Skilful precipitation nowcasting using deep generative models of radar

Precipitation nowcasting, the forecasting of precipitation up to two hours ahead with high resolution, is critical for various sectors reliant on weather-dependent decision-making, including emergency services, energy management, retail, flood early-warning systems, air traffic control, and marine services. Effective nowcasting demands accurate predictions across spatial and temporal scales, proper accounting for uncertainty with probabilistic verification, and robust performance on rarer but impactful heavy precipitation events. Traditional methods like ensemble numerical weather prediction (NWP) systems, while capable of producing probabilistic forecasts and uncertainty estimates, often yield poor forecasts for the 0-2 hour timeframe due to limitations in model spin-up and non-Gaussian data assimilation. Alternative methods using composite radar observations have emerged, leveraging the high-resolution radar data available at intervals of a few minutes. Established probabilistic nowcasting methods like STEPS and PySTEPS utilize ensembles to account for uncertainty, modeling precipitation using the advection equation with a radar source term. These methods estimate motion fields via optical flow and employ smoothness penalties. However, deep learning approaches have shown potential to surpass traditional methods by directly predicting precipitation rates without relying on physical constraints, thus aiming to capture complex nonlinear phenomena like convective initiation and heavy precipitation more effectively. While demonstrating improved accuracy for low-intensity rainfall, these deep learning systems produce blurry predictions at longer lead times, hindering their operational usefulness, particularly for crucial high-intensity rainfall events.

Existing deep learning methods for precipitation nowcasting have shown promise in predicting low-intensity rainfall but struggle with longer lead times and high-intensity events. Their inherent lack of physical constraints leads to blurry, spatially inconsistent forecasts. Methods like RainNet, convolutional LSTM networks, and MetNet, while offering advancements, often fall short in capturing the detail and uncertainty associated with higher intensity precipitation at longer forecast horizons. The focus on location-specific predictions rather than probabilistic field predictions also limits operational applicability. The current state-of-the-art methods often struggle with representing uncertainty effectively across various spatial and temporal scales. This motivates the exploration of alternative methods like deep generative models, which offer inherent probabilistic capabilities to better manage uncertainty and improve the quality of nowcasts.

This research proposes a novel approach to precipitation nowcasting using deep generative models (DGMs). The core of the methodology lies in a conditional generative model that predicts future radar fields (N) based on past radar fields (M). The model incorporates latent random vectors (Z) and parameters (θ), represented by P(XM+1:M+N|X1:M) = ∫P(XM+1:M+N|Z, X1:M, θ)P(Z|X1:M)dZ. This framework is implemented as a conditional generative adversarial network (GAN), where a generator creates future precipitation samples given past radar observations. The learning process is guided by two discriminators: a spatial discriminator focusing on spatial consistency, penalizing blurry predictions, and a temporal discriminator enforcing temporal consistency by penalizing jumpy predictions. A regularization term further enhances accuracy by minimizing deviations between real and predicted radar sequences at the grid-cell resolution. The generator utilizes a fully convolutional latent module, enabling predictions over larger precipitation fields than those used during training. The model is trained on a substantial corpus of precipitation events extracted from a radar stream, employing importance sampling to ensure sufficient representation of heavy precipitation. Training utilized radar observations from the UK (2016-2018), with the model evaluated on a 2019 test set. A weekly train-test split and validation on US data are also included in the supplementary information. The model's efficiency enables rapid, full-resolution nowcasts, with a single prediction taking just over a second using an NVIDIA V100 GPU.

The proposed deep generative model for radar (DGMR) outperforms existing methods in terms of both quantitative metrics and expert meteorologist evaluations. Comparative analysis against PySTEPS (a widely used probabilistic nowcasting system), UNet, and an axial attention model reveals DGMR's superiority. The critical success index (CSI) demonstrates significantly improved location accuracy across various rainfall thresholds. The radially averaged power spectral density (PSD) shows that DGMR, unlike UNet and the axial attention model, preserves the small- and medium-scale precipitation variability found in the observations, avoiding the blurring effect seen in other deep learning models. The continuous ranked probability score (CRPS), used for probabilistic verification, shows DGMR’s superior performance across multiple spatial aggregations, especially for maximum precipitation rates. A simple recalibration post-processing technique further enhances DGMR's performance. Economic value analysis, using a decision-analytic model and considering multiple rainfall accumulation levels, reveals that DGMR provides the highest economic value compared to the baselines. Crucially, a double-blind evaluation by 56 expert meteorologists from the Met Office showed a strong preference for DGMR (89% and 90% for medium and heavy rainfall respectively) for both accuracy and usefulness, highlighting its superiority in providing valuable physical insights not captured by alternative methods.

The findings demonstrate that the proposed deep generative model offers a significant advance in precipitation nowcasting. The superior performance across statistical, economic, and cognitive evaluation metrics highlights its potential for operational use. The ability of DGMR to capture both the spatial and temporal variability of precipitation, while maintaining a realistic representation of uncertainty, is a key advantage over existing deep learning methods that often resort to blurring to manage uncertainty. The expert meteorologist evaluation provides strong support for the model's practical value and its ability to provide actionable insights for decision-making. The model's speed and efficiency make it suitable for real-time operational applications.

This study presents a novel deep generative model for precipitation nowcasting that significantly outperforms existing methods. Using a combination of statistical, economic, and cognitive evaluations, the superior performance and practical value of the model are demonstrated. While challenges remain in predicting heavy rainfall at long lead times, this work lays a strong foundation for future improvements and highlights the importance of integrating machine learning and environmental science for improved forecasting.

While the deep generative model shows significant improvements, limitations remain. Predicting heavy precipitation accurately at longer lead times remains challenging. The model's performance is highly dependent on the quality and quantity of training data. Further research is needed to explore the generalizability of the model across different geographical regions and climatic conditions.