Cross-regional generalization of satellite-based radar synthesis with deep learning

For many applications and regions, radar coverage remains a challenge in weather prediction. Remote regions, oceanic areas, and mountain-shadowed terrain suffer from sparse or absent radar observations, and even where infrastructure exists, data accessibility and downtime create further gaps. Satellite imagery offers a promising opportunity to fill these spatial and temporal voids. Meteosat Third Generation (MTG) satellites provide unprecedented spatial resolution and temporal revisit rates across Europe and Africa, making them particularly well-suited for learning continuous radar-like reflectivity fields.

We present a UNet-based approach to synthesize 2D radar composites from MTG satellite channels and lightning data (LINET). Our architecture employs wavelet decomposition for multi-scale feature extraction, going beyond standard lowpass filtering to better capture the spatial structure of precipitation. To preserve the sharpness of convective features, we employ a loss function that explicitly emphasizes sharp edges in the target and penalizes the synthesized output accordingly. The bottleneck combines Swin-style local window attention with a global pooled attention branch, enabling the model to simultaneously capture fine-grained local structure and mesoscale spatial context. We also employ an efficient channel attention mechanism for very wide feature maps.

A central focus of this work is the geographic transferability of the trained model. Using radar observations from Europe, we systematically evaluate cross-regional generalization by training and validating on subsets of countries and testing on held-out regions. This design allows us to assess how well learned satellite-to-radar mappings transfer across different climatic regimes and precipitation characteristics, with implications for deploying such models in radar-sparse or radar-free regions globally.