Data-driven Precipitation Nowcasting Using Satellite Imagery

TL;DR

This work tackles the challenge of real-time, high-resolution precipitation nowcasting in regions lacking radar infrastructure. It introduces the Neural Precipitation Model (NPM), a two-stage framework that predicts future satellite frames from geostationary imagery and then translates those predictions into radar-like precipitation maps, enhanced by day/hour positional embeddings and a spatio-temporal large-kernel attention backbone. The authors also release Sat2Rdr, a large geostationary satellite dataset with corresponding radar observations, enabling fair evaluation of radar-free nowcasting methods. Their results show that NPM outperforms several baselines on CSI metrics across lead times, demonstrates robustness in radar-sparse regions (including a North Korea flood case), and highlights the practical potential for real-time flood alerts in developing regions. Overall, this approach offers a scalable, satellite-only pathway toward affordable, real-time precipitation nowcasting with broad societal impact.

Abstract

Accurate precipitation forecasting is crucial for early warnings of disasters, such as floods and landslides. Traditional forecasts rely on ground-based radar systems, which are space-constrained and have high maintenance costs. Consequently, most developing countries depend on a global numerical model with low resolution, instead of operating their own radar systems. To mitigate this gap, we propose the Neural Precipitation Model (NPM), which uses global-scale geostationary satellite imagery. NPM predicts precipitation for up to six hours, with an update every hour. We take three key channels to discriminate rain clouds as input: infrared radiation (at a wavelength of 10.5 $μm$), upper- (6.3 $μm$), and lower- (7.3 $μm$) level water vapor channels. Additionally, NPM introduces positional encoders to capture seasonal and temporal patterns, accounting for variations in precipitation. Our experimental results demonstrate that NPM can predict rainfall in real-time with a resolution of 2 km. The code and dataset are available at https://github.com/seominseok0429/Data-driven-Precipitation-Nowcasting-Using-Satellite-Imagery.

Figures (6)

• Figure 1: (a) Our +6 hour forecasting results without radar. (b) NASA IMERG observation for March 18, 2024, Papua New Guinea flood case. (Note that NASA GPM IMERG-late run is accessible only 14 hours after observation.)
• Figure 2: Overview of NPM Architecture. First, the Satellite Prediction Model takes season-aware sampled satellite sequences and then predicts future frames. Second, the Satellite-to-Radar Model generates precipitation from predicted satellite sequences.
• Figure 3: Performance comparison of CSI 1 mm, CSI 4 mm, and CSI 8 mm by month. Categorical CSI (higher is better). CSI plots are for light (1 mm/h), moderate (4 mm/h), and heavy (8 mm/h) precipitation.
• Figure 4: Comparison between the Radar2Radar state-of-the-art method, PreDiff, and our Sat2Rdr approach. (a) shows ours, (b) shows PreDiff, and (c) shows radar-observed ground truths.
• Figure 5: Precipitation forecasting results of 2024-07-26 heavy rainfall case in North Korea. (a) is the prediction result of NPM, (b) is the global satellite-based precipitation data from NASA GPM IMERG-late run, and (c) is the observation data from KMA radar.