Accurate observations of Ocean Surface Wind Speed (OSWS) are vital for predicting extreme weather and understanding ocean–atmosphere interactions. However, spaceborne sensors (e.g., ASCAT, SMAP) often experience data loss due to harsh weather and instrument malfunctions. Existing inpainting methods often rely on reanalysis data that is released with delays, which restricts their real-time capability. Additionally, deep-learning-based methods, such as Transformers, face challenges due to their high computational complexity. To address these challenges, we present the Multimodal Wind Speed Inpainting Dataset (MWSID), which integrates 12 auxiliary forecasting variables to support real-time OSWS inpainting. Based on MWSID, we propose the MMamba framework, combining the Multimodal Feature Extraction module, which uses mutual information (MI) theory to optimize feature selection, and the OSWS Reconstruction module, which employs Attention-Mamba-2 within a Residual-in-Residual-Dense architecture for efficient OSWS inpainting. Experiments show that MMamba outperforms MambaIR (state-of-the-art) with an RMSE of 0.5481 m/s and an SSIM of 0.9820, significantly reducing RMSE by 21.10% over Kriging and 8.22% over MambaIR in high-winds (>15 m/s). We further introduce MMamba-L, a lightweight 0.22M-parameter variant suitable for resource-limited devices. These contributions make MMamba and MWSID powerful tools for OSWS inpainting, benefiting extreme weather prediction and oceanographic research.

Tropical cyclones (TCs) rank among the most destructive natural hazards globally, with core damaging potential originating from regions of intense wind shear and steep wind speed gradients within the eyewall and spiral rainbands. Accurately characterizing these fine-scale structural features is therefore critical for understanding TC intensity evolution, wind hazard distribution, and disaster mitigation. Recently, the deep learning-based downscaling methods have shown significant advantages in efficiently obtaining high-resolution wind field distributions. However, existing methods are mainly used to downscale general wind fields, and research on downscaling extreme wind field events remains limited. There are two main difficulties in downscaling TC wind fields. The first one is that high-quality datasets for TC wind fields are scarce; the other is that general deep learning frameworks lack the ability to capture the dynamic characteristics of TCs. Consequently, this study proposes a novel deep learning framework, CycloneWind, for downscaling TC surface wind fields: (1) a high-quality dataset is constructed by integrating Cyclobs satellite observations with ERA5 reanalysis data, incorporating auxiliary variables like low cloud cover, surface pressure, and top-of-atmosphere incident solar radiation; (2) we propose CycloneWind, a dynamically constrained Transformer-based architecture incorporating three wind field dynamical operators, along with a wind dynamics-constrained loss function formulated to enforce consistency in wind divergence and vorticity; (3) an Adaptive Dynamics-Guided Block (ADGB) is designed to explicitly encode TC rotational dynamics using wind shear detection and wind vortex diffusion operators; (4) Filtering Transformer Layers (FTLs) with high-frequency filtering operators are used for modeling wind field small-scale details. Experimental results demonstrate that CycloneWind successfully achieves an 8-fold spatial resolution reconstruction in TC regions. Compared to the best-performing baseline model, CycloneWind reduces the Root Mean Square Error (RMSE) for the U and V wind components by 9.6% and 4.9%, respectively. More significantly, it achieves substantial improvements of 23.0%, 22.6%, and 20.5% in key dynamical metrics such as divergence difference, vorticity difference, and direction cosine dissimilarity.