With the advancement of meteorological instruments, increasingly abundant data has become available.
However, due to intrinsic limitations such as environmental sensitivity and orbital constraints, raw observations often contain temporal or spatial gaps, making it essential to employ data synthesis techniques to fill in missing information.
Existing approaches typically focus on single-variable or single-region tasks and primarily rely on deterministic modeling.
This limits unified synthesis across variables and regions, overlooks cross-variable complementarity, and often leads to over-smoothed results.
To address these challenges, we introduce SynWeather, the first dataset designed for
Unified Multi-region and Multi-variable Weather Observation Data Synthesis.
SynWeather covers four representative regions—the Continental United States, Europe, East Asia, and Tropical Cyclone areas—and provides high-resolution observations of key weather variables, including Composite Radar Reflectivity, Hourly Precipitation, Visible Light, and Microwave Brightness Temperature.
In addition, we present SynWeatherDiff, a general and probabilistic weather synthesis model built upon the Diffusion Transformer framework to mitigate the over-smoothing issue.
Experiments on the SynWeather dataset demonstrate that our model outperforms both task-specific and general baselines.
Moreover, SynWeatherDiff produces fine-grained and accurate estimates, especially in high-value weather regions.
Through this dataset and baseline model, we aim to advance downstream meteorological tasks and promote the development of general models for weather variable synthesis.
