A Spatial-temporal Deep Probabilistic Diffusion Model for Reliable Hail Nowcasting with Radar Echo Extrapolation

TL;DR

Hail nowcasting demands high-resolution, short-lead forecasts derived from radar signals. The authors introduce SteamCast, a Spatial-Temporal diffusion model that fuses radar echo data with a Spatiotemporal Encoding scheme within a Stable Diffusion framework to generate 30-minute nowcasts. Key contributions include a target/reference patch attention mechanism, SpEn for temporal and spatial conditioning, and an efficient subpatch processing strategy, achieving superior performance over state-of-the-art methods on multiple metrics. The work demonstrates practical improvements for disaster preparedness and precision agriculture, with future directions highlighting multi-source data integration for enhanced reliability.

Abstract

Hail nowcasting is a considerable contributor to meteorological disasters and there is a great need to mitigate its socioeconomic effects through precise forecast that has high resolution, long lead times and local details with large landscapes. Existing medium-range weather forecasting methods primarily rely on changes in upper air currents and cloud layers to predict precipitation events, such as heavy rainfall, which are unsuitable for hail nowcasting since it is mainly caused by low-altitude local strong convection associated with terrains. Additionally, radar captures the status of low cloud layers, such as water vapor, droplets, and ice crystals, providing rich signals suitable for hail nowcasting. To this end, we introduce a Spatial-Temporal gEnerAtive Model called SteamCast for hail nowcasting with radar echo extrapolation, it is a deep probabilistic diffusion model based on spatial-temporal representations including radar echoes as well as their position/time embeddings, which we trained on historical reanalysis archive from Yan'an Meteorological Bureau in China, where the crop yield like apple suffers greatly from hail damage. Considering the short-term nature of hail, SteamCast provides 30-minute nowcasts at 6-minute intervals for a single radar reflectivity variable, across 9 different vertical angles, on a latitude-longitude grid with approximately 1 km * 1 km resolution per pixel in Yan'an City, China. By successfully fusing the spatial-temporal features of radar echoes, SteamCast delivers competitive, and in some cases superior, results compared to other deep learning-based models such as PredRNN and VMRNN.

Figures (3)

• Figure 1: Comparison of our SteamCast in (c) with the DGM-based method in (a) and the physics-informed method in (b). Enhancing the capability of capturing spatial-temporal features is one of the most important aspects of reliable precipitation nowcasting. Methods (b) and (c) utilize a physical model and a well-designed spatial-temporal DGM to enhance the capability of capturing spatial-temporal features.
• Figure 2: SteamCast architecture details. SteamCast adopts a SD architecture with slight yet critical modifications to effectively and flexibly represent spatial-temporal radar echoes. Left: The pre-trained encoder in SD captures features from the target patch of radar echoes, the target patch consists of $M$ nowcasted time steps. Middle: In U-Net, we apply self-attention within the target patch to encourage target-to-target consistency. We also apply a cross-attention block between the target patch and $T$ reference patches to enhance reference-to-target consistency. In each attention block, SpEn is employed for the key and query, enabling the attention map to capture relative positions and time steps. Right: The pre-trained encoder in SD and projection layer capture features from $T$ reference patches, each patch contains $N$ historical time steps.
• Figure 3: Visualization results of hail nowcasting. Input and GT denote target patches of 5 historical and 5 future time steps, respectively. PredRNN, CMS, and VMRNN represent nowcasting results from the respective competitors. SpEn represents the hail nowcast of SteamCast. NoEmbd and TimeEmbd represent hail nowcasts from our two variants, as discussed in Sec. \ref{['sec:ablation']} of the ablation studies.