Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning Accurate Storm-Scale Evolution from Observations

Jaideep Pathak, Mohammad Shoaib Abbas, Peter Harrington, Zeyuan Hu, Noah Brenowitz, Suman Ravuri, Alberto Carpentieri, Jussi Leinonen, Corey Adams, Oliver Hennigh, Nicholas Geneva, Dale Durran, Mike Pritchard

TL;DR

Stormscope presents an observation-driven, diffusion-transformer approach to kilometer-scale storm forecasting using GOES-16 ABI and MRMS data. It delivers 10-minute, 6 km forecasts with ensemble trajectories that quantify uncertainty and are competitive with HRRR up to 6 hours, even without heavy data assimilation. Across nearcasting and nowcasting horizons, Stormscope shows deterministic and probabilistic gains over strong baselines, particularly in radar reflectivity and infrared channels, while leveraging minimal synoptic conditioning. The work demonstrates scalable training and inference through domain parallelism and suggests a path to democratize mesoscale forecasting in regions lacking extensive modeling infrastructure.

Abstract

Accurate short-term prediction of clouds and precipitation is critical for severe weather warnings, aviation safety, and renewable energy operations. Forecasts at this timescale are provided by numerical weather models and extrapolation methods, both of which have limitations. Mesoscale numerical weather prediction models provide skillful forecasts at these scales but require significant modeling expertise and computational infrastructure, which limits their accessibility. Extrapolation-based methods are computationally lightweight but degrade rapidly beyond 1-2 hours. This presents an opportunity for data-driven forecasting directly from observations using geostationary satellites and ground-based radar, which provide high-frequency, high-resolution observations that capture mesoscale atmospheric evolution. We introduce Stormscope, a family of transformer-based generative diffusion models trained on high-resolution, multi-band geostationary satellite imagery and ground-based weather radar over the continental United States. Stormscope produces forecasts at a temporal resolution of 10 minutes and 6 kilometer spatial resolution, which are competitive with state-of-the-art mesoscale NWP models for lead times up to 6 hours. Its generative architecture enables large ensemble forecasts of explicit mesoscale dynamics for robust uncertainty quantification. Evaluated against extrapolation methods and operational mesoscale NWP models, Stormscope achieves leading performance on standard deterministic and probabilistic verification metrics across forecast horizons from 1 to 6 hours. As Stormscope relies on globally available satellite observations (and radar where available), it offers a pathway to extend skillful mesoscale forecasting to oceanic regions and countries without strong operational mesoscale modeling programs.

Learning Accurate Storm-Scale Evolution from Observations

TL;DR

Stormscope presents an observation-driven, diffusion-transformer approach to kilometer-scale storm forecasting using GOES-16 ABI and MRMS data. It delivers 10-minute, 6 km forecasts with ensemble trajectories that quantify uncertainty and are competitive with HRRR up to 6 hours, even without heavy data assimilation. Across nearcasting and nowcasting horizons, Stormscope shows deterministic and probabilistic gains over strong baselines, particularly in radar reflectivity and infrared channels, while leveraging minimal synoptic conditioning. The work demonstrates scalable training and inference through domain parallelism and suggests a path to democratize mesoscale forecasting in regions lacking extensive modeling infrastructure.

Abstract

Accurate short-term prediction of clouds and precipitation is critical for severe weather warnings, aviation safety, and renewable energy operations. Forecasts at this timescale are provided by numerical weather models and extrapolation methods, both of which have limitations. Mesoscale numerical weather prediction models provide skillful forecasts at these scales but require significant modeling expertise and computational infrastructure, which limits their accessibility. Extrapolation-based methods are computationally lightweight but degrade rapidly beyond 1-2 hours. This presents an opportunity for data-driven forecasting directly from observations using geostationary satellites and ground-based radar, which provide high-frequency, high-resolution observations that capture mesoscale atmospheric evolution. We introduce Stormscope, a family of transformer-based generative diffusion models trained on high-resolution, multi-band geostationary satellite imagery and ground-based weather radar over the continental United States. Stormscope produces forecasts at a temporal resolution of 10 minutes and 6 kilometer spatial resolution, which are competitive with state-of-the-art mesoscale NWP models for lead times up to 6 hours. Its generative architecture enables large ensemble forecasts of explicit mesoscale dynamics for robust uncertainty quantification. Evaluated against extrapolation methods and operational mesoscale NWP models, Stormscope achieves leading performance on standard deterministic and probabilistic verification metrics across forecast horizons from 1 to 6 hours. As Stormscope relies on globally available satellite observations (and radar where available), it offers a pathway to extend skillful mesoscale forecasting to oceanic regions and countries without strong operational mesoscale modeling programs.
Paper Structure (22 sections, 1 equation, 25 figures, 2 tables)

This paper contains 22 sections, 1 equation, 25 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Results
  3. Qualitative Evaluation
  4. Quantitative Evaluation
  5. Nearcasting (0-12 hr)
  6. Nowcasting (0-2hr)
  7. Conclusion
  8. Methods
  9. Data
  10. Satellite and Radar Observations
  11. Numerical Weather Prediction (NWP) Conditioning
  12. Model Details
  13. Nowcasting and Nearcasting Approaches
  14. Model Architecture
  15. Training and Inference
  16. ...and 7 more sections

Figures (25)

  • Figure 1: An example forecast of composited visible satellite radiance fields from Stormscope visualized along with the corresponding verification data from the GOES satellite observation. Panels (a), (c), (e), and (g) show the Stormscope forecasts at 10 min, 120 min, 360 min, and 720 min lead times with the corresponding verification visualized in panels (b), (d), (f), and (h) respectively. The color visualization is composed of a linear combination of the visible radiance forecasts (a, c, e) and observations (b, d, f) using the Blue (0.47), Red (0.64), and Veggie (0.86) channels. The initialization timestamp of this forecast was 06-25-2024 at 12:00 UTC. See Fig. \ref{['fig:visualization_ir']} and Fig. \ref{['fig:visualization_mrms']} for the corresponding IR channel forecast and the radar forecast for the same initialization time.
  • Figure 2: An example forecast from Stormscope's IR channel measuring 10.35 GOES brightness temperature visualized along with the corresponding verification data. Panels (a), (c), (e), and (g) show the Stormscope forecasts at 10 min, 120 min, 360 min and 720 min lead times with the corresponding verification visualized in panels (b), (d), (f), and (h) respectively. The initialization timestamp of this forecast was 2024-06-25 at 12:00 UTC. See Fig. \ref{['fig:visualization_viz']} and Fig. \ref{['fig:visualization_mrms']} for the corresponding composite visible radiance forecast and the radar forecast for the same initialization time respectively.
  • Figure 3: Comparison of Stormscope radar reflectivity forecasts and verification data across the contiguous United States. Panels (a), (c), (e), and (g) display the Stormscope reflectivity forecasts at lead times of 10 min, 120 min, 360 min, and 720 min, respectively. The corresponding radar verification data for each lead time is shown in panels (b), (d), (f), and (h). The forecast was initialized on 2024-06-25 at 12:00 UTC. The color scale at the bottom indicates radar reflectivity in decibels relative to 1^6^3(dBZ), with values ranging from 0 to 60 dBZ, highlighting the spatial evolution and intensity of precipitation systems.
  • Figure 4: Comparison of Stormscope ensemble forecasts, verification data, and HRRR simulated brightness temperature (SBT) for a developing Mesoscale Convective System (MCS) over the central United States. The forecast initialization date is 05-15-2024 at 00:00 UTC. The rows display the evolution of the system at lead times of 1, 3, and 6 hours. The first three columns show individual ensemble members (Ens 1–3) from the Stormscope model, illustrating the predicted spatial distribution and intensity of the MCS. The fourth column provides the GOES verification data (Clean IR window channel, 10.35) for the corresponding times. The final column shows the HRRR model’s simulated brightness temperature forecast. Note that while Stormscope and verification data utilize the 10.35 channel, the HRRR SBT is based on a 10.7$\mu m$ operator; despite this slight wavelength difference, HRRR serves as a baseline numerical forecast. The color scale at the bottom indicates brightness temperature in Kelvin (K), where lower temperatures correspond to higher cloud tops and more intense convective activity.
  • Figure 5: Comparison of radar reflectivity forecasts at different lead times for a forecast initialized on 3-14-2024 at 00:00 UTC over the central United States. The columns represent (left) the Stormscope forecast, (middle) the verification observations, and (right) the pySTEPS forecast (see section \ref{['sec:nowcast']}). The rows correspond to forecast lead times of: (a–c) 10 min, (d–f) 20 min, (g–i) 60 min, and (j–l) 120 min.
  • ...and 20 more figures