The problem is more consequential than it might first appear. A single weather radar can only see effectively out to roughly 200 to 300 kilometers, and terrain, buildings, and vegetation can block parts of its view. To monitor large-scale weather systems continuously, meteorological agencies fuse observations from networks of radars into mosaic products. In China, the operational Severe Weather Automatic Nowcasting system, run by the Hubei Meteorological Observatory, generates composite reflectivity mosaics from nine S-band radars at a spatial resolution of one kilometer and a temporal resolution of six minutes. These products underpin quantitative precipitation estimation and the nowcasting of severe convective weather. When a region of the mosaic goes dark, the accuracy of rainfall estimates and storm warnings degrades immediately.

Existing repair strategies have struggled to keep pace with the variety of ways data can go missing. Physical approaches, such as correcting beam blockage using dual-polarization measurements, address only specific obstruction problems. Optical flow methods estimate the motion field of radar echoes and extrapolate observed information into the gap, which works reasonably well when storms evolve smoothly and the missing region is spatially continuous. But during rapidly developing convection, when echoes can form, dissipate, or intensify abruptly, the assumption that echoes simply translate breaks down and reconstruction accuracy collapses. Meanwhile, deep learning approaches such as CNN-BiConvLSTM and DSA-UNet have shown strong modeling power, yet most of them depend on a missing-data mask, an explicit map of which pixels are absent. In real operational mosaics, especially when the spatial distribution of gaps is highly uncertain, such masks are difficult to obtain, which limits practical deployment.

The new method, called BiConvLSTM-UNet, sidesteps that requirement entirely. It treats missing echo restoration as a sequence reconstruction task: given a corrupted sequence of radar frames, the model learns the inherent spatiotemporal patterns of echo evolution and reconstructs the complete sequence, implicitly identifying and filling the gaps. The architecture follows a U-shaped encoder-decoder design. The encoder stacks depthwise separable convolution modules that progressively reduce spatial resolution while expanding the channel dimension, building hierarchical spatial features. At the bottleneck, a bidirectional ConvLSTM module models the feature sequence in both forward and backward time directions, so the model can draw on context from frames both before and after a gap. The decoder then restores spatial resolution using PixelShuffle upsampling combined with skip connections that merge high-level semantics with fine spatial detail, and a ReLU1 activation confines outputs to the physically valid range between zero and one.

Training the model required careful data engineering. The team used composite reflectivity mosaics from March to September of 2022 and 2023, covering two full rainy seasons over Hubei Province and totaling 102,720 frames. They focused on meteorologically significant echoes between 10 and 75 dBZ, clipped values outside that range, and discarded frames with almost no echo coverage. After testing sequence lengths of 5, 15, and 30 frames, they settled on 15 frames per sample as the best balance between informational richness and computational efficiency, yielding a final dataset of 60,615 frames split into training, validation, and test sets. To simulate realistic data loss, each frame was divided into five sectors, and random subregions of 100 by 100, 150 by 150, or 200 by 200 grid points were blanked out, with one to three frames per sample randomly affected so the model would learn to handle both isolated and consecutive gaps.

The loss function proved equally important. The researchers combined a weighted L1 loss, in which pixel weights increase with echo intensity so the model pays more attention to strong reflectivity, with a multi-scale structural similarity loss that preserves both global morphology and local texture. Ablation experiments confirmed that each component matters. Removing the structural similarity term dropped the peak signal-to-noise ratio from 25.608 to 24.445 and the structural similarity index from 0.765 to 0.741, and for intense echoes above 40 dBZ the critical success index fell to just 0.078. Replacing the weighted L1 term with a standard unweighted version reduced the critical success index at the 40 dBZ threshold from 0.192 to 0.163 and the probability of detection from 0.359 to 0.215, showing that intensity weighting is what enables the model to recover fierce convective cores rather than smoothing them away.

Head-to-head comparisons against ConvLSTM, DSA-UNet, and classical optical flow showed BiConvLSTM-UNet achieving the best overall performance across peak signal-to-noise ratio, structural similarity, critical success index, and probability of detection, while preserving sharper boundaries and more spatially coherent echo structures. ConvLSTM produced robust but locally unsmooth reconstructions, DSA-UNet recovered echoes proactively but tended toward over-smoothed edges and a high false alarm ratio, and optical flow, faithful to echo boundaries when storms moved steadily, failed badly under rapidly evolving convection. One honest caveat emerged: BiConvLSTM-UNet showed a moderately elevated false alarm ratio, reflecting a recall-oriented strategy that prefers to reconstruct plausible echoes rather than miss them, particularly in low-reflectivity or ambiguous areas.

Because the model reconstructs the entire sequence rather than only the gaps, it risks subtly altering pixels that were actually observed. To guard against this, the team added a post-processing step that keeps original observed values everywhere the input data exists and substitutes model predictions only where the input reads zero. The effect was dramatic: in non-missing regions, the root mean square error fell from 0.704 to 0.266 and the mean absolute error from 0.260 to 0.013, while a newly introduced boundary continuity metric showed that the smoothness of intensity transitions across gap edges was preserved rather than degraded. Generalization tests extended the picture further. When the missing length grew from one to three frames to four random frames, performance declined only marginally, but four consecutive missing frames caused consistent deterioration across all metrics, indicating that uninterrupted loss of temporal context fundamentally challenges the model, which then adopts a conservative bias favoring omission over false activation.

The authors also tested the model outside its comfort zone, evaluating it on data from the non-rainy season months of 2023. It maintained moderate performance, with a peak signal-to-noise ratio of 25.838 and a structural similarity of 0.749, but detection of strong echoes weakened as thresholds rose, a consequence of the distributional shift between sparse, dispersed cold-season echoes and the dense rainy-season data it was trained on. The team proposes expanding training data across seasons, incorporating complementary variables such as vertical wind shear and convective available potential energy, and refining loss formulations with spatially adaptive weighting to address these limits. The study does not claim mask-free reconstruction is inherently superior to mask-based methods; rather, it demonstrates for the first time that high-fidelity radar echo restoration is feasible using only the intrinsic spatiotemporal structure of the echo sequences themselves. For operational meteorology, where missing masks are rarely available and every minute of a storm’s evolution counts, that is a meaningful step toward weather radar mosaics that heal themselves in real time.

Subject of Research: Deep learning-based restoration of missing radar echoes in weather radar mosaic data

Article Title: Research on deep learning-based missing echo restoration method for weather radar mosaic data

Article References: Guo, H., Du, M., Fan, X., Wu, C., Lai, A., & Ma, H. (2026). Research on deep learning-based missing echo restoration method for weather radar mosaic data. Atmospheric Measurement Techniques, 19(19), 6341-6356. https://doi.org/10.5194/amt-19-6341-2026

Image Credits: AI Generated

DOI: 10.5194/amt-19-6341-2026

Keywords: weather radar, radar mosaic, deep learning, BiConvLSTM-UNet, data reconstruction, nowcasting, severe convective weather, ConvLSTM, U-Net, optical flow, precipitation estimation, Atmospheric Measurement Techniques

The model, named PI-TGNet, was developed by Dr. Hengde Zhang, a senior engineer at the National Satellite Meteorological Center of China, together with his jointly supervised doctoral students. Its defining feature is written directly into its name: physics-inspired. Rather than allowing a neural network to mine raw data for statistical correlations alone, the researchers embedded physical guidance into the model’s architecture so that it learns the characteristic wind-speed changes that reliably precede thunderstorm gusts. This design choice reflects a growing recognition in the atmospheric science community that purely data-driven approaches, however powerful, can falter when confronted with rare or rapidly evolving phenomena whose training examples are limited. By anchoring the learning process to known physical behavior, PI-TGNet gains a form of structural intuition about where and when convective winds are likely to turn destructive.

What makes the system genuinely multi-source is its fusion of three complementary observing platforms. Ground-based automatic weather stations provide high-frequency point measurements of wind speed and direction at the surface, exactly where gusts do their damage, but the stations are scattered unevenly across the landscape and can easily miss a gust that passes between them. Weather radar offers a rich spatial picture of the convective systems themselves, revealing the reflectivity structures, outflow boundaries, and downdraft signatures that generate surface gusts, yet radar says little directly about the wind a person would actually feel at ground level. Fengyun-4 geostationary satellites contribute continuous wide-area observations of cloud tops and storm dynamics, filling gaps where neither radar nor station coverage is adequate. PI-TGNet weaves these three streams together into a single coherent product: regional thunderstorm-gust identification maps refreshed every 10 minutes at a spatial resolution of 1 kilometer.

The technical mechanism that binds these heterogeneous data together is a cross-attention architecture. Attention mechanisms, borrowed from modern machine learning, allow a model to weigh the relevance of different pieces of information dynamically rather than treating all inputs equally. In PI-TGNet, cross-attention connects the point observations from weather stations with the spatial fields from radar and satellites, so that a sudden wind-speed spike recorded at a single station can direct the model’s focus to the convective structures visible in radar and satellite imagery over the surrounding region. Conversely, an intensifying storm cell seen from space can prime the network to scrutinize nearby station traces for the abrupt changes that signal an imminent gust. This dialogue between sparse point data and dense spatial data is precisely what a human forecaster attempts, often under severe time pressure, when piecing together a gust warning.

Dr. Zhang summarized the philosophy behind the design in remarks accompanying the release: “By incorporating physical guidance into the deep learning model, we enable it not only to recognize abrupt changes in wind-speed curves, but also to focus on the key structures of convective systems.” The statement captures the dual character of the model. On one hand, it behaves like a vigilant observer of surface anemometers, alert to the sharp inflections in wind-speed curves that empirical threshold methods detect only crudely. On the other, it acts like a radar-savvy storm analyst, attending to the organization and evolution of the thunderstorm complexes that produce those curves in the first place. The fusion of both perspectives within a single spatiotemporal network is what distinguishes PI-TGNet from earlier attempts at automated gust detection.

The performance figures reported in the study are striking by the standards of severe-convective-weather verification. When tested on thunderstorm-gust datasets compiled from eastern and southern China, two regions with dense observing networks and frequent warm-season convection, the model achieved a probability of detection of 0.9520, meaning it successfully flagged more than 95 percent of the gust events in the evaluation data. Equally important, it attained a critical success index of 0.8151, a composite measure that penalizes both missed events and false alarms. A critical success index above 0.8 indicates that the vast majority of the model’s gust identifications corresponded to real events, a level of precision that empirical threshold schemes, which typically trade sensitivity for specificity, have struggled to approach. For operational meteorologists, high scores on both metrics translate directly into warnings that are simultaneously more complete and less prone to eroding public trust through unnecessary alarms.

The practical implications extend across several high-stakes sectors. Aviation is perhaps the most exposed: microbursts and gust fronts have historically been implicated in takeoff and landing accidents, and even with modern wind-shear detection systems at major airports, the suddenness of thunderstorm outflows keeps them on the list of most-feared weather hazards. Ground transportation faces its own exposure, as crosswinds capable of overturning high-sided vehicles can materialize along a highway corridor with little lead time. Outdoor work, from construction cranes to festival staging to maritime operations in coastal waters, similarly depends on warnings that arrive early enough to matter. A 10-minute update cycle at kilometer-scale resolution, generated automatically from data that national meteorological services already collect, could give duty forecasters a continuously refreshed picture of gust risk rather than a patchwork of threshold alarms.

The choice of a 1-kilometer grid is also significant in the context of nowcasting trends. Thunderstorm gusts are inherently small-scale phenomena, and identification products coarser than the storms themselves tend to smear the risk across broad areas, diluting its usefulness. By producing maps at a resolution comparable to the width of a typical gust front, PI-TGNet aligns its output with the spatial scale at which decisions are actually made, whether that decision is rerouting air traffic, pausing crane operations, or issuing a public alert for a specific district. The 10-minute cadence matches the tempo of convective evolution, in which a storm can mature from innocuous cell to wind-producing system within the span of a single conventional observation cycle.

The research team is explicit that the current results mark a beginning rather than an endpoint. Their stated plans include evaluating the model under different regional and meteorological conditions, a necessary step because convection in eastern China’s humid coastal plains behaves differently from convection in other climatic regimes, and a network trained on one environment may not transfer cleanly to another. The team also intends to incorporate additional atmospheric physics into the framework, deepening the physical constraints that guide the network’s learning. Both efforts target the same goal: generalizability. A gust-identification system is only as valuable as the range of conditions over which it can be trusted, and the researchers frame their future work as building toward more refined and reliable technical support for regional severe-convective-weather monitoring and early warning.

The broader significance of the study lies in its demonstration of how physics-informed machine learning can address one of the hardest classes of forecasting problems: rare, fast, small, and dangerous. Thunderstorm gusts sit at the intersection of every factor that makes severe weather difficult to warn against, and the fact that a multi-source deep learning model can push detection probabilities above 0.95 while keeping false alarms in check suggests a viable path forward. As geostationary satellites deliver ever richer imagery, radar networks expand, and surface stations multiply, the raw ingredients for such systems will only improve. The remaining work, as Dr. Zhang and his students recognize, is to ensure that the algorithms fusing those ingredients remain grounded in the physics of the storms themselves, so that when the wind rises without warning, the warning arrives first.

Subject of Research: Physics-inspired multi-source deep learning for thunderstorm gust identification

Article Title: Multi-source observations improve thunderstorm gust identification

Article References: Multi-source observations improve thunderstorm gust identification. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: thunderstorm gusts, deep learning, PI-TGNet, weather radar, Fengyun-4 satellites, automatic weather stations, nowcasting, severe convective weather, cross-attention mechanism, early warning, China Meteorological Administration, Atmospheric and Oceanic Science Letters