The radar reflectivity data used in this study were obtained from composite reflectivity mosaic products (MCR) generated by the operational Severe Weather Automatic Nowcasting (SWAN) system operated by the Hubei Meteorological Observatory. These radar mosaic products are derived from nine S-band weather radars deployed across Hubei Province, with a spatial resolution of 1 km and a temporal resolution of 6 min. The dataset encompasses radar observations from March to September in both 2022 and 2023, covering two successive years and capturing the evolutionary features of radar echoes associated with distinct weather systems during the principal rainy season in Hubei Province, amounting to a total of 102 720 frames. To enable efficient model training, a 400 km × 400 km study area was delineated, centered on Wuhan City, Hubei Province (28.72–32.72° N, 112.42–116.42° E). This region features dense and spatially overlapping radar coverage, rendering occurring echo absence highly improbable – thereby minimizing the confounding effect of systematic data missingness. The geographic extent of the study area and a representative example of radar mosaic reflectivity data are presented in Fig. 1.

To establish a valid dataset, the following data preprocessing steps were performed:

First, numerical processing was applied to the radar reflectivity data. This study only focuses on significant precipitation, hail, and other meteorological echoes (with a range of 10 to 75 dBZ) that are of concern in meteorological services. Values below 10 dBZ were set to 0 dBZ, and values exceeding 75 dBZ were truncated to 75 dBZ. The lower threshold is selected to exclude weak or non-precipitating echoes, thereby enhancing the model's focus on meteorologically significant signals, whereas the upper threshold encompasses virtually all intense convective echoes. This procedure provides a consistent data range for subsequent min–max normalization.

Second, appropriate training data were selected. As this study focuses on reconstructing radar echo sequences to fill data gaps, the choice of radar data and sequence length substantially influences experimental outcomes. To prioritize the restoration of missing radar echoes, datasets containing significant echo signals were emphasized. Frames devoid of radar echoes or with minimal coverage – specifically, those in which echo grid points constituted less than 5 % of the total grid – were excluded. Additionally, small isolated echoes and noise, defined as contiguous echo regions comprising fewer than 20 grid points, were set to 0 dBZ. Furthermore, from a temporal modeling standpoint, it is essential to determine a suitable number of consecutive frames per sample to effectively utilize the dynamic evolution of radar echoes for plausible reconstruction of missing areas. To reconcile practical constraints with model training efficiency, three sequence lengths – 5, 15, and 30 frames per sample – were evaluated.

For 5-frame sequences, despite the advantages of higher training efficiency and a larger dataset size, the limited information content within each individual sample increases the difficulty of reconstruction learning in the presence of missing data. However, for 30-frame sequences, although they encapsulate substantial temporal dynamics in radar echoes, the considerable data volume per sample presents considerable challenges to training efficiency, dataset scalability, and hardware constraints. Ultimately, a sequence length of 15 frames per sample was identified as optimal, achieving a balance between informational richness, computational efficiency, and hardware limitations.

Following the aforementioned data preprocessing, the final raw dataset comprised 60 615 frames of radar reflectivity data. A subset of 32 115 frames, spanning March to September 2023, served as the training set for model development. For the purpose of guiding learning rate adjustments, implementing early stopping, and selecting optimal model weights during training, 12 990 frames from March to May 2022 were allocated as the validation set. An independent validation set of 4755 frames from September 2022 was utilized to assess the model's proficiency in reconstructing missing regions and to optimize hyperparameters, including loss function weights. Finally, 10 755 frames from June to August 2022 were designated as the test set for the subsequent quantitative evaluation of experimental results.

The radar mosaic data are generated by integrating observations from multiple weather radars situated at distinct geographical locations at approximately the same time. In the event of system failure, data delay, or data loss affecting any of these radars, the corresponding coverage area will exhibit missing data. To accurately simulate such data loss scenarios, a masking procedure was designed in this study. Specifically, each frame of radar reflectivity data was partitioned into five sectors: upper-left, lower-left, upper-right, lower-right, and central, with each sector spanning a 200 × 200 grid. Upon selection of a sector for masking, a subregion of dimensions 100 × 100, 150 × 150, or 200 × 200 grid points was defined within it, and all radar reflectivity values inside this subregion were set to 0 dBZ (Fig. 2).

It is important to note that during the construction of the loss dataset, both the selection of the loss region and the dimensions of the loss subregion were randomized for each radar echo sample; however, within individual samples, these parameters remained consistent. When employing subregion sizes of 100 × 100 or 150 × 150, all frames from the same sample utilized identical dimensions, while the precise location of the subregion was still randomized, as illustrated in Fig. 2. This methodology was implemented to thoroughly simulate a wide range of missing-data conditions. Furthermore, to simulate unforeseen circumstances, 1–3 frames per sample were randomly subjected to data loss, ensuring that the model could robustly reconstruct missing regions regardless of which frames were affected and whether the losses occurred consecutively. Using this approach, a corresponding loss dataset was constructed to serve as a foundational resource for subsequent model training.

In operational practice, radar mosaic data can suffer from localized gaps in observation within their nominal coverage areas, owing to issues such as radar hardware malfunctions and delays in data transmission. These gaps typically manifest as spatially contiguous regions and persist across successive time steps, significantly compromising the spatiotemporal continuity of radar echo evolution. Consequently, the accuracy of subsequent quantitative precipitation estimation and nowcasting of severe convective weather events is adversely impacted. As illustrated in Fig. 3, while the complete radar observations (Fig. 3c) show full coverage, the mosaic from the preceding time step (Fig. 3b) exhibits a distinct absence of radar echo in the lower-left sector. In contrast, the mosaic from two-time steps earlier (Fig. 3a) covers an even more limited area, with extensive data loss evident in both the upper-left and lower-left regions.

From a temporal modeling standpoint, this study regards the problem of restoring missing radar echoes as a sequence reconstruction task. The complete radar echo sequence is defined as: 1Y={Yt-T+1,…,Yt-1,Yt},Yt∈RH×W where R denotes the complete dataset, while H and W represent the spatial dimensions of the radar mosaics, both set to 400 grid points in this study.

In contrast to conventional restoration techniques that depend on explicit missing masks to guide the reconstruction process, the present study does not presuppose knowledge of the spatial locations of missing regions during training, inference, or post-processing. Instead, the recovery of missing radar echoes is consistently framed as a sequence reconstruction task. It is important to emphasize that the post-processing procedure does not entail additional reconstruction; rather, it serves to constrain and refine the model's output, minimizing potential artifacts in non-missing areas. By learning both the temporal dynamics and spatial structure of radar echo sequences, the model reconstructs the degraded sequence X̃, thereby restoring missing regions through an implicit modeling approach. The target output of the model is defined as: 2Y^t-T+1:t=fX̃t-T+1:t

Here, f(⋅) represents the sequence reconstruction model, Y^ denotes the reconstruction result, and X̃ denotes the radar echo sequence in which some frames have missing data.

The proposed problem formulation enables the reconstruction of missing radar echoes by incorporating constraints such as overall sequence consistency, and, without requiring prior annotations of the missing areas, establishes a foundation for subsequent model design and application in complex missing-data environments.

In response to the aforementioned research problem, we present a sequence reconstruction model – termed BiConvLSTM-UNet (the detailed architecture is shown in Fig. 8) – designed for the restoration of missing data in radar echo sequences. The model accepts a corrupted radar echo sequence (X̃) as input and reconstructs the data by jointly capturing spatiotemporal evolution patterns and structural features inherent in radar echoes, thereby recovering missing information. Although the model's output spans the entire spatial domain, its optimization objective focuses not on reconstructing non-missing regions, but on inferring missing content under constraints that enforce overall sequence consistency.

As illustrated in Fig. 4, the BiConvLSTM-UNet model employs a U-Net architecture comprising three principal components: an encoder, a bottleneck, and a decoder. The encoder incorporates three cascaded depthwise separable convolution modules (IR-Blocks). To preserve local spatial structures during initial feature extraction, the first layer omits downsampling. Subsequent IR-Blocks progressively reduce spatial resolution by setting the stride of the depthwise convolution to 2, thereby expanding the effective receptive field and increasing channel dimensionality. This process facilitates the construction of hierarchical spatial feature representations, establishing a robust foundation for sequence reconstruction. The bottleneck processes high-level features from the final IR-Block of the encoder, further refining them through an additional IR-Block. A BiConvLSTM module is then applied to perform bidirectional temporal modeling on the feature sequence. By concatenating forward and backward outputs along the feature dimension, the model integrates contextual information, capturing the dynamic evolution of radar echoes across the sequence and improving the representation of temporal continuity and evolutionary trends.

The decoder utilizes an upsampling module based on depthwise separable convolution and PixelShuffle (Qin et al., 2020), referred to as the PS-Block. This module first reconstructs spatial resolution through pixel rearrangement, followed by reorganization of the upsampled features across both spatial and channel dimensions via depthwise separable convolution (Howard et al., 2017). Skip connections are then employed to integrate corresponding encoder-derived features, with subsequent refinement through IR-Blocks to combine high-level semantics and low-level spatial details. This procedure is iteratively applied to progressively restore spatial resolution. Finally, an output normalization module (P_C2D) employs pointwise convolution to project the decoded features, while the ReLU1 activation constrains outputs to the interval [0,1], ensuring numerical validity and physical interpretability of the reconstructed radar echo sequences. The formulation of ReLU1 is given as follows: 3f(x)=min⁡(max⁡(0,x),1)

In the absence of prior knowledge regarding missing regions, the BiConvLSTM-UNet model achieves end-to-end reconstruction of incomplete radar echo sequences by jointly modeling temporal and spatial correlations, thereby enabling spatiotemporally consistent radar echo reconstruction.

In this study, the model training employs a hybrid loss function combining weighted L1 loss and Multi-Scale Structural Similarity (MS-SSIM) loss, which jointly constrains the reconstruction of radar echo sequences by enforcing both pixel-wise accuracy and structural fidelity. The weighted L1 loss is specifically utilized to ensure that reconstructed radar reflectivity values closely approximate the ground truth at the pixel level. To enhance the model's attention to regions of intense radar echoes, a pixel-wise weighting scheme (w) is adopted, where the weight increases proportionally with the echo intensity – i.e., stronger reflectivity factors (ZH) are assigned larger weights (Table 1). These weight values, determined empirically, aim to capture the relative significance of varying echo intensities in the reconstruction task, rather than representing rigorously optimized parameters.

The weighted L1 loss is defined as: 4LWL1=1N∑i=1NwiYi-Y^i where N represents the total number of pixels in the sequence, Yi denotes the ground-truth value of the ith pixel in the complete radar echo sequence, Y^i is the reconstructed value of the ith pixel in the reconstructed radar echo sequence, and Wi is the weight assigned to the ith pixel. As the weighted L1 loss is computed across the entire radar echo sequence during training, pixels in non-missing regions are continuously engaged in the supervision process with non-zero weights, thereby imposing effective constraints on the model outputs. Consequently, without the introduction of an explicit mask, the model is capable of both minimizing unnecessary alterations in non-missing regions and directing its focus toward learning the reconstruction of pixels in missing regions.

The MS-SSIM loss enforces constraints on reconstructed radar echoes from a multi-scale structural perspective, thereby ensuring consistency in both global morphology and local textural details. This loss function effectively captures the spatial continuity and hierarchical characteristics inherent in radar echo structures. The MS-SSIM loss is formulated as follows: 5LMS-SSIM=1-1T∑j=1TMS-SSIMYj,Y^j where T represents the total number of frames in a radar echo sequence, Yj corresponds to the ground-truth radar echo data of the jth frame, and Y^j denotes the reconstructed radar echo data of the jth frame.

In summary, the combined loss function is defined as: 6L=λ1LWL1+λ2LMS-SSIM where λ1 and λ2 represent the weighting coefficients for the weighted L1 loss and the MS-SSIM loss, respectively. These two loss components operate in a complementary manner, with the weighted L1 loss managing pixel-level accuracy and the MS-SSIM loss enforcing structural consistency across multiple scales (Zhao et al., 2016). The integration of both terms ensures that the reconstruction of missing regions in radar echo data maintains high numerical precision as well as spatially coherent structural integrity.

The experimental setup utilized a Python 3.10 environment with the PyTorch 2.6.0 framework and CUDA 12.4. Model training was carried out on two NVIDIA A100 GPUs with a batch size of 2, over a maximum of 200 epochs. Optimization was performed using the Adam optimizer with an initial learning rate of 1 × 10⁻⁴.

During the training process, a learning rate scheduling strategy was implemented, wherein the Structural Similarity Index (SSIM) metric, evaluated on the full validation sequence, served as the reference. The learning rate was halved whenever the improvement in validation SSIM failed to exceed a threshold of 1 × 10⁻⁴ relative to the historical optimum for five consecutive epochs, with a minimum learning rate constraint set at 1 × 10⁻⁶.

To mitigate overfitting and enhance training efficiency, an early stopping criterion was implemented. Specifically, training was halted if the SSIM on the validation set exhibited no substantial improvement – defined as an increase greater than 1 × 10⁻⁴ – for a patience threshold of 10 consecutive epochs, or if the learning rate had decayed to its predefined minimum (1 × 10⁻⁶). Early stopping was, however, disabled during the initial 30-epoch warm-up phase.

The training employs a combined loss function incorporating weighted L1 and MS-SSIM, which constrains the reconstructed outputs of the model in terms of both pixel-wise accuracy and structural fidelity. The loss weights (λ1 and λ2) were optimized through a grid search over a predefined range. The final configuration, with λ1=3 and λ2=3, achieves an optimal trade-off between training stability and reconstruction quality.

1.

PSNR (Peak Signal-to-Noise Ratio): Used to quantify pixel-level discrepancies between reconstructed and complete radar echo data, with higher values signifying superior numerical reconstruction fidelity. It is calculated as follows: 7PSNR=10⋅log10(MAX)2MSE where MAX denotes the maximum radar reflectivity factor (set to 75 dBZ in this study and normalized to unity), and MSE represents the mean squared error.

Since the BiConvLSTM-UNet model reconstructs the entire radar echo sequence – including both missing and non-missing regions – its reconstruction may inadvertently alter originally observed (i.e., non-missing) echo structures. To mitigate this undesirable propagation of reconstruction artifacts into reliable observational data, a targeted post-processing step is applied at the model output stage. Specifically, it selectively preserves the original non-missing echo values from the input (lossy) radar data while replacing only the missing regions with the model's predictions. This strategy ensures fidelity to observed evidence in intact areas, minimizes spurious modifications outside missing zones, and thereby enhances the physical plausibility and spatial consistency of the final reconstructed sequence. The post-processing operation is formulated as: 12Ypost(i,j)=Y^(i,j)if  X̃(i,j)=0 and Y^(i,j)≠0,X̃(i,j) otherwise  where Ypost(i,j) denotes the final post-processed radar echo value at pixel (i,j); X̃(i,j) denotes the corresponding value from the input (lossy) radar echo data; Y^(i,j) denotes the model's reconstructed radar echo value at the same location; and (i,j) indexes the spatial coordinates in the two-dimensional radar echo grid.

This study addresses missing radar echo regions through sequence-level reconstruction. To systematically isolate and evaluate the contributions of the BiConvLSTM module and the composite loss function to reconstruction performance, we conduct controlled ablation experiments (Table 2). All ablated variants retain the identical backbone architecture, data split (training/validation/test), optimization configuration (optimizer, learning rate schedule, batch size), and inference protocol – ensuring that observed performance differences arise solely from the targeted component modifications. Specifically, Scheme 4 substitutes the bidirectional ConvLSTM with a unidirectional ConvLSTM to assess the impact of temporal context integration in both forward and backward directions; Schemes 2 and 3, respectively remove the weighted L1 term and the MS-SSIM term from the loss function, enabling quantitative attribution of each loss component's role in preserving radar echo morphology and dynamic coherence; Scheme 5 substitutes the weighted L1 term with the standard L1 term to evaluate the impact of weighting on model performance.

Quantitative evaluation of reconstruction performance within missing radar echo regions is conducted at three reflectivity thresholds: 20, 30, and 40 dBZ – corresponding to light, moderate, and strong precipitation regimes, respectively. Pixel-level fidelity is assessed using PSNR and SSIM, computed exclusively over masked missing regions to isolate reconstruction accuracy and structural preservation (Zhang et al., 2026). Event-based detection capability is evaluated using CSI, POD, and FAR, all calculated on binary masks derived from the missing-region reconstructions and ground-truth observations – thereby quantifying the model's ability to correctly recover echo presence/absence across different intensity levels (Zhao et al., 2024). For all metrics: ↑ denotes higher-is-better; ↓ denotes lower-is-better.

As shown in Table 2, Scheme 1 achieves either the best or second-best performance across most evaluation metrics, demonstrating that the BiConvLSTM architecture and the adopted loss function jointly contribute to effective reconstruction of missing radar echo regions. A direct comparison between Scheme 1 (with BiConvLSTM) and Scheme 4 (without BiConvLSTM) reveals consistent improvements in both CSI and POD, particularly for strong echoes (≥ 30 and ≥ 40 dBZ). These gains indicate that bidirectional temporal modeling enables more comprehensive utilization of contextual information from both preceding and succeeding frames – thereby enhancing the fidelity and physical plausibility of reconstructed radar echo structures.

From the perspective of the loss function, removing the MS-SSIM loss (Scheme 3) leads to substantial declines in both PSNR and SSIM – dropping from 25.608 to 24.445 and from 0.765 to 0.741, respectively – demonstrating that this term is essential for preserving spatial structural fidelity in reconstructed missing regions. Moreover, for strong radar echoes (≥ 40 dBZ), Scheme 3 exhibits marked performance degradation relative to the full model (Scheme 1), with CSI and POD falling to 0.078 and 0.090, respectively. This underscores that accurate recovery of intense radar echo structures critically relies on explicit structural regularization; in its absence, morphological coherence deteriorates significantly. In contrast, eliminating the weighted L1 loss (Scheme 2) yields only marginal reductions in PSNR and SSIM. Nevertheless, a direct comparison between Scheme 1 and Scheme 2 reveals a consistent weakening in echo intensity preservation across all intensity thresholds, accompanied by measurable declines in both CSI and POD. These results indicate that the weighted L1 loss effectively constrains pixel-wise intensity deviation, mitigates over-smoothing of radar echo amplitudes, and thereby enhances quantitative and perceptual reconstruction quality in missing regions. Compared with Scheme 5, Scheme 1 yields consistently higher PSNR and SSIM values, demonstrating that the incorporation of weighted loss enhances overall reconstruction fidelity. Specifically, at the reflectivity threshold of 40 dBZ, the CSI increases from 0.163 to 0.192, and the POD rises from 0.215 to 0.359, indicating that the weighted loss function prioritizes high-reflectivity regions, thereby improving the model's capability to accurately reconstruct intense convective echoes.

It should be noted that, across all echo intensity thresholds, Scheme 1 achieves relatively high CSI and POD values – but at the cost of a moderately elevated FAR. This trade-off suggests that the scheme prioritizes recall-oriented reconstruction: by recovering plausible echo signals in missing regions, it effectively reduces missed detections; however, this sensitivity also leads to an increased incidence of false alarms.

The ablation study demonstrates that the BiConvLSTM module significantly improves temporal modeling capability; meanwhile, the MS-SSIM loss and weighted L1 loss serve complementary functions – MS-SSIM enforces structural consistency, whereas the weighted L1 loss preserves echo intensity fidelity – thereby jointly enhancing the model's performance in reconstructing missing radar echo regions.

To further verify the BiConvLSTM-UNet model's capability in reconstructing missing radar echo regions, we conducted a comparative evaluation against other advanced deep learning approaches – specifically ConvLSTM (Shi et al., 2015) and DSA-UNet (Gong et al., 2023) – as well as a classical motion-based method, optical flow (Ayzel et al., 2019). All models are trained and evaluated under identical experimental conditions – including the same dataset, preprocessing procedure, and evaluation metrics – to ensure methodological fairness and result comparability.

As can be seen in Table 3, the ConvLSTM model achieves the lowest FAR across all reflectivity thresholds – demonstrating its superior capability in false alarm suppression. Its PSNR, SSIM, and CSI scores rank immediately below those of BiConvLSTM-UNet, with negligible inter-metric disparity, indicating robust and balanced performance in missing radar echo reconstruction. However, qualitative analysis in Fig. 5 reveals a consistent limitation: ConvLSTM struggles to reconstruct sharp, spatially coherent boundaries in missing echo regions, resulting in locally unsmooth structural transitions and diminished spatial continuity.

DSA-UNet achieved high POD values across all models, reflecting its strong capability in proactively recovering valid radar echoes. Qualitative analysis in the example results (Fig. 5) further demonstrates that, relative to BiConvLSTM-UNet, DSA-UNet delivers more complete and spatially coherent reconstruction in partially missing echo regions – yielding visually plausible outputs. However, boundary reconstruction remains a limitation: DSA-UNet tends to produce over-smoothed transitions and loss of fine-scale echo morphology, particularly at region edges. Its relatively high FAR further suggests a tendency toward echo overestimation under ambiguous or low-signal conditions. In contrast, the optical flow method better preserves the spatial morphology and structural integrity of radar echo boundaries in reconstructed regions, resulting in subjectively realistic visual outputs. Yet, as a purely motion-based extrapolation technique, it inherently assumes echo translation and approximate intensity conservation – assumptions that fail to account for physical processes such as echo genesis, dissipation, or intensity modulation during temporal evolution. Consequently, optical flow consistently underperforms on quantitative metrics – including PSNR, SSIM, and CSI – with degradation most severe in rapidly evolving convective systems where non-translational dynamics dominate.

Quantitative evaluation and qualitative analysis jointly demonstrate that BiConvLSTM-UNet achieves superior overall performance across key metrics – including PSNR, SSIM, CSI, and POD – and exhibits enhanced fidelity in reconstructing the structural boundaries of missing radar echo regions. However, its elevated FAR reveals a consistent tendency toward echo overestimation, particularly in low-reflectivity or spatially ambiguous areas.

BiConvLSTM-UNet performs holistic reconstruction of missing radar echo sequences – a strategy that, while effective for gap filling, may inadvertently perturb pixel values in non-missing regions. To quantify this side effect and rigorously assess the efficacy of post-processing in preserving structural consistency within observed (i.e., non-missing) echo regions, we restrict analysis exclusively to frames containing missing data. Within these frames, we conduct a pixel-wise comparative evaluation between the raw model outputs and the corresponding post-processed results – specifically within the non-missing regions (Table 4). Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) serve as complementary metrics: RMSE emphasizes sensitivity to localized outliers and large-magnitude deviations, thereby highlighting instances of severe inconsistency; MAE provides a robust measure of average deviation across all non-missing pixels, reflecting global fidelity and stability (Chai and Draxler, 2014). The definitions are as follows: 13RMSE=∑i,jMi,jYi,j-y^i,j2∑i,jMi,j14MAE=∑i,jMi,jYi,j-y^i,j∑i,jMi,j

Here, Yi,j denotes the observed radar echo intensity (in dBZ), y^i,j represents the model-predicted echo value – either from raw reconstruction or post-processed output – and the binary mask (Mi,j) indicates data availability: Mi,j=1 for non-missing (observed) regions and Mi,j=0 for missing regions.

Furthermore, to quantitatively assess the effect of the post-processing procedure on boundary continuity within missing regions, this study introduces the Boundary Continuity Difference (BCD) metric. BCD quantifies the local reflectivity intensity consistency across the boundary of a missing region; specifically, it is computed over a symmetric boundary band – centered on the missing-region boundary and extending 5 pixels inward and outward, respectively. It is defined as follows: 15BCD=1N∑i=1NIin(i)-Iout(i)

Here, Iin and Iout denote the reflectivity values within the inner and outer boundary zones of the missing region, respectively, and N represents the total number of pixels along the boundary included in the computation. A lower BCD value indicates a smaller reflectivity difference across the boundary, implying better boundary continuity and stronger spatial coherence.

The results in Table 4 show that, the model's holistic reconstruction yields RMSE and MAE values of 0.704 and 0.260, respectively, within non-missing regions – indicating unintended distortion of valid radar echo data. Upon integrating the proposed post-processing strategy, these metrics decline substantially to 0.266 (RMSE) and 0.013 (MAE), reflecting a marked reduction in spurious modifications to non-missing regions. Concurrently, the BCD value decreases from 6.41 in reconstruction results to 6.32 after post-processing, demonstrating that the proposed strategy preserves boundary continuity and does not degrade the smoothness of intensity transitions along the boundaries of missing regions.

Crucially, this improvement is achieved without compromising reconstruction accuracy in missing regions. Collectively, these results indicate that the post-processing strategy maintains structural fidelity in observed regions without degrading reconstruction accuracy in missing regions, thereby enhancing the overall quality of the restored radar mosaic data. Figure 6 compares the non-missing regions of radar echo data before and after post-processing.

To evaluate model generalization under variations in missing length and temporal continuity, we designed a controlled experimental protocol comprising distinct training and testing regimes. During training, missing patterns were sampled at random across 1–3 frames – thereby jointly simulating both isolated (non-contiguous) and brief contiguous gaps. During testing, two targeted scenarios were introduced: (i) random 4-frame missing (preserving temporal dispersion) and (ii) consecutive 4-frame missing (enforcing maximal temporal continuity) – to systematically probe reconstruction robustness under extended gap duration and heightened structural discontinuity. Critically, the 1–3 frame missing patterns used in this evaluation were freshly regenerated – decoupled from those employed in ablation and comparative experiments – to ensure independent assessment and mitigate confounding bias in generalization performance.

As presented in Table 5, extending the random missing length from 1–3 frames to 4 frames induces only marginal declines in PSNR and SSIM for the BiConvLSTM-UNet model, confirming its robustness to modest increases in missing duration. In contrast, CSI and POD exhibit more pronounced degradation – particularly under intense radar echo thresholds (≥ 30 and ≥ 40 dBZ) – while FAR remains stable or even decreases across several thresholds. This pattern suggests that the model preserves detection specificity under extended random gaps, albeit with reduced sensitivity to strong echoes. Under consecutive 4-frame missing, however, all metrics – PSNR, SSIM, CSI, and POD – show consistent and statistically meaningful deterioration across reflectivity thresholds, indicating that uninterrupted temporal context loss fundamentally impairs the model's capacity to recover spatiotemporal coherence in radar echo fields. Notably, FAR continues to decrease, reflecting a conservative reconstruction bias: the model increasingly favors omission over false activation when deprived of contiguous temporal cues – a behavior that trades recall for precision under severe structural uncertainty.

As shown in Fig. 7, the BiConvLSTM-UNet model maintains robust reconstruction performance across varying missing patterns – specifically, under both extended missing length and reduced temporal continuity – indicating substantive generalization capability. Nevertheless, performance degrades systematically with increasing missing intensity: under random 4-frame missing, metrics decline modestly; under consecutive 4-frame missing, however, all evaluation metrics (PSNR, SSIM, CSI, POD) exhibit pronounced and consistent deterioration. This graded, interpretable degradation – monotonic with respect to missing severity and qualitatively aligned with the underlying spatiotemporal modeling assumptions – further validates the model's behavior as physically plausible and methodologically sound.