Alaoui et al. (2026) A novel deep learning semi-observational analog framework for diagnosing convection over Morocco using satellite imagery and ERA5-reanalysis

Identification

• Journal: Atmospheric Research
• Year: 2026
• Date: 2026-04-09
• Authors: Badreddine Alaoui, Driss Bari, Chakib Bounoun
• DOI: 10.1016/j.atmosres.2026.108969

Research Groups

• National Center of Meteorological Research, General Directorate of Meteorology, Casablanca, Morocco
• Computer Science, School of Information Sciences, Rabat, Morocco

Short Summary

This study introduces a novel semi-observational analog framework (AnOb) that leverages deep learning and integrates satellite imagery with ERA5 reanalysis data to diagnose convection parameters over Morocco. The AnOb system demonstrates enhanced performance compared to climatology and persistence, showing promising potential for detecting severe convective events.

Objective

• To develop and evaluate a new semi-observational analog approach (AnOb) for diagnosing convection-related parameters by integrating deep learning computer vision algorithms with high-resolution satellite imagery and ERA5 reanalysis data.

Study Configuration

• Spatial Scale: 40°N–20°N and 20°E–4°W, covering Morocco.
• Temporal Scale: Trained over a four-year period (2020–2023) for lead times of 12 hours, 15 hours, and 18 hours, with 2024 reserved for testing.

Methodology and Data

• Models used: Fine-tuned pre-trained deep learning architectures: VGG16, Xception, and Inception-ResNetV2.
• Data sources: High-resolution RGB convection satellite imagery (EUMETSAT) and ERA5 reanalysis data for Convective Available Potential Energy (CAPE), Total Column Water Vapor (TCWV), Vertical Velocity (W600), and Lifted Index (LI).

Main Results

• The AnOb system exhibited enhanced performance compared to climatology and persistence for both the full test period and extreme convection events.
• Hourly mean biases for convection parameters were:
  • CAPE: between -200 J/kg and 200 J/kg.
  • Lifted Index (LI): 0 K to 1 K.
  • Vertical Velocity (W) at 600 hPa: -0.3 Pa/s to 0.3 Pa/s.
• For convection detection, the system achieved:
  • Probability of Detection (POD): 0.6
  • False Alarm Ratio (FAR): 0.4
  • Brier Score (BS): 0.4
  • Continuous Ranked Probability Score (CRPS): 0.4
• These metrics indicate a promising potential for detecting severe convective events.

Contributions

• Introduction of a novel deep learning semi-observational analog framework (AnOb) for diagnosing convection, which moves beyond traditional point-wise analog approaches.
• Integration of deep learning computer vision algorithms (VGG16, Xception, Inception-ResNetV2) to assess weather state similarity from satellite imagery.
• Combination of high-resolution RGB satellite imagery with ERA5 reanalysis data for a more comprehensive characterization of convective processes.
• Development of a cost-effective, high-performance complementary tool for operational meteorological applications, particularly for regions with complex topography like Morocco.

Funding

Not specified in the provided text.

Citation

@article{Alaoui2026novel,
  author = {Alaoui, Badreddine and Bari, Driss and Bounoun, Chakib},
  title = {A novel deep learning semi-observational analog framework for diagnosing convection over Morocco using satellite imagery and ERA5-reanalysis},
  journal = {Atmospheric Research},
  year = {2026},
  doi = {10.1016/j.atmosres.2026.108969},
  url = {https://doi.org/10.1016/j.atmosres.2026.108969}
}