Aug 2026· Journal of Meteorological Research· Vol 40, pp. 1137 - 1150· 0 citations· 39 references
TL;DR
Overall, the SAECNN provides a computationally efficient approach for downscaling and improving coarse precipitation forecasts of EPS, while enhancing probabilistic forecasting skills.
Accurate short-term precipitation forecasting, which targets predictions up to 3 days ahead, has long been challenging, mainly due to extreme sample imbalance and complex multiscale physical processes. Existing regression- or classification-based deep learning methods often produce over-smoothed precipitation fields, lack physical guidance, and provide limited capability for uncertainty quantification, while radar-based nowcasting methods are unsuitable for multi-day forecasting. To address these limitations, we propose RainCast, a high-resolution framework for hourly precipitation forecasting over China up to 72 hours ahead at 0.05° resolution. RainCast incorporates two key designs: (1) a physics-guided feature extractor, which emulates continuity-equation-based diagnostics to extract vorticity, divergence, and vertical-motion-related signals from circulation fields; and (2) a multi-head output design that supports both deterministic forecasts with a regression head and probabilistic multi-member forecasts with an ensemble head, thereby reducing over-smoothing. Experiments show that RainCast consistently outperforms baselines. For heavy rainfall events (50 mm/24 h), RainCast improves CSI by up to 62.82% over GFS, improves CRPS by 19.35% over GEFS. And interpretability analysis further identifies 600-hPa temperature as a key signal for East Asian precipitation forecasting, likely related to hydrometeor phase transitions near the freezing–melting layer.
Guanlong Ma, Weiqiu Chen, Yang Zhao et al.· Proceedings of the 32nd ACM...· 0 citations
Accurate forecasting of extreme precipitation remains a critical challenge, owing to its heavy‐tailed distribution and severe class imbalance that limit conventional deep learning approaches. This study presents a two‐stage cascade‐diffusion framework for extreme precipitation forecasting over eastern China (110°–130°E, 20°–40°N). The first stage employs a cascade binary classification module comprising 15 independent U‐Net models. Each model was trained for a specific precipitation threshold and integrated via hierarchical stacking, producing a coarse‐resolution (0.25°) diagnostic field with reliable spatial localization across intensity levels. The second stage employs a diffusion‐based post‐processing module for simultaneous bias correction and probabilistic spatial downscaling from 0.25° to 0.1°. This step recovers distributional continuity and preserves the heavy‐precipitation tail via higher‐percentile ensemble aggregation (P80, P90). The framework uses ERA5 reanalysis and GPM IMERG over 2013–2023, with 2023 as an independent test year. Results demonstrate that the cascade module achieves Critical Success Index (CSI) improvements of 25%–54% over regression and multi‐task learning baselines at the 50–100 mm 6h
−1
thresholds. The diffusion post‐processing module further improves the distributional fidelity, with P80 and P90 aggregation both achieving strong detection skill at high thresholds. While P90 attains the highest CSI at extreme thresholds, P80 provides a more balanced performance with better‐controlled frequency bias. Variable importance analysis via sequential feature selection reveals physically interpretable threshold‐dependent pressure‐level contributions. Near‐surface levels dominate across all intensities, while upper‐tropospheric levels provide progressively greater marginal skill for extreme events. These results establish the cascade‐diffusion framework as a practically viable approach for operational extreme precipitation forecasting.
Leyan Dai, H. Yuan, Yiheng Liu et al.· Journal of Geophysical Resea...· 0 citations
High-resolution precipitation fields are essential for hydrological impact assessment, yet global climate model outputs are too coarse and biased for direct use. AI-based statistical downscaling with diffusion models offers a promising approach, but the mechanism by which large-scale atmospheric predictors condition generation remains largely unexplored. We investigate three conditioning strategies for a denoising diffusion probabilistic model applied to daily precipitation downscaling: channel concatenation of upsampled coarse predictors, cross-attention conditioning with a learned convolutional encoder, and cross-attention conditioning with the frozen encoder of the pretrained Prithvi WxC weather foundation model. All strategies are evaluated against an unconditioned baseline under identical conditions using probabilistic, distributional, spectral, and extreme-event metrics for the Colorado River Basin. Concatenation conditioning achieves the lowest point-wise CRPS and MSE, but tends to produce over-smoothed fields that suppress high-intensity events. In contrast, cross-attention conditioning provides substantially better distributional realism and modest improvements in spectral fidelity. Improvements are greatest for extremes: the Prithvi-WxC conditioned model retains over half of>100mm/day events, although estimates are uncertain due to limited samples. When trained on the full dataset, the learned convolutional model performs similarly to the foundation model-conditioned approach while requiring lower computational resources. However, the Prithvi-WxC-conditioned model achieves comparable performance with only five years of training data. These results indicate that cross-attention conditioning offers advantages over simple concatenation for probabilistic precipitation downscaling, and that pre-trained foundation model representations may offer benefits in data-limited settings.
Victor Nascimento Ribeiro, Jorge Guevara, J. Moraga et al.· 0 citations
Extreme rainfall events (ERES) are among the most challenging hydro-meteorological phenomena to forecast because the complex, nonlinear atmospheric processes involved span multiple spatial and temporal scales and are further intensified by rising climate variability. While Numerical Weather Prediction (NWP) models offer physically consistent representations of atmospheric dynamics, they still struggle to resolve localized convection, rapidly evolving storm systems, and rare, high-intensity precipitation events. This study systematically reviews recent advances in Artificial Intelligence (AI) for extreme rainfall prediction published between 2020 and 2026, using a transparent literature search strategy, predefined screening criteria, quality assessment, and a structured literature review matrix. The review synthesizes evidence from a final evidence base of 139 studies across several dimensions, including deep learning (DL) methodologies, multimodal data fusion, forecasting horizons, operational deployment, uncertainty quantification, and emerging intelligent forecasting paradigms. The analysis reveals a transition from deterministic DI models to hybrid, physics-informed, uncertainty-aware, and operational forecasting systems that increasingly integrate AI with physical knowledge and heterogeneous environmental observations. Despite notable progress, persistent challenges remain, including data scarcity, class imbalance, model generalization, physical consistency, uncertainty estimation, transferability, and the lack of standardized evaluation frameworks. This review provides a unified synthesis of current methodologies. It identifies key research gaps, highlighting the need for evaluation frameworks that address physical plausibility, uncertainty quantification, reliability, transferability, and operational relevance. The findings indicate that future extreme rainfall prediction systems should advance beyond accuracy-focused evaluation toward integrated, trustworthy, uncertainty-aware, and physically consistent forecasting approaches.
Braiton U. Mukhalela, S. Viriri, D. Ndzi et al.· Frontiers in Artificial Inte...· 0 citations
Precipitation nowcasting over the immediate 10-90 min period is important for flood management and real-time decision-making in urban regions. Conventional short-range forecasting with high-resolution numerical weather prediction requires frequent data assimilation, model initialization, and spin-up, introducing computational latency. Machine learning provides an alternative by learning storm evolution directly from high-frequency observations and producing forecasts quickly after training. This is particularly relevant for Mumbai, India, where monsoon convection, land-sea interactions, and localized intense rainfall make short-term prediction difficult. Here, we develop a compact radar-only nowcasting framework that combines multi-elevation reflectivity, Doppler radial velocity, and radial-velocity-gradient proxy features within an encoder-decoder U-Net. Using the most recent radar volume scan, the model predicts 12 future composite reflectivity fields at 7.5-min intervals up to 90 min lead time. The derived velocity magnitude, divergence-like, directional-shear, and vorticity-like channels represent kinematic signatures associated with convergence and boundary interactions without requiring full wind-field retrieval. A high-reflectivity attention module improves sensitivity to convective cores, and physics-guided attribution examines whether the learned sensitivities are meteorologically meaningful. The model is trained using Mumbai Doppler radar observations from May to August 2023 and evaluated on temporally independent events. At 90 min lead time, Critical Success Index values are 0.437, 0.332, and 0.193 for $\geq$10, $\geq$20, and $\geq$30 dBZ thresholds, respectively. Compared with persistence, the model gives lower RMSE and higher spatial correlation at longer lead times. Once trained, it runs on a standard computer, generating nowcasts within seconds for real-time use.
Akshay Sunil, Muhammed Rashid, Raja Sekhar Sivaraju et al.· 0 citations
This study proposes a novel deep learning framework that uses a U-Net to generate an initial high-resolution SST estimate, which is subsequently refined using a residual corrective approach, and progressively refines initial U-Net predictions by incorporating dynamically scaled residuals at each step, enabling accurate capture of broad patterns and fine-grained features such as eddies and fronts.
Onkar Jadhav, Tim French, I. Janeković et al.· 1 citation
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
