Synergistic Fusion of GNSS-PWV and Radar for Precipitation Nowcasting: An AI-Empowered Spatio-Temporal Attention Network

Highlights

What are the main findings?

• High-precision Precipitable Water Vapor (PWV) derived from ground-based GNSS networks demonstrates a strong correlation (>0.97) with reanalysis data, providing valuable information on pre-convective moisture accumulation.
• The proposed STEA-Swin deep learning model, which fuses GNSS-PWV and radar data, improved the Critical Success Index (CSI) by 18.5% and the Probability of Detection (POD) by 21.5% for torrential rain events compared to single-source radar models.

What is the implication of the main finding?

• Integrating thermodynamic water vapor precursors with kinematic radar echoes decisively mitigates the “blurring effect” and intensity underestimation issues common in pure deep learning extrapolation models.
• This multi-modal AI framework significantly enhances short-term heavy precipitation nowcasting capabilities, providing a robust and highly accurate early warning approach for urban flash flood mitigation and disaster resilience.

Abstract

Extreme weather events exacerbated by global warming pose severe threats to urban safety, underscoring the urgent need for highly accurate precipitation nowcasting. Short-term local heavy precipitation remains a particular challenge for traditional forecasting due to its suddenness and high disaster potential. To address this, we propose a multi-modal fusion framework that integrates ground-based GNSS-derived Precipitable Water Vapor (GNSS-PWV) and ground-based Radar Composite Reflectivity (CR). While GNSS-PWV keenly captures pre-convective atmospheric water vapor accumulation, radar CR details the morphological distribution of hydrometeors. Specifically, we developed the Spatio-Temporal Enhanced Attention Swin U-Net (STEA-Swin) model to synergize these heterogeneous datasets over the Beijing–Tianjin–Hebei region. High-precision PWV was retrieved from 250 Continuously Operating Reference Stations (CORS) using the dual-frequency ionosphere-free Precise Point Positioning (PPP) method, achieving a strong correlation (>0.97) with ERA5 reanalysis data. Validated against measured data from the 2025 flood season, the STEA-Swin model achieved a Probability of Detection (POD) of 0.68 for torrential rain events at a +1 h forecast lead time. Notably, compared to single-source models, the Critical Success Index (CSI) and POD for torrential rain improved by 18.5% and 21.5%, respectively. These findings demonstrate that coupling deep learning with ground-based GNSS-derived atmospheric thermodynamic information can significantly enhance early warning capabilities, providing a promising technical approach for regional disaster prevention and climate resilience.

1. Introduction

Under the exacerbating influence of global climate change, the hydrological cycle has intensified significantly, leading to a dramatic escalation in the frequency and destructiveness of Extreme Precipitation Events (EPEs). The Beijing–Tianjin–Hebei (BTH) urban agglomeration, a pivotal socio-economic hub in Northeast Asia, faces particularly severe challenges due to its unique geographical configuration. The interaction between the moisture-laden East Asian summer monsoon and the steep topographic gradients of the Yanshan and Taihang Mountains creates a volatile atmospheric environment [1]. This coupling frequently triggers short-duration, high-intensity convective storms characterized by rapid evolution, non-linear dynamics, and extreme spatiotemporal heterogeneity [2]. Furthermore, rapid urbanization in the BTH region has exacerbated the urban heat island and rain island effects, altering local microphysics and making storm tracks even more unpredictable [3]. Recent catastrophic flooding events have underscored that accurate precipitation nowcasting is not merely a meteorological pursuit but a critical necessity for urban resilience, flash flood mitigation, and emergency response optimization [4]. In addition, the BTH region is characterized by the coexistence of mountain–plain transitional terrain, dense urban infrastructure, and frequent summer extreme rainfall events, making it a representative study area for investigating short-term heavy precipitation and its related disaster risks.

Current operational nowcasting systems primarily rely on two pillars: Numerical Weather Prediction (NWP) and radar echo extrapolation. While NWP models based on fluid dynamics equations are authoritative for medium-range forecasts, they inherently suffer from the spin-up effect. Due to imperfect initial conditions and parameterization schemes, NWP models often struggle to effectively initialize convective cells within the first hour of the forecast window, leading to a blind spot in immediate warning systems [5]. Conversely, radar-based extrapolation methods treat precipitation strictly as a computer vision problem. Although computationally efficient, these kinematic approaches fundamentally ignore atmospheric thermodynamics. They track the movement of existing hydrometeors based on the assumption of Lagrangian persistence but fail to predict the initiation or dissipation of storm cells driven by unseen water vapor instability [6]. Consequently, the prediction skill of pure extrapolation degrades rapidly after 30–45 min, as confirmed by benchmarking on large-scale datasets and operational evaluations [7].

To bridge the gap between kinematic observation and dynamic evolution, the integration of thermodynamic precursors is essential. Precipitation is physically driven by the phase change of atmospheric water vapor. The Global Navigation Satellite System (GNSS) has emerged as a robust, all-weather technique for sensing atmospheric moisture conditions [8]. High-precision Precipitable Water Vapor (PWV) derived from GNSS observations provides valuable information on moisture accumulation prior to precipitation [9]. Recent studies have demonstrated that PWV variations are closely associated with convective initiation and heavy rainfall processes [10]. However, GNSS observations are typically sparse and represent vertically integrated atmospheric conditions, whereas radar provides high-resolution spatial structures of precipitation systems [11]. This discrepancy between sparse thermodynamic observations and dense radar fields presents a major challenge for short-term precipitation forecasting.

From an observational perspective, different monitoring techniques provide complementary information. Satellite observations can capture large-scale cloud-top evolution but mainly reflect upper-atmospheric properties [12]. Weather radar remains the primary tool for observing precipitation morphology and storm evolution, although it is affected by terrain, attenuation, and beam blockage [13]. In contrast, GNSS meteorology provides continuous monitoring of atmospheric water vapor and enables the detection of pre-convective environmental changes that are not directly observable from radar echoes [14]. Previous studies have shown that PWV anomalies and temporal gradients can serve as effective indicators for heavy precipitation events [15].

With the rapid development of artificial intelligence, precipitation nowcasting methods have evolved from traditional extrapolation approaches to deep learning frameworks [16]. Convolutional Neural Networks (CNNs) have been widely applied due to their capability in extracting spatial features, but their limited receptive fields restrict the modeling of long-range dependencies [17]. Recurrent neural networks such as ConvLSTM have been introduced to capture spatiotemporal evolution, yet they often suffer from training instability and limited parallel efficiency [18]. Encoder–decoder architectures such as U-Net have demonstrated strong capability in multi-scale feature extraction and have become widely used in precipitation nowcasting [19]. However, standard convolution-based approaches still rely on local operations, which may constrain their ability to represent long-distance interactions in organized convective systems [20].

More recently, attention-based architectures have provided new opportunities for precipitation nowcasting. Transformer-based models enable more effective modeling of long-range dependencies and complex spatiotemporal interactions [21]. In the broader context of weather forecasting, deep learning frameworks have shown strong potential in capturing nonlinear atmospheric dynamics [22]. Furthermore, recent developments in global weather prediction using deep learning highlight the capability of data-driven approaches in representing complex atmospheric processes [23]. These advances indicate that combining multi-scale feature learning with global dependency modeling is essential for improving nowcasting performance. However, most existing approaches still rely primarily on radar data and treat precipitation forecasting as a sequence prediction problem, while insufficiently incorporating thermodynamic constraints related to water vapor evolution [24].

From the perspective of multi-source fusion, recent studies have explored the integration of GNSS-PWV and radar observations, demonstrating that thermodynamic precursor information can enhance prediction capability beyond radar-only systems [25]. Nevertheless, the mismatch between sparse GNSS observations and dense radar grids introduces challenges in data representation and fusion, which may lead to uncertainties in model performance [26]. Therefore, constructing an effective framework that can jointly exploit the temporal evolution of GNSS-derived moisture information and the spatial morphology of radar echoes remains a key challenge in short-term precipitation nowcasting [27].

Accordingly, this study aims to investigate whether the integration of GNSS-derived atmospheric moisture information can improve the forecasting performance of short-term local heavy precipitation, particularly during the initiation and rapid development stages of convective systems in the Beijing–Tianjin–Hebei region. To address this problem, a Spatio-Temporal Enhanced Attention U-Net model is proposed, and a multi-source collaborative forecasting framework based on GNSS PWV and radar composite reflectivity (CR) is developed.

To solve the aforementioned problems, this paper proposes a Spatio-Temporal Enhanced Attention U-Net model, constructing a multi-source collaborative forecast framework of GNSS PWV and Radar CR to carry out research on short-term local heavy precipitation forecasting in the Beijing–Tianjin–Hebei region.

The main contributions of this paper include the following:

• High-Fidelity Ground-Based GNSS Atmospheric Data Retrieval: We constructed a high-precision meteorological dataset for the Beijing–Tianjin–Hebei region by retrieving GNSS PWV (30 min resolution, <3 mm accuracy) using the dual-frequency ionosphere-free PPP method. This provides a reliable ground-based GNSS-derived characterization of atmospheric thermodynamic precursors, serving as a critical input for weather resilience analysis.
• AI-Empowered Multi-Modal Data Fusion: We developed the STEA-Swin model, an AI-empowered framework that synergizes the 1D temporal evolution of GNSS-PWV data with 2D spatial radar echoes. By embedding Swin Transformer blocks and spatio-temporal attention mechanisms, the model effectively mines the nonlinear coupling between GNSS water vapor precursors and precipitation dynamics.
• Optimization for Extreme Weather Detection: To support disaster early warning, we introduced a composite Edge-Aware loss function. This optimization balances numerical precision with geometric structural similarity, successfully mitigating the “blurring effect” in deep learning nowcasting and ensuring high detection rates for high-impact torrential rain events, thereby supporting robust decision-making for urban flood prevention.

The remainder of this paper is organized as follows: Section 2 details the research data, GNSS PWV inversion method, and the design ideas of the Spatio-Temporal Enhanced Attention U-Net model; Section 3 presents the experimental design; Section 4 analyzes the results; Section 5 summarizes the research conclusions and looks forward to future work directions.

2. Materials and Methods

2.1. Retrieval of GNSS PWV Based on Dual-Frequency Ionosphere-Free Combination PPP

Dual-frequency GNSS observation data (sampling rate 30 s) from 250 CMONC CORS stations in the BTH region were used, spanning from May to August 2025. The observation data include carrier phase and pseudorange observations from GPS satellites, as well as station metadata (e.g., station coordinates, antenna types). Before retrieval calculation, strict quality control was performed on the raw data, including cycle slip detection and repair based on the TurboEdit algorithm and outlier removal. IGS-provided Antenna Phase Center Offset (PCO) and Variation (PCV) files were used for refined antenna phase center correction to ensure the purity and high precision of input data [28].

Unlike traditional differential positioning techniques or single observation equation solutions, this study adopted the more flexible and high-precision un-differenced un-combined Precise Point Positioning (PPP) technology and constructed the GNSS observation model as a dynamic state space estimation problem based on Extended Kalman Filtering. For station r and satellite S, to eliminate the first-order term influence of ionospheric delay (which typically accounts for over 99% of total ionospheric error), the linearized dual-frequency ionosphere-free combination (IF) observation equation is expressed in the following compact matrix form [29]:

$$y_{IF}=Ax+ϵ$$

(1)

where $y_{IF}$ is the observation vector composed of pseudorange $P_{IF}$ and carrier phase $L_{IF}$ after ionosphere-free combination, effectively suppressing signal interference during active ionospheric periods; $A$ is the design matrix, containing the geometric direction cosines of the satellite relative to the station and the mapping function values of the tropospheric wet delay; $ϵ$ is the measurement noise, with its stochastic model constructed using an elevation-angle-dependent weighting strategy. The state vector $x$ to be estimated is defined as

$$x={\left[r_r^T,c\mathrm{\delta }{t}_r,ZTD,N_{IF}^T\right]}^T$$

(2)

In this model, $r_r$ is the station coordinate correction used to absorb minute displacements caused by tidal loading or tectonic motion; $c\delta t_r$ is the receiver clock error, estimated as a white noise process; $ZTD$ is the core meteorological parameter to be retrieved—Zenith Total Delay, modeled as a random walk process; and $N_{IF}$ is the floating ambiguity after ionosphere-free combination. To achieve millimeter-level solution precision, IGS precise orbit and clock products were introduced to correct satellite-side orbit and clock errors, and the GMF projection function was used to map slant path delays to the zenith direction, solving the issue of utilizing low elevation angle observation data.

After obtaining high-precision Zenith Total Delay ($ZTD$) through the PPP filter, it needs to be separated into Zenith Hydrostatic Delay (ZHD) and Zenith Wet Delay (ZWD). ZHD mainly depends on station air pressure and is calculated using the classic Saastamoinen model [30]:

$$ZHD={\displaystyle \frac{0.0022768\cdot P_s}{1-0.00266\cos \left(2\mathrm{\phi }\right)-0.00028H}}$$

(3)

where $P_s$ is the station surface pressure, $\phi$ is the station latitude, and $H$ is the station geodetic height. Since the hydrostatic delay has high spatiotemporal stability, the wet delay component, which changes drastically and is closely related to precipitation, can be separated by subtracting this component from the total delay:

$$\mathrm{ZWD}=\mathrm{ZTD}-\mathrm{ZHD}$$

(4)

Finally, the inversion of Precipitable Water Vapor (PWV) is achieved through the coupling of Zenith Wet Delay (ZWD) and a thermodynamic conversion factor. According to the water vapor state equation of Askne and Nordius [31], the conversion relationship between PWV and ZWD is expressed as

$$PWV={\displaystyle \frac{ZWD}{{\mathrm{\rho }}_w\cdot \mathrm{\kappa }\left(\overline{T_m}\right)}}$$

(5)

where ${\rho }_w$ is a dimensionless conversion function related to the density of liquid water $\kappa \left(\overline{T_m}\right)$ and the atmospheric weighted mean temperature $\overline{T_m}$. Its rigorous mathematical expression integrates the atmospheric refractive index empirical formula:

$$\mathrm{\kappa }\left(\overline{T_m}\right)={10}^{-6}\cdot R_v\cdot \left(k_2^{\prime }+{\displaystyle \frac{k_3}{\overline{T_m}}}\right)$$

(6)

In the equation, physical constants are taken from international standards [32]: $R_v$ is the specific gas constant for water vapor, and $k_2^{\prime }$ and $k_3$ are atmospheric refractivity constants. The atmospheric weighted mean temperature $\overline{T_m}$ is estimated using a regionalized linear regression model using station measured surface temperature to ensure inversion accuracy in the BTH region. The conversion factor $\kappa$ can adaptively adjust the water vapor inversion sensitivity under different temperature conditions, thereby ensuring inversion robustness under severe convective weather.

It should be noted that GNSS-derived PWV primarily reflects column-integrated atmospheric moisture and is used in this study as a thermodynamic precursor signal (1–2 h prior to precipitation) rather than for capturing minute-scale convective variability.

2.2. Radar Composite Reflectivity Data Processing

Radar Composite Reflectivity (CR) mosaic data for the Beijing–Tianjin–Hebei region was used, with a time resolution of 6 min, covering the entire study area. The spatial resolution after resampling was set to 0.1° × 0.1°. This resolution is selected to ensure consistency between GNSS station density and radar grids and to maintain stable multi-modal fusion. Composite reflectivity reflects the maximum echo intensity in the vertical column at a given grid point, effectively characterizing the vertical distribution of liquid water content in the cloud body.

Raw data is in polar coordinate format and contains ground clutter interference, making it unsuitable for direct use as tensor input for deep learning models. To achieve spatio-temporal synergy between radar data and GNSS PWV and meet Swin-UNet input specifications, the following preprocessing workflow was executed on the radar data:

• Noise Suppression Based on Statistical Thresholds: Due to interference from ground clutter, insects, or non-meteorological echoes, raw radar mosaics often contain noise. Based on the quality control standards of Zhang et al. [33], a threshold method was used to remove outliers. According to the 3σ criterion, data points with reflectivity less than −30 dBZ or greater than 70 dBZ were marked as invalid values and reset to 0 dBZ to eliminate the influence of noise on convolutional feature extraction. It is noted that 0 dBZ does not indicate absence of precipitation but typically corresponds to very weak echoes. This preprocessing step is applied solely for noise suppression. In addition, reflectivity values were linearly normalized to a fixed range before tensor construction. This normalization improves convergence stability and ensures compatibility with batch-based optimization. By standardizing the dynamic range across time steps, the model focuses on structural evolution rather than absolute magnitude discrepancies caused by different radar calibration states.
• Spatio-Temporal Registration of Heterogeneous Data: GNSS PWV data are site-based discrete time series, while radar data are gridded spatial fields; there is an inconsistency in spatiotemporal benchmarks between the two. In the time dimension, to achieve synchronization of multi-modal data, we adopted a minimum time difference matching strategy. Assuming the set of radar observation times is $\mathcal{T}\ast radar$, for any GNSS observation time $t\ast gnss$, the nearest neighbor radar frame $t_{radar}^{\ast }$ in the time window is selected as the matching object [34]:

$$t^{\ast }\ast radar=\mathrm{argmin}\ast t\in \mathcal{T}\ast radar\left|t-t\ast gnss\right|$$

(7)

Through this matching mechanism, multi-source data were unified to a consistent 6 min time step. In the spatial dimension, the raw radar mosaic usually uses Lambert conformal conic projection or a polar coordinate system, which needs to be uniformly converted to the WGS84 geographic coordinate system consistent with GNSS stations. The nearest neighbor interpolation method was used to re-map the radar grid to the target grid point $\left(lo{n}_i,la{t}_j\right),$ ensuring that every tensor pixel point accurately corresponds to a geographical spatial position, laying the geometric foundation for subsequent channel stacking. Nearest-neighbor interpolation is adopted to avoid introducing excessive smoothing errors, although it may limit spatial continuity.

3.

Data Augmentation: Due to terrain blocking or radar scanning strategies, data holes may exist in some areas. To ensure spatial continuity of CNN inputs, a bilinear interpolation algorithm was used to fill local missing areas. Assuming the four known grid points around the target missing point $P\left(x,y\right)$ are $Q_{11},Q_{12},Q_{21},Q_{22}$, its pixel value is calculated by weighting:

$$Z\left(x,y\right)\approx \frac{1}{(x_2-x_1)(y_2-y_1)}{\displaystyle \sum _{i=1}^2}{\displaystyle \sum _{j=1}^2}{w}_{ij}Z\left(Q_{ij}\right)$$

(8)

where the weight $w_{ij}$ depends on the Euclidean distance between the target point and each reference point. In addition, addressing the long-tail distribution problem caused by the extreme scarcity of severe convective samples in meteorological datasets, we introduced an online data augmentation strategy during the model training phase. By randomly rotating and horizontally/vertically flipping the input tensor, the diversity of heavy precipitation samples was artificially expanded. This not only enabled the model to learn rotation-invariant features but also significantly improved the model’s robustness and generalization ability when dealing with precipitation systems of different scales and shapes.

2.3. Design of Spatio-Temporal Enhanced Attention U-Net Model

This paper proposes a Spatio-Temporal Enhanced Attention U-Net deep learning model (hereinafter referred to as STEA-Swin) based on Swin Transformer, as shown in Figure 1 [35]. Based on the classic U-Net architecture, this model utilizes Swin Transformer blocks to replace traditional CNNs as the backbone network to capture long-range semantic dependencies. It introduces the Convolutional Block Attention Module (CBAM) [36] to optimize skip connections and designs a mixed loss function containing edge perception and geometric correction to solve common blurring effects and extreme value underestimation problems in precipitation forecasting. It should be noted that the primary contribution of this study lies in the physics-guided multi-modal fusion of GNSS-derived PWV and radar observations, rather than solely in architectural modifications. The proposed design explicitly incorporates complementary thermodynamic and kinematic information to enhance precipitation predictability.

The STEA-Swin model adopts a full-process end-to-end encoder–decoder structure, aiming to achieve a one-step mapping from multi-modal meteorological observation data to high-resolution precipitation fields. The input layer receives multi-modal data tensors from the past 10 time steps, with a shape of $\left(\mathrm{B},\mathrm{H},\mathrm{W},{\mathrm{T}}_{\mathrm{in}}\times \mathrm{C}\right)$. Each time step corresponds to 6 min, and the input sequence covers 10 steps (i.e., 60 min history). Input features include radar reflectivity, GNSS PWV, Digital Elevation Model (DEM), and time encoding. The encoder consists of multiple Swin Transformer Blocks and Patch Merging layers, progressively reducing feature map resolution and increasing channel numbers through downsampling to extract meteorological semantic features at different scales. The bottleneck layer abstracts deep features at the lowest resolution, capturing the evolution laws of large-scale weather systems. The decoder consists of Swin Transformer Blocks and Patch Expanding layers, progressively restoring the spatial resolution of feature maps, mapping deep semantics back to high-precision meteorological grid space. Finally, the output layer adopts a dual-task head, where a regression branch is used to predict continuous precipitation intensity, and a classification branch is introduced as an auxiliary task to enhance the model’s ability to distinguish different precipitation levels, thereby improving overall prediction performance, particularly for heavy precipitation events.

To solve the limited receptive field problem of traditional CNNs, this model employs Swin Transformer as the basic building unit. This module contains Window-based Multi-head Self-Attention (W-MSA) and Shifted Window Multi-head Self-Attention (SW-MSA), achieving global information interaction while maintaining linear computational complexity.

Assuming the input feature is $z^{l-1}$, the calculation process of the Swin Transformer block is as follows:

$$\begin{array}{c}{\hat{z}}^l=\mathrm{W}-\mathrm{MSA}\left(\mathrm{LN}\left(z^{l-1}\right)\right)+z^{l-1}\hfill \\ z^l=\mathrm{MLP}\left(\mathrm{LN}\left({\hat{z}}^l\right)\right)+{\hat{z}}^l\hfill \\ {\hat{z}}^{l+1}=\mathrm{SW}-\mathrm{MSA}\left(\mathrm{LN}\left(z^l\right)\right)+z^l\hfill \\ z^{l+1}=\mathrm{MLP}\left(\mathrm{LN}\left({\hat{z}}^{l+1}\right)\right)+{\hat{z}}^{l+1}\hfill \end{array}$$

(9)

where $\mathrm{LN}$ denotes Layer Normalization [37], used to stabilize deep network training; $\mathrm{MLP}$ is the Multi-Layer Perceptron. W-MSA restricts self-attention calculation within non-overlapping local windows. Compared to global attention, the complexity $\mathrm{O}\left({\mathrm{N}}^2\right)$ is reduced to linear with the image size $\mathrm{O}\left(\mathrm{N}\right)$, significantly improving computational efficiency. SW-MSA introduces cross-window connections through cyclic shift operations, enhancing information interaction between adjacent windows, enabling the model to simultaneously pay attention to the spatiotemporal evolution of local convective cells and wide-area precipitation systems.

At the classic skip connections of U-Net, simple feature concatenation often leads to background noise (e.g., ground clutter) from the shallow layers of the encoder interfering with the recovery process of the decoder. Therefore, this paper introduces the Convolutional Block Attention Module (CBAM) to adaptively fuse shallow detail features from the encoder and deep semantic features from the decoder. CBAM contains two sub-modules: Channel Attention and Spatial Attention. Channel Attention compresses spatial dimensions through global average pooling and max pooling, enabling the model to automatically identify and focus on feature channels most important for precipitation prediction (e.g., high-weight strong echo center channels or high PWV gradient zone channels) and suppress irrelevant meteorological element interference. Spatial Attention generates a two-dimensional spatial attention map through pooling operations along the channel dimension, guiding the model to focus on spatially important information areas, such as the leading edge of rain bands or windward slope areas affected by terrain lifting. The fusion formula for the two sub-modules is:

$$F^{\prime }=M_c\left(F\right)\otimes F, F^{″}=M_s\left(F^{\prime }\right)\otimes F^{\prime }$$

(10)

where $F$ is the concatenated feature map, $M_c$ and $M_s$ are the channel and spatial attention maps, respectively, and $\otimes$ represents element-wise multiplication. This design significantly suppresses background noise and strengthens the feature response to core areas of extreme precipitation.

Addressing the problem in meteorological prediction tasks where traditional Mean Squared Error (MSE) loss functions tend to generate smooth, blurry images, leading to the loss of high-frequency details such as heavy precipitation center extremes and cloud cluster edge textures, this model designs a composite loss function named EdgeAwareLoss. It is composed of three weighted parts, aiming to constrain the model from three dimensions: numerical precision, geometric shape, and edge texture:

• Huber Loss (L_huber) [38]: Used for regression of basic pixel intensity. Com-pared to MSE which is overly sensitive to outliers, Huber Loss behaves as squared error when the error is small and as linear error when the error is large (e.g., extreme precipitation points). Therefore, it has stronger robust-ness to outliers in the data, preventing overall gradient instability caused by the model overfitting extreme points.
• Log-Cosh Dice Loss (L_dice) [39]: Used to optimize the geometric shape mat-ching of precipitation areas. The Dice coefficient is usually used for segmentation tasks. After introducing Log-Cosh smoothing, its formula is

$$L_{\mathit{dice}}=\mathit{log}\left(\mathit{cosh}\left(1-{\displaystyle \frac{2\sum y_{\mathit{true}}{y}_{\mathit{pred}}}{\sum y_{\mathit{true}}+\sum y_{\mathit{pred}}}}\right)\right)$$

(11)

This loss function has a continuous and smooth derivative, is easy to backpropagate, and can effectively improve the Critical Success Index (CSI) of precipitation forecasts, ensuring that the predicted rain area shape matches the actual situation more closely.

3.

Sobel Edge Loss (L_edge) [40]: To solve the blurring effect, the model uses the Sobel operator to extract gradients of the prediction map and ground truth map in horizontal (G_x) and vertical (G_y) directions, respectively, forcing the model to learn the gradient changes of the precipitation field:

$$L_{edge}={\displaystyle \frac{1}{N}}\sum \left|\nabla y_{true}-\nabla y_{pred}\right|$$

(12)

This loss term forces the predicted image to maintain sharpness at precipitation boundaries, enabling more accurate depiction of the contours of severe convective weather.

The total loss function is defined as

$$L_{total}=L_{huber}+{\mathrm{\lambda }}_1{L}_{dice}+{\mathrm{\lambda }}_2{L}_{edge}$$

(13)

In the implementation of this embodiment, after hyperparameter optimization, weights are set to ${\mathrm{\lambda }}_1=1.0, {\mathrm{\lambda }}_2=0.5$ to balance pixel regression precision and structural similarity. Through auxiliary classification loss, the semantic understanding ability of the model is further constrained, ensuring that while numerical prediction is accurate, precipitation level classification also maintains high accuracy.

The overall architecture of this study is shown in Figure 2, mainly including four parts: multi-source data preprocessing layer, spatio-temporal feature encoding layer, multi-scale feature decoding layer, and dual-task forecast output layer. The data preprocessing layer performs spatio-temporal registration and normalization on GNSS PWV, Radar CR, and terrain data (DEM), constructing multi-modal input tensors through channel stacking. The spatio-temporal feature encoding layer uses Swin Transformer as the backbone network, capturing long-distance spatio-temporal dependencies through shifted window attention mechanisms, and using Patch Merging for downsampling. The multi-scale feature decoding layer restores spatial resolution through Patch Expanding and introduces CBAM attention modules at skip connections to adaptively fuse shallow texture details with deep semantic information, solving the information asymmetry problem. The forecast output layer designs regression and classification dual output heads, combined with joint optimization using the edge perception loss function, finally outputting high-resolution precipitation intensity maps and graded warning results for future periods, achieving high-precision, high-definition deduction of future precipitation processes.

3. Experimental Design

3.1. Study Area and Dataset Construction

This study selected the Beijing–Tianjin–Hebei region as the experimental area and constructed a multi-source meteorological fusion dataset containing GNSS atmospheric water vapor, radar echoes, and ground precipitation. The data selection time span was from 1 May 2025 to 31 August 2025, covering the main flood season in the region. Regarding dataset division, data from May to July were used as the training set for model feature learning; data from August were used for model evaluation, with 20% divided into the verification set and 80% into the test set. This time-based splitting strategy is adopted to simulate real-world nowcasting scenarios, where models are trained on historical data and applied to predict future events, thereby avoiding temporal information leakage and providing a more realistic assessment of generalization capability.

Experimental data specifically included: GNSS Precipitable Water Vapor (PWV) data retrieved from 250 CORS stations in the Beijing–Tianjin–Hebei region; Radar Composite Reflectivity mosaic data for the same area; and measured precipitation data from 250 ground meteorological stations serving as ground truth labels for model training and verification. As shown in Figure 3, GNSS stations are uniformly distributed in major urban areas of Beijing, Tianjin, and Hebei, and effectively cover areas prone to heavy precipitation such as mountainous regions, ensuring that PWV data has good spatial representativeness within the study area. To verify the validity of radar data, Figure 4 shows the radar composite reflectivity mosaic at 6:00 on 28 July 2025. This moment corresponds to a typical torrential rain event, where the central and southern parts of Beijing–Tianjin–Hebei showed obvious strong reflectivity centers (>50 dBZ), highly consistent with the actual torrential rain falling area.

To ensure the reliability and physical consistency of multi-source inputs, a systematic data preprocessing workflow was implemented prior to model training. First, a sliding window algorithm was applied to remove abnormal jumps in the GNSS PWV time series, effectively suppressing observational noise and ensuring temporal continuity. Second, interpolation techniques were employed to fill local gaps in radar data, improving spatial completeness. Finally, all multi-source data were uniformly resampled to a 6 min temporal resolution and a spatial resolution of 0.1° × 0.1°, constructing a standardized high-resolution spatiotemporal tensor suitable for deep learning-based fusion. This preprocessing strategy ensures consistency in spatiotemporal alignment across heterogeneous data sources and provides stable inputs for subsequent model training.

The choice of 0.1° spatial resolution represents a trade-off between GNSS station density and radar coverage consistency, ensuring stable multi-modal fusion while acknowledging limitations in resolving sub-kilometer convective structures.

3.2. Experimental Environment and Training Strategy

This experiment was carried out on a high-performance computing platform. The hardware environment was equipped with Intel Xeon Gold 6348 processors and NVIDIA A100 (80 GB) graphics acceleration cards, equipped with 128 GB of memory to meet the throughput needs of large-scale spatiotemporal data. The software environment was based on Python 3.9, integrating PyTorch 2.0 and TensorFlow 2.10 deep learning frameworks.

In this study, the model training configuration and optimization strategy were carefully designed to ensure stable convergence and effective learning of multi-source spatiotemporal features. The model was trained using the AdamW optimizer with an initial learning rate of 0.0001 and a batch size of 16 over 200 epochs. A custom edge-perception joint loss function was adopted, comprising Huber Loss for regression, Log-Cosh Dice Loss for precipitation classification, and Sobel Edge Loss for capturing fine-scale spatial gradients. A weighted combination strategy assigned the regression task a weight of 1.0 and the classification task a weight of 0.5, emphasizing the model’s ability to accurately represent both precipitation intensity and location morphology.

The temporal resolution of 6 min and spatial resolution of 0.1° were chosen to balance the trade-off between capturing high-frequency convective dynamics and maintaining numerical stability in multi-source fusion. DEM data were utilized to assist the model in learning terrain-induced precipitation patterns, particularly in mountainous regions, enhancing the spatial representativeness of predictions.

These experimental settings, including data preprocessing, dataset construction, and model training configuration, are all defined within the Materials and Methods framework to ensure methodological clarity and reproducibility, in accordance with standard scientific reporting practices.

It is important to contextualize the chosen 10 km spatial resolution (equivalent to a 75 × 75 grid over the study domain) within the hardware limitations of deep learning frameworks. While a 75 × 75 grid is computationally trivial for traditional Numerical Weather Prediction (NWP) models operating iteratively on CPU clusters, spatio-temporal deep learning architectures—such as the Swin-UNet employed in this study—impose fundamentally different memory allocation demands. In our model, multi-step and multi-channel data sequences must be processed simultaneously within the limited Video Random Access Memory (VRAM) of a GPU. Maintaining a 1 km resolution (a 750 × 750 grid) across multiple frames would drastically escalate the memory footprint, primarily because the computational complexity and memory requirement of the self-attention mechanism scale quadratically with the spatial resolution ($\mathcal{O}\left({\mathrm{N}}^2\right)$, where $\mathrm{N}$ is the number of spatial tokens). Such high-resolution sequence modeling leads to immediate Out-of-Memory (OOM) failures on current hardware configurations. Consequently, the 10 km resolution represents a necessary computational compromise, ensuring that the model can successfully ingest extended multi-frame sequences and maintain an adequately large global receptive field without exceeding GPU hardware constraints.

4. Results and Analysis

4.1. GNSS PWV Inversion Accuracy Comparison

To ensure that the water vapor features input into the STEA-Swin model have high physical authenticity and data reliability, this study first rigorously verified the accuracy of the solved GNSS Precipitable Water Vapor (PWV) products. The period from May to August 2025 was selected as the verification period. This period covers the typical summer main flood season, with frequent precipitation and drastic water vapor changes, which can fully test the performance of data under extreme weather backgrounds. The Fifth Generation Atmospheric Reanalysis (ERA5) published by the European Centre for Medium-Range Weather Forecasts (ECMWF) was selected as the reference standard for testing GNSS PWV retrieval accuracy [41]. As the successor to ERA-Interim, ERA5 is based on an advanced 4D-Var data assimilation system, deeply fusing multi-source heterogeneous data such as global satellite remote sensing, radiosondes, and ground observations, significantly improving the physical consistency and accuracy of atmospheric parameter estimation. This dataset provides high-quality atmospheric variable fields with global full coverage, a spatial resolution of 0.25° × 0.25°, and a time resolution of 1 h, capable of finely characterizing the rapid change process of atmospheric water vapor. Given its excellent spatio-temporal continuity and widely verified accuracy, ERA5 PWV is currently recognized by the international meteorology and geodesy communities as reliable ground truth data for evaluating the performance of ground-based GNSS water vapor retrieval algorithms and analyzing regional water vapor climate characteristics [42]. This study extracted ERA5 PWV sequences strictly matched with GNSS stations in space and time based on bilinear interpolation, providing an objective basis for subsequent accuracy verification and error analysis [43].

The overall accuracy verification results are shown in Figure 5 (Scatter Density Plot). Statistical analysis shows that the GNSS PWV retrieved in this paper and ERA5 PWV exhibited extremely high consistency. Specific quantitative indicators show that the Pearson correlation coefficient (R) between the two is as high as 0.974, indicating an extremely strong linear positive correlation; the Root Mean Square Error (RMSE) is controlled at 3.63 mm, and the Mean Bias (Bias) is only −0.46 mm. In the density scatter plot, the vast majority of sample points converge tightly around the 1:1 diagonal, and the core area representing high data density (red hotspot area) almost completely coincides with the diagonal. This result powerfully proves that the GNSS PWV data retrieved in this paper has no obvious systematic bias, and no overall overestimation or underestimation phenomenon occurs. It possesses extremely high retrieval precision and reliability, capable of accurately quantifying the numerical characteristics of regional atmospheric water vapor content.

Figure 6 and Figure 7 further reveal error laws from the dimensions of station dispersion and spatial distribution. From Figure 6 (Station Error Statistical Histogram), among the selected 170+ observation stations, the Mean Bias (blue bar chart) of the vast majority of stations fluctuates slightly around the 0 value, mainly concentrated within ±4 mm; although the RMSE (black broken line) of individual stations reaches a peak of 6–8 mm, the overall trend is stable, maintaining a low level of around 3 mm. Combined with the analysis of Figure 8 (Spatial Error Distribution Map), the spatial distribution of RMSE and Bias within the Beijing–Tianjin–Hebei region shows certain landform relevance. Figure 8a shows that low RMSE values (blue areas) are widely distributed in plain areas, while higher RMSE values (orange-red areas) are mainly concentrated in the junction zone of the Yanshan and Taihang Mountains in the central part of the study area. This is mainly due to the significant elevation difference between GNSS stations and ERA5 grid points under complex terrain, increasing the difficulty of local water vapor vertical distribution correction. Nevertheless, Figure 8b shows that the spatial distribution of Bias remains relatively uniform, with most areas showing cyan-green representing low bias, indicating that the overall retrieval algorithm maintains good robustness under complex terrain.

Figure 7 shows the details of the time series comparison between GNSS PWV and ERA5 PWV at a representative station (SZSL station) during the study period. The horizontal axis covers the complete time span from late April to early September. It can be seen that as the season progresses, PWV values show a significant seasonal climbing trend (rising from 5–15 mm in spring to over 40 mm in midsummer). Throughout the time series, the fluctuation trends of GNSS PWV (black solid line) and ERA5 PWV (red dashed line) are highly consistent, and the curves of the two almost overlap in most periods. It is particularly noteworthy that during the water vapor peak period in mid-to-late July (e.g., around 07–24), GNSS PWV accurately captured the high water vapor content exceeding 45 mm and its violent short-term fluctuation characteristics, without obvious Phase Lag at key turning points of wave peaks and troughs. This characteristic strongly confirms that the GNSS PWV data obtained in this paper has excellent sensitivity in reflecting high-frequency, short-term weather processes and can provide credible water vapor transport information rich in high-frequency details for the STEA-Swin model.

4.2. Evaluation of Precipitation Forecast Accuracy in Beijing–Tianjin–Hebei Region

To comprehensively and objectively evaluate the forecast performance of the STEA-Swin model under different precipitation intensities, this study selected four core indicators common in the meteorological field: Probability of Detection (POD), False Alarm Ratio (FAR), Critical Success Index (CSI), and Heidke Skill Score (HSS). These metrics were jointly selected to ensure a balanced evaluation framework: POD emphasizes the sensitivity to precipitation occurrence, CSI reflects overall event prediction accuracy, FAR quantifies false alarm suppression capability, and HSS evaluates forecast skill relative to random chance, thereby providing a comprehensive assessment of both detection capability and reliability.

Strict ablation experiments and horizontal comparison experiments were designed, and the model design is shown in

Table 1. Experimental Groups.

Groups	Description
F_STEA-Swin	STEA-Swin model (fusing radar and pwv)
STEA-Swin	STEA-Swin model (using radar data only)
F_ ConvLSTM	ConvLSTM model (fusing radar and pwv)
ConvLSTM	ConvLSTM model (using radar data only)
F_ GRU_Seq2Seq	GRU_Seq2Seq model (fusing radar and pwv)
GRU_Seq2Seq	GRU_Seq2Seq model (using radar data only)

. We compared the STEA-Swin model proposed in this paper with two types of benchmark models:

• Data Source Ablation Comparison: To explore the necessity of multi-source data fusion, we set up a model of the same architecture relying only on radar echo data (STEA-Swin model (using radar data only)) to verify the gain of introducing GNSS PWV on precipitation forecasting.
• Architecture Performance Comparison: To verify the advanced nature of the Swin Transformer architecture, we selected classic Recurrent Neural Network models in the meteorological field as competing products: the classic spatio-temporal sequence prediction model combining convolution and LSTM (ConvLSTM model) and the sequence-to-sequence model based on Gated Recurrent Units (GRU_Seq2Seq model), to test the advantages of the model in this paper in capturing long-distance spatio-temporal dependencies and alleviating the gradient vanishing problem.

From the heatmap data in Figure 9, the F_STEA-Swin model proposed in this paper demonstrates optimal comprehensive performance across four key meteorological evaluation indicators: POD, FAR, CSI, and HSS. At the torrential rain level, the POD of the fused model F_STEA-Swin reached 0.676, an increase of 21.5% compared to the single data source STEA-Swin (0.556); CSI increased from 0.345 to 0.409 (an increase of approximately 18.5%). This strongly proves that the synergistic effect of GNSS PWV and Radar CR can effectively supplement the water vapor evolution information of radar during the generation and dissipation of strong convection, thereby correcting the intensity underestimation problem of single radar extrapolation.

Horizontal comparison of different network architectures shows that the STEA-Swin series exhibits advantages over traditional Convolutional Recurrent Networks (ConvLSTM) and Recurrent Neural Networks (GRU_Seq2Seq), and this performance gap tends to widen as precipitation intensity increases. In the CSI indicator for torrential rain, STEA-Swin (0.409) is significantly superior to ConvLSTM (0.312) and GRU_Seq2Seq (0.256). This is mainly attributed to the hierarchical attention mechanism of Swin Transformer introduced by STEA-Swin, which can establish global spatio-temporal dependencies, effectively overcoming the gradient vanishing and error accumulation problems common in RNN-type models during long-sequence prediction, thus maintaining higher structural integrity in the deduction of complex and changeable severe convective weather.

Observing the change of CSI indicators with precipitation levels, the performance of all models declines as rain intensity increases. However, traditional models (such as ConvLSTM) showed a sharp decline in CSI during the transition from heavy rain to torrential rain (0.474 → 0.312, a drop of 34%); while the attenuation of F_STEA-Swin was more gentle (0.485 → 0.409, a drop of only 15%). This indicates that the model in this paper has higher sensitivity to extreme values and stronger anti-interference robustness. Regarding the trade-off between false alarms and missed alarms: under the torrential rain level, F_STEA-Swin controlled the FAR at 0.465, significantly lower than the 0.588 of F_GRU_Seq2Seq.

In this study, the proposed architecture is specifically designed to generate multi-step precipitation forecasts. Rather than producing a single future snapshot, the model sequentially outputs predictions at a temporal interval of 1 h, extending to a maximum lead time of 3 h. Consequently, a single inference process yields three distinct temporal frames corresponding to lead times of T + 1 h, T + 2 h, and T + 3 h. This multi-step output design allows for a comprehensive evaluation of the model’s capability to capture the continuous spatio-temporal evolution and intensity variation of convective precipitation systems over the short-term nowcasting window. Building upon this multi-step design, the prediction performance was further analyzed at multiple lead times (+1 h, +2 h, and +3 h) to evaluate forecast skill decay, as illustrated in Figure 10. The results indicate that although all models exhibit a natural decline in forecasting skill with increasing lead time due to the chaotic nature of convective systems, the STEA-Swin model maintains relatively higher stability across all horizons. The degradation rate of CSI and POD for the proposed model is significantly slower compared to baseline models, demonstrating its stronger capability in preserving spatio-temporal coherence during longer prediction intervals.

To quantitatively substantiate the claim regarding the temporal stability and performance consistency of the proposed models across extended lead times (+1 h, +2 h, and +3 h), a multi-step verification was conducted, specifically focusing on the challenging ‘ Torrential Rain’ regime which aligns with the convective event illustrated in Figure 10. At the +1 h lead time, the optimal fusion model, F_STEA-Swin, achieves a prominent CSI of 0.409 and a POD of 0.676, notably outperforming the baseline F_ConvLSTM (CSI: 0.312, POD: 0.408) and the single-modality ablation model STEA-Swin (CSI: 0.345, POD: 0.556). As the forecast horizon extends to +3 h, atmospheric chaotic dynamics inevitably trigger performance decay across all architectures. However, F_STEA-Swin displays superior resilience against error accumulation, maintaining a CSI of 0.318 (a moderate performance decay of 22.2% from +1 h). In comparison, the performance of F_ConvLSTM and the single-modality STEA-Swin degrades more sharply, with their +3 h CSIs dropping to 0.222 (decay of 28.7%) and 0.211 (decay of 38.8%), respectively. These quantitative statements rigorously confirm that the synergistic multi-modal fusion and spatio-temporal attention mechanism within F_STEA-Swin preserve higher stability and robustness during long-sequence nowcasting, a trend visually corroborated by the radar echo evolution profiles demonstrated in Figure 10.

Figure 10 illustrates the precipitation forecasting performance of the F_STEA-Swin, STEA-Swin, and F_ConvLSTM models against the ground truth for a severe rainfall event (initialized at 10:00 on 19 August 2025) at +1 h, +2 h, and +3 h lead times. Utilizing a discrete color scale to distinctly delineate precipitation grades, it is visually evident that F_STEA-Swin delivers the most robust performance. It accurately captures the spatial distribution, sharp intensity gradients, and temporal evolution of the convective system. Notably, the predicted heavy rainfall centers (magenta areas, >30 mm/h) remain highly consistent with the observations even at the +3 h lead time. In comparison, the STEA-Swin model exhibits a significant underestimation of precipitation intensity, with the extreme rainfall cores noticeably decaying by T + 2 h and T + 3 h. Conversely, the F_ConvLSTM model suffers from severe structural smoothing and spatial displacement; it fails to preserve the fine-scale boundaries of the convective system, resulting in an overly diffused prediction area and intensity attenuation. These results underscore the superior capability of F_STEA-Swin in modeling the complex spatio-temporal dynamics of heavy precipitation.

It is worth noting that the reported performance improvements (21.5% in POD and 18.5% in CSI) are currently presented as point estimates. Due to data and computational constraints during this cycle, a full bootstrap-based uncertainty evaluation was not performed, which represents a limitation of the current study. Future work will investigate the statistical variability across more heterogeneous weather events.

In summary, by effectively integrating multi-source physical features and leveraging the global receptive modeling capability of Transformers, the F_STEA-Swin model not only maintains high accuracy in conventional precipitation forecasting but also decisively mitigates the common blurring issue in deep learning extrapolations. It makes breakthrough progress in the accurate hit of heavy precipitation cores and the effective suppression of false alarms, validating its superiority as an advanced new-generation nowcasting algorithm.

5. Conclusions and Future Work

In the context of strengthening urban climate resilience against global warming-induced extreme weather, precise nowcasting of short-term heavy precipitation is critical for effective disaster mitigation. Traditional forecasting methods often struggle to capture the rapid evolution of convective systems due to limited spatiotemporal resolution or single-source data constraints. In this study, we focus on a seasonal-scale nowcasting framework rather than long-term climate resilience assessment, aiming to improve short-term forecasting accuracy through multi-source data fusion. By integrating ground-based GNSS-derived Precipitable Water Vapor (PWV) with radar composite reflectivity within an AI-driven framework, this study demonstrates the potential of combining thermodynamic and dynamic atmospheric information to enhance precipitation nowcasting performance.

The primary conclusions of this research validate the efficacy of advanced multi-GNSS data processing and multi-modal AI fusion. We successfully retrieved high-resolution PWV from 250 CORS stations in the BTH region using the dual-frequency ionosphere-free PPP method, achieving a high correlation coefficient of 0.974 and an RMSE of 3.63 mm against ERA5 reanalysis data, indicating that ground-based GNSS observations can reliably capture regional atmospheric water vapor variability. It should be emphasized that the PWV data used in this study are derived from ground-based GNSS observations rather than satellite-borne sensors, ensuring high temporal resolution and local sensitivity to moisture evolution.

To leverage this data, we proposed the STEA-Swin model, which addresses the heterogeneity between 1D satellite water vapor time series and 2D radar spatial fields. By integrating Swin Transformer blocks with a dual-attention mechanism and a composite Edge-Aware loss function, the model effectively captured the nonlinear coupling between macroscopic water vapor accumulation and microscopic hydrometeor evolution. Consequently, during the 2025 flood season, the framework significantly outperformed single-source baselines, achieving a 68% prediction accuracy for torrential rain events and increasing the CSI by approximately 18.5% compared to radar-only models, while effectively reducing false alarms. These improvements are consistently observed across multiple forecast lead times (+1 h to +3 h), indicating that the proposed model maintains relatively stable predictive skill under increasing forecast horizons.

Overall, the results highlight that incorporating physically meaningful atmospheric variables, such as GNSS-derived PWV, into deep learning frameworks can enhance both detection capability and structural accuracy of precipitation forecasts. Rather than claiming a fully operational or universally robust system, this study provides evidence that multi-source data fusion is a promising direction for improving short-term heavy precipitation nowcasting at the regional scale.

Despite these encouraging results, several limitations remain. First, the current study is based on data from a single flood season (May–August 2025), which restricts the evaluation of long-term generalization and inter-annual variability. Second, the spatial resolution (0.1°) may not fully resolve fine-scale convective structures, particularly in urban environments. Third, potential uncertainties introduced by terrain-induced PWV interpolation errors and radar vertical structure effects (e.g., bright band) require further investigation.

Future work will focus on addressing these limitations and extending the applicability of the proposed framework. We plan to expand the dataset to multi-year observations to evaluate model robustness under diverse synoptic conditions and extreme events. In addition, we will incorporate additional atmospheric observations, such as satellite brightness temperature data and vertical profile information, to construct a more comprehensive three-dimensional representation of the pre-convective environment. To further improve model reliability, future research will explore the integration of physical constraints, such as atmospheric conservation laws, into the learning process, as well as adaptive modeling techniques to better capture multi-scale precipitation dynamics. Finally, the framework will be extended to other regions with high exposure to extreme precipitation, enabling region-specific optimization and broader validation of its generalization capability.