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

February 14, 2026

A paper entitled “Deterministic nowcasting of geostationary satellite temperature using 3D U-Net diffusion model” was recently published in Scientific Reports. The study was led by Vesta Afzali Gorooh, with co-authors Luca Delle Monache, Duncan Axisa, Agniv Sengupta, Zhenhai Zhang, and F. Martin Ralph [all at the Center for Western Weather and Water Extremes (CW3E) at Scripps Institution of Oceanography, University of California, San Diego]. The work supports the Advanced Precipitation and Streamflow Prediction priority identified in CW3E’s Strategic Plan by demonstrating a technique to better leverage information from geostationary satellites for hazard prediction.

The paper offers a generative modeling framework for satellite-based cloud nowcasting that couples a denoising diffusion probabilistic model with a 3D U-Net backbone to forecast infrared (IR) brightness temperatures from the Spinning Enhanced Visible and Infrared Imager (SEVIRI) sensor onboard the METEOSAT Second Generation (MSG) satellite with Indian Ocean Data Coverage (IODC). The model ingests six hours of IR history and produces six-hour nowcasts at 15-minute temporal resolution, providing accurate short-lead forecasts of convective cloud evolution.

Deterministic evaluation over an independent July–September 2022 test period shows that the 3D U-Net diffusion model systematically outperforms baseline approaches (3D U-Net, ConvLSTM, and Optical Flow extrapolation) across a broad range of lead times, with a particularly strong advantage out to approximately two hours. The diffusion architecture yields lower errors and higher correlations, achieves the highest Structural Similarity Index (SSIM) at all leads, and attains the lowest Continuous Ranked Probability Score (CRPS) when evaluated as a nine-member diffusion ensemble (Figure 1), demonstrating both improved accuracy and sharper, more reliable brightness temperature (Tb) fields relative to conventional deterministic and extrapolation benchmarks.

Figure 1. Structural Similarity Index (SSIM, top) and Continuous Ranked Probability Score (CRPS, bottom) as a function of lead time for 3D U-Net model (blue) and 3D U-Net Diffusion (red), ConvLSTM (green), and Optical Flow (yellow) with 95% confidence intervals from 1000 bootstrapping. Figure 4 from Afzali Gorooh et al. (2026).

Figure 2 presents spectral and spatial diagnostics comparing the diffusion model with baseline deep learning and Optical Flow methods. The diffusion framework preserves substantially more high-frequency variance while avoiding the purely texture-advective behavior characteristic of Optical Flow. Case studies also highlight more coherent cloud-top structures, sharper brightness-temperature gradients, and improved localization of evolving cold features, resulting in more physically realistic satellite scenes at short and intermediate lead times.

Overall, the results demonstrate that diffusion-based generative modeling offers a robust pathway to advance satellite nowcasting beyond conventional deterministic deep learning approaches. By improving both structural realism and predictive skill, this framework offers strong potential for next-generation operational applications in convective monitoring, hazard forecasting, and decision support systems.

Figure 2. Radially averaged power spectral density (RAPSD) of observed and forecasted Tbs for 1–6 h lead times (a–f). Spectra are averaged over the test period (July–September 2022) for regions with Tbs less than 275 K. The black dashed line shows observations, while red, blue, green, and yellow lines represent 3D U-Net Diffusion, 3D U-Net, ConvLSTM, and Optical Flow forecasts, respectively. Figure 6 from Afzali Gorooh et al. (2026).

Citation:

Afzali Gorooh, V., Delle Monache, L., Axisa, D., Sengupta, A., Zhang, Z., & Ralph, F. M. (2026). Deterministic nowcasting of geostationary satellite infrared brightness temperature using 3D U-Net diffusion model. Scientific Reports, 16, 4191. https://doi.org/10.1038/s41598-025-34207-9