Earth observation embeddings are effective sub-grid descriptors for probabilistic weather downscaling
Earth-observation embeddings improved station-level downscaling skill beyond hand-built terrain features.
The paper adds compressed 10 m TESSERA surface embeddings to a probabilistic downscaler using coarse ERA5 fields at about 25 km resolution. Across five regions, it reports CRPS skill gains of 11.5% for 2 m temperature and 6.2% for 10 m wind speed on stations held out across space and time. The authors say topography explains more of the temperature structure, while the embeddings add surface information for wind. The gains also held when the coarse input came from Aurora forecasts and when predicting at newly deployed stations with no regional history. ArXiv · AI/CL/LG's note
The paper adds compressed 10 m TESSERA surface embeddings to a probabilistic downscaler using coarse ERA5 fields at about 25 km resolution. Across five regions, it reports CRPS skill gains of 11.5% for 2 m temperature and 6.2% for 10 m wind speed on stations held out across space and time. The authors say topography explains more of the temperature structure, while the embeddings add surface information for wind. The gains also held when the coarse input came from Aurora forecasts and when predicting at newly deployed stations with no regional history. ArXiv · AI/CL/LG's note
score 4