Satellite-based Offshore Wind Nowcasting: addressing spatio-temporal irregularities of scatterometer observations with Deep Learning

We present the first satellite-based wind forecast model, WindCastNet, and demonstrate that it outperforms state-of-the-art regional weather forecast models by lead times of up to ~2.5 hours in forecasts of 10-meter wind speed. Our model, WindCastNet, is a space-time aware convolutional long short-term memory network (ConvLSTM) that enables short-term wind nowcasting over offshore regions using exclusively satellite scatterometer observations (ASCAT and HSCAT) as input. Unlike geostationary satellites and ground radar, satellite scatterometers provide observations that are sparse and irregular in space and time, covering variable spatial domains and providing only a moderate training set size. 

WindCastNet overcomes these challenges by encoding spatial coverage masks, geographic coordinates, and inter-observation time intervals as explicit input channels, enabling the model to remain robust under varying observational configurations. From a training perspective, the limited availability of satellite data - approximately 13,000 usable overpasses between 2021 and 2025 - is addressed through a two-stage strategy: pretraining on ERA5 reanalysis fields to learn overall dynamics of 10-meter wind fields, followed by fine-tuning on scatterometer measurements. This transfer learning approach proves essential to achieve skillful forecasts, as demonstrated by the training diagnostics. Although the model is optimized for a 3-hour horizon, its recurrent architecture allows inference at longer lead times. 

Evaluation against HARMONIE-AROME (MEPS) over the North Sea shows 55% lower RMSE at 1h and 10% lower at 2h lead time, while HARMONIE-AROME exhibits lower forecast RMSE beyond 3h lead time. Compared to persistence, the model achieves 65%, 32%, and 21% RMSE reduction at 1h, 2h, and 3h, respectively. 
We further characterize the WindCastNet behavior across different meteorological regimes and discuss improvement potential, particularly at domain boundaries and for phenomena originating outside the training domain. Finally, we dicuss introducing physical constraints, outline pathways toward uncertainty, quantification and probabilistic extensions as natural next steps for operational integration.