southern Beijing and producing the faithful vertical depiction of melting layer characteristics. This superior performance is quantified by the smallest errors among the three experiments (MBE = 1.36, RMSE = 5.14).

To further quantify model performance, we employed scatter density plots (Fig. ). In the BTH case, Ex14 exhibited the poorest agreement with observations, characterized by a wide scatter dispersion and a pronounced systematic offset below the diagonal. Ex26 showed marked improvement, with a more symmetric density distribution concentrated closer to the 1 : 1 line compared to Ex14. However, it retained a tendency to underestimate reflectivity values above 25 dBZ, particularly in high-intensity precipitation regimes. Ex35's scatter density distribution is more concentrated around the diagonal, with better alignment at higher reflectivity, indicating better overall correspondence between observed and reconstructed reflectivity.

However, a notable issue in the reconstruction is the occurrence of reflectivity values below the surface in high-altitude regions, particularly areas with elevations above 1.1 km. For instance, in Fig. 7c, which represents the observations, there are invalid values near points A and B due to the influence of mountainous terrain (elevations exceeding 1.1 km). However, in the reconstruction results, reflectivity values are still present below 1.1 km (Fig. f, i, l), highlighting this notable inconsistency. This issue highlights a limitation in the model's applicability to complex terrains and mountainous regions. The potential solutions for this issue, such as designing the loss function to account for surface constraints, incorporating terrain height as an input feature, and applying post-processing techniques, will be addressed in future work.

Based on the above evaluation, the Ex35 model demonstrates superior performance in reconstructing the horizontal distribution, vertical structure, and intensity of reflectivity for both oceanic and terrestrial precipitation. Here, we utilize the Ex35 model to reconstruct the reflectivity across the entire MWRI-RM swath to gain a more comprehensive view of precipitation systems. The spatial distributions of brightness temperatures of the oxygen absorption band channels for two cases are shown in Figs.  and , where we can clearly observe the impact of precipitation on the channel brightness temperature, with lower brightness temperature, especially in the 118 GHz channel, due to the strong effect of ice particle scattering in convective areas. Figure a and c show the reconstructed reflectivity of Typhoon Khanun across the entire MWRI-RM swath. Compared to the actual observations from PMR-Ku, we can observe a very complete three-dimensional structure of the typhoon. For instance, the well-formed spiral rainbands accompanied by intense reflectivity in the southwest of Typhoon Khanun's eyewall have been well reconstructed, a feature not evident in the PMR-Ku observations. Similarly, Fig. b and d present a more complete reflectivity distribution of a terrestrial precipitation system. In addition to the precipitation echoes observed by PMR-Ku over southern Beijing, central Hebei, and Tianjin (TJ, 39.2° N, 117.0° E), the MWRI-RM swath reveals precipitation echoes over northern Hebei, northern Shanxi, and parts of Inner Mongolia adjacent to Shanxi.

To validate these reconstructed reflectivities, we compare them against a composite of ground-based radar observations. Figure e shows a CAPPI product compiled from 99 CINRAD/SA radars in and around the study area. Within the regions covered by ground-based radars, the spatial patterns of reflectivity and the locations of major reflectivity maxima in the reconstructed field correspond closely to the observations of ground-based radars. For instance, the model accurately reproduces the strong echoes and their approximate positions in southern Beijing and central Hebei, reflecting areas of intense precipitation similar to those captured by the radars. This spatial alignment demonstrates that the reconstructed precipitation reflectivity distributions are consistent with ground-based radar measurements in terms of key precipitation features (e.g., high-reflectivity cores).

It is worth noting that ground-based radar coverage is limited in remote regions, such as northern Shanxi (SX, 39.8° N, 113.0° E) and Inner Mongolia (IM, 40.0° N, 112.0° E), where the radars are sparse or absent. In these areas, the reconstructed field extends beyond the radar coverage, suggesting precipitation where radars cannot confirm its presence. While this part of the reconstruction cannot be directly validated against ground-based observations, the demonstrated consistency within the radar-covered domain increases confidence that the satellite-based retrieval is capturing realistic precipitation distributions elsewhere. Hence, the Ex35 model's ability to align with radar-derived precipitation structures in well-observed areas, coupled with its broader swath coverage, indicates its potential to fill observational gaps and offer a more complete depiction of precipitation systems than ground-based radars alone.