Skillful Nowcasting of Convective Clouds With a Cascade Diffusion Model

TL;DR

This work tackles the challenge of nowcasting convective clouds from satellite data by proposing SATcast, a diffusion-based cascade model conditioned on FuXi-predicted atmospheric fields and historical FY-4A imagery. The two-phase autoregressive cascade (predicting $T+1$ to $T+4$ in phase 1 and $T+5$ to $T+8$ in phase 2, with further extension by reusing outputs) mitigates error accumulation and extends skill to $24$ hours, outperforming persistence and optical flow baselines on a large test set. Ablation studies confirm the importance of multimodal conditioning, cascade structure, and fine-tuning, while permutation-importance analyses reveal the evolving contribution of satellite observations and FuXi variables over lead times. The model demonstrates strong generalization across channels and shows promise for operational nowcasting in data-sparse regions, potentially enabling probabilistic forecasts with future extensions.

Abstract

Accurate nowcasting of convective clouds from satellite imagery is essential for mitigating the impacts of meteorological disasters, especially in developing countries and remote regions with limited ground-based observations. Recent advances in deep learning have shown promise in video prediction; however, existing models frequently produce blurry results and exhibit reduced accuracy when forecasting physical fields. Here, we introduce SATcast, a diffusion model that leverages a cascade architecture and multimodal inputs for nowcasting cloud fields in satellite imagery. SATcast incorporates physical fields predicted by FuXi, a deep-learning weather model, alongside past satellite observations as conditional inputs to generate high-quality future cloud fields. Through comprehensive evaluation, SATcast outperforms conventional methods on multiple metrics, demonstrating its superior accuracy and robustness. Ablation studies underscore the importance of its multimodal design and the cascade architecture in achieving reliable predictions. Notably, SATcast maintains predictive skill for up to 24 hours, underscoring its potential for operational nowcasting applications.

Figures (6)

• Figure 1: Comparison of the RMSE, correlation coefficient, MSSSIM, and CSI spatially averaged over (86$^\circ$ to 150$^\circ$ E in longitude and 1$^\circ$ to 41$^\circ$ N in latitude. The results from persistence model, SATcast, and optical flow. The results using testing data from 2022 September to 2022 December. The threshold is 240 K in the calculation of CSI. The red shading represents the one standard deviation range of SATcast.
• Figure 2: Spatial distribution of brightness temperature from observation and SATcast predictions, shown at 4-hour time intervals from T+4 to T+24.
• Figure 3: Comparison of RMSE, correlation coefficient, MSSSIM, and CSI spatially averaged over (86$^\circ$ to 150$^\circ$ E in longitude and 1$^\circ$ to 41$^\circ$ N in latitude across five models: Persistence, SATcast, SATcast-NoC, SATcast-NoT, SATcast-NoF, and optical flow. The results are based on testing data from September to December in 2022. A threshold of 240 K is used in the calculation of ACC. The red shading represents the one standard deviation range of SATcast.
• Figure 4: Spatial distribution of brightness temperature at T+1, T+3, and T+8 from observation (first row), SATcast (second row), SATcast-NoC (third row), and SATcast-NoF predictions (fourth row), where T denotes the starting time of the prediction, the UTC time of T is 2022-09-14-06:00.
• Figure 5: Heatmap of N-MSE from the permutation feature importance. The x-axis represents the different features, and the y-axis represents the lead time. Panels are divided into a) higher importance and b) lower importance. FY-4A denotes the satellite imagery, and FuXi represents all 13 variables shown in the right panel. The three-digits codes on the y-axis of the right figure indicate pressure levels: 250 hPa, 500 hPa, 700 hPa, and 850 hPa. Variable abbreviations are as follows: z (geopotential), t (temperature), u (u component of wind), v (v component of wind), q (specific humidity), t2m (temperature at 2 meters), u10m (u component of wind at 10 meters), v10 (v component of wind at 10 meters), msl (mean sea-level pressure).