Bosc et al. (2026) Predicting thunderstorm risk probability at very short time range using deep learning

Identification

• Journal: Natural hazards and earth system sciences
• Year: 2026
• Date: 2026-03-31
• Authors: Mélanie Bosc, Adrien Chan-Hon-Tong, Aurélie Bouchard, Dominique Béréziat
• DOI: 10.5194/nhess-26-1603-2026

Research Groups

• DPHY, ONERA, Université Paris-Saclay, Palaiseau, France
• DTIS, ONERA, Université Paris-Saclay, Palaiseau, France
• Sorbonne Université, CNRS, LIP6, Paris, France

Short Summary

This study develops a deep learning methodology to predict thunderstorm risk probability at very short time ranges (every 5 minutes up to 1 hour ahead) for aviation safety. It utilizes an adapted Convolutional Neural Network with attention mechanisms, fed by satellite observations and Numerical Weather Prediction outputs, to generate well-calibrated lightning risk maps.

Objective

• To develop a deep learning-based methodology for forecasting calibrated lightning risk probability maps at very short time ranges (up to 1 hour, every 5 minutes) at mesoscale, without relying on radar data, to enhance air safety.

Study Configuration

• Spatial Scale:
  • Studied area: [15° N; 40° N] latitude and [100° W; 65° W] longitude (centered over the Gulf of Mexico and Florida).
  • Training/Testing area: 256 × 256 pixels, corresponding to approximately [17.3° N; 37.7° N] latitude and [93° W; 72° W] longitude.
  • Spatial resolution: 0.08° × 0.08° (approximately 8.8 km).
• Temporal Scale:
  • Forecast horizons: Every 5 minutes, up to 1 hour (60 minutes).
  • Input sequence length: 6 timesteps (30 minutes).
  • Data collection period: January, February, and December (2020-2023), from 00:00 to 05:00 UTC.
  • Satellite data temporal resolution: 5 minutes (ABI), 20 seconds aggregated to 5 minutes (GLM).
  • NWP data temporal resolution: 3 hours (GFS), oversampled and reused for 5-minute timesteps.

Methodology and Data

• Models used:
  • Primary: ED-DRAP (Encoder-Decoder Deep Residual Attention Prediction Network) with spatial and sequential attention mechanisms.
  • Comparison models: ConvLSTM, PredRNN, U-Net, Persistence.
• Data sources:
  • Satellite observations (GOES-R/GOES-16, NOAA):
    • Advanced Baseline Imager (ABI): Brightness Temperature (BT) from the 13th infrared band (10.3 µm).
    • Geostationary Lightning Mapper (GLM): Groups product (aggregated over 5-minute intervals).
  • Numerical Weather Prediction (NWP) outputs (Global Forecast System - GFS, NCEP):
    • Lifted Index (LI): bestLI (lowest LI value across altitude levels).
    • Relative Humidity (RH): maxRH (maximum RH value across altitude levels).

Main Results

• The proposed methodology successfully predicts lightning risk probability maps with well-calibrated outputs.
• Achieved an F1 score of 0.65 for 5-minute predictions and 0.5 for 30-minute predictions (using a 0.5 threshold).
• Demonstrated very low Expected Calibration Error (ECE) of less than 10% for 5-minute predictions, and around 10% for 50-minute predictions, significantly outperforming other models (ConvLSTM, PredRNN, U-Net) which had ECE values of at least 30%.
• ED-DRAP consistently showed a better overall precision-recall balance and higher F1 scores across all forecast horizons compared to ConvLSTM and PredRNN.
• Generated physically interpretable lightning risk probability maps, allowing users to adjust thresholds based on desired precision and recall (e.g., a 0.05 threshold results in only 5% missed lightning; a 0.2-0.3 threshold yields approximately 80% recall and 35% precision).
• The method demonstrated robustness by generalizing well to a new geographical region (Panama) and different seasonal/diurnal conditions, maintaining comparable F1-Scores and calibrated probabilistic maps.

Contributions

• Developed a novel deep learning methodology for very short-term (up to 1 hour) calibrated lightning risk probability forecasting, specifically tailored for aviation safety.
• Successfully adapted and applied the ED-DRAP network, incorporating spatial and sequential attention mechanisms, to effectively process spatio-temporal meteorological data for lightning prediction.
• Addressed the critical issue of output calibration in AI-driven weather forecasting, achieving significantly lower calibration errors compared to existing methods.
• Effectively managed the class imbalance problem inherent in lightning data through the strategic combination of Cross-Entropy and Dice Loss functions.
• Integrated heterogeneous meteorological data sources (satellite observations and NWP outputs) to provide comprehensive input for the deep learning model, without reliance on radar data.
• Demonstrated the method's robustness and generalization capability across different geographical regions and time periods, highlighting its potential for operational deployment.

Funding

• ALBATROS project, European Union Horizon Europe (Grant Agreement No. 101077071).
• HORIZON EUROPE Climate, Energy and Mobility (grant no. 101077071).

Citation

@article{Bosc2026Predicting,
  author = {Bosc, Mélanie and Chan-Hon-Tong, Adrien and Bouchard, Aurélie and Béréziat, Dominique},
  title = {Predicting thunderstorm risk probability at very short time range using deep learning},
  journal = {Natural hazards and earth system sciences},
  year = {2026},
  doi = {10.5194/nhess-26-1603-2026},
  url = {https://doi.org/10.5194/nhess-26-1603-2026}
}