Reliable cloud motion estimation and interpolation of cyclones through fractional order variational technique

Abstract Accurate short-term cyclone state prediction is essential for minimizing damage and enhancing disaster preparedness. This study introduces a data-driven framework that utilizes satellite imagery and temporal correlations of sequential atmospheric patterns to predict cyclone motion. The temporal correlation is performed using the optical flow, which is further used to support the early warning system. Optical flow is employed to capture motion fields, which are used to identify high-alert regions of interest through velocity components along x - and y -axes. For optical flow estimation, a novel variational model is proposed that integrates a non-local weighted parameter, Charbonnier norm, and curvature-domain descriptor-based data term with a fractional-order derivative-based smoothness term. The non-local weighted account dynamic discontinuities in the motion field, while the curvature-domain descriptor mitigates illumination inconsistencies across satellite frames. Minimization and discretization of the variational functional are achieved using Euler–Lagrange equations and the Grünwald–Letnikov fractional derivative. Experiments are performed on various types of cyclone datasets from Himawari-8 and INSAT-3D geostationary satellites, covering both the Western North Pacific and North Indian Ocean basins, that demonstrate the robustness of the proposed model. Since velocity components alone are insufficient for analyzing cloud motion under intermediate weather conditions, a sigmoid-based interpolation technique is further employed. In addition to that, a fractional anisotropic diffusion-based denoising component is seamlessly integrated into the proposed optical flow and denoising framework to effectively suppress noise and preserve structural coherence across consecutive frames. The proposed model is validated under varying noise conditions, and statistical analysis is done on a synthetically illuminated satellite dataset variant that confirms the model’s superiority. An ablation study highlights the contribution of each algorithmic component.