Deterministic nowcasting of geostationary satellite infrared brightness temperature using 3D U-Net diffusion model.

Area of Science:

• Atmospheric Science
• Remote Sensing
• Artificial Intelligence

Background:

• Accurate nowcasting of infrared brightness temperatures (Tb) is crucial for weather prediction.
• Existing methods like deep learning and optical flow have limitations in capturing complex atmospheric dynamics.

Purpose of the Study:

• To develop and evaluate a generative modeling approach for satellite-based infrared nowcasting.
• To improve the accuracy and structural fidelity of short-term (up to 6 hours) Tb forecasts.

Main Methods:

• Coupled a denoising diffusion probabilistic model with a 3D U-Net backbone.
• Utilized SEVIRI geostationary satellite observations (10.8 μm).
• Benchmarked against 3D U-Net, ConvLSTM, and Optical Flow using deterministic, perceptual, and probabilistic diagnostics.

Main Results:

• The diffusion model demonstrated lower errors and higher correlation compared to all baselines across most forecast lead times.
• Achieved superior performance in structural similarity (SSIM) and continuous ranked probability score (CRPS).
• Showed improved retention of high-frequency variance and more coherent structures in forecasts.

Conclusions:

• The proposed diffusion model coupled with 3D U-Net significantly enhances the quality of satellite infrared nowcasts.
• It outperforms traditional extrapolation and deep learning baselines, especially for short to intermediate lead times.
• The model provides more structurally faithful and accurate predictions of atmospheric phenomena.