Deep learning for precipitation nowcasting: A survey from the perspective of time series forecasting

TL;DR

This survey comprehensively analyzes deep learning for precipitation nowcasting from a time-series forecasting perspective, detailing data sources (radar, satellites, and fusion), preprocessing, and objective/evaluation frameworks. It classifies models into recursive (non-adversarial and adversarial) and multiple-strategy (UNet, diffusion, transformer) categories, and compares representative approaches across benchmark datasets, highlighting trade-offs in sharpness, temporal dependency, and uncertainty quantification. Key contributions include a structured taxonomy, critical discussion of loss functions and metrics tailored to precipitation, and a synthesis of methodology and performance on public benchmarks with insights into practical challenges such as long lead times, multi-sensor fusion, and evaluation standardization. The authors also outline promising directions—physics-informed diffusion/transformer architectures, sensor fusion protocols, and standardized evaluation pipelines—that could accelerate robust, real-world nowcasting systems.

Abstract

Deep learning-based time series forecasting has dominated the short-term precipitation forecasting field with the help of its ability to estimate motion flow in high-resolution datasets. The growing interest in precipitation nowcasting offers substantial opportunities for the advancement of current forecasting technologies. Nevertheless, there has been a scarcity of in-depth surveys of time series precipitation forecasting using deep learning. Thus, this paper systemically reviews recent progress in time series precipitation forecasting models. Specifically, we investigate the following key points within background components, covering: i) preprocessing, ii) objective functions, and iii) evaluation metrics. We then categorize forecasting models into \textit{recursive} and \textit{multiple} strategies based on their approaches to predict future frames, investigate the impacts of models using the strategies, and performance assessments. Finally, we evaluate current deep learning-based models for precipitation forecasting on a public benchmark, discuss their limitations and challenges, and present some promising research directions. Our contribution lies in providing insights for a better understanding of time series precipitation forecasting and in aiding the development of robust AI solutions for the future.

Figures (8)

• Figure 1: Paradigms in precipitation nowcasting. We classify forecasting models at the first level based on two training strategies: the recursive strategy and the multiple strategy. While the recursive strategy predicts future time steps sequentially at each step $t$, the multiple strategy predicts future frames simultaneously. At the second sub-level, models of the recursive strategy are classified into non-adversarial and adversarial-based categories, and multiple strategy models are categorized into UNet, Diffusion, and Transformer.
• Figure 2: Visualization of the SEVIR dataset from April 30, 2019, at 18 UTC. The images depicting weather events captured over the contiguous US. (a) Vertically integrated liquid of NEXRAD radar. (b) GOES-16 satellite channel 2 visible (VIS). (c) GOES-16 satellite channel 9 water vapor (WV). (d) GOES-16 satellite channel 13 infrared (IR). (Source) The images were obtained from the SEVIR official page.
• Figure 3: Overview of the recursive strategy. Methods in the recursive strategy can be categorized into non-adversarial and adversarial. We group the applications based on the above category and then order them chronologically. N effectively learn the temporal dependency using recurrent frameworks. A realistically predict future frames based on GANs. The subcategories were classified based on the core keywords intended to address previous issues in precipitation nowcasting.
• Figure 4: Examples of the model architectures with the recursive strategy
• Figure 5: Overview of multiple strategy. Applications are categorized into UNet, Diffusion, and Transformer. Models are characterized based on the methods they report results on and are ordered chronologically. U effectively capture channel-wise dependency in multivariate input data. D realistically predict future frames. T robust long-term dependency. The subcategories were classified based on the core keywords intended to address previous issues in precipitation nowcasting.