Integrating spherical geometry and spectral dynamics for improved satellite-based cloud cover forecasting

This is the authors' abstract. We don't add key points for this paper.

As a key regulator of Earth’s energy balance, accurate cloud cover forecasting is essential for assessing radiative forcing and precipitation. However, recent deep-learning-based forecasting models often neglect the intrinsic spherical geometry of the Earth and the long-term seasonal dynamics of cloud evolution. To bridge this gap, this paper proposes the Spherical-Spectral Trend Network (S 2 T-Net), a unified geospatial–temporal framework that reconciles spherical geometry with macro-scale atmospheric dynamics. Specifically, we first implement a Spherical Geometric Embedding (SGE) method, which maps pixel-wise geographic coordinates onto a 3D unit sphere via trigonometric transformations. By injecting explicit geographical priors derived from both latitude and longitude, SGE enables the network to account for view-angle-dependent spatial anisotropy and scale variation under geostationary observation while preserving the native satellite raster. Building upon this, we develop the Frequency-Domain Evolution Distillation (FDED) mechanism. FDED utilizes a learnable temporal memory pool and parameterized spectral filters to capture periodic oscillatory features of cloud movement. By explicitly disentangling spectral amplitude and phase, FDED independently preserves structural integrity and models kinematic advection. This iterative distillation of salient spectral trends provides physically consistent priors to guide current observations, allowing the model to perceive macro-scale trajectories and seasonal cycles typically obscured in standard spatial-domain processing. Consequently, S 2 T-Net effectively integrates static spherical geographic priors with dynamic cloud evolution. Evaluated on the full-disk Fengyun satellite dataset and eight high-resolution sub-regions with diverse climatic conditions, S 2 T-Net demonstrates strong and robust spatiotemporal forecasting performance across multiple prediction horizons. It achieves the highest accuracy among the evaluated methods for long-term full-disk forecasting at the 12-hour horizon, while also showing competitive F1 and IoU performance across regional evaluations. The source code is publicly available at: https://github.com/mxx123321/cloud_forecasting