Downscaling Precipitation Forecast with Super-High Resolution

Downscaling weather forecast to high resolution (~3km) subsgrid scales using conventional numerical weather models faces the critical challenge from the physical parameterization of subgrid-scale processes, which in term leads to high computational cost as well as low skills scores.

In this study, a new downscaling method called Deep Hybrid Super Resolution (DHSR) is proposed to generate a super-high spatial resolution of numerical weather forecast. To effectively train the deep learning model, multi-source data fusion of the Shanghai local WRF model on 9km-resolution, the radar observation data as well as the satellite data are used as input.

Our model can downscale the input fields from 9km to 4km, even to 1km during day time with much less computational cost than conventional numerical weather models. It also shows significant improvement compared with other traditional statistical downscale methods.