论文详情
Paper Details

Accurate nowcasting of tropical cyclone satellite cloud imagery (TCSCI) is critical for disaster prevention and mitigation, but it remains challenging due to coupled uncertainties and multiscale dynamics in TCSCI evolution. While data-driven deep learning (DL)-based deterministic models mitigate biases inherent in traditional physics-driven approaches, their deterministic nature inherently limits uncertainty quantification. To address this, we propose TCSat-Diff, an ensemble framework that integrates spectral deterministic prediction with a frequency-adaptive probabilistic diffusion model. Specifically, we introduce a physics-inspired Reynolds decomposition strategy that separates TCSCI evolution into a large-scale trend flow and a small-scale fluctuating flow. The trend flow is predicted using the adaptive Fourier neural operator (AFNO) with mean-squared error (MSE) loss for smooth large-scale forecasts, while the fluctuating flow is captured by a conditional latent diffusion model (CLDM) for highly uncertain small-scale fluctuations. To further improve frequency-aware uncertainty modeling, we develop a signal-to-noise ratio (SNR)-guided noise frequency modulation (NFM) module and a time–frequency response (TiFR) module into the CLDM, enabling adaptive spectral focus during denoising. Experiments on 1099 tropical cyclones (TCs) from the Northwest Pacific demonstrate that TCSat-Diff outperforms previous deterministic and probabilistic baselines. Ablation studies confirm the effectiveness of the decomposition strategy and frequency-aware modules. Notably, cross-regional generalization evaluation on 480 TCs from the Southern Hemisphere reveals excellent generalization capability across different hemispheres. These results highlight the potential of TCSat-Diff for global TCSCI nowcasting, supporting early warning and risk reduction for TC hazards.