Speaker
Description
Commercial Microwave Links (CMLs) provide path-integrated rainfall estimates, but their irregular geometry poses a substantial challenge for deriving gridded rainfall maps.
In this contribution, we propose a generative AI approach for reconstructing rainfall fields from CML-derived rainfall estimates. The central idea is to learn characteristic precipitation structures from reference radar data and condition this with CML data. We investigate whether these learned spatial patterns can support the reconstruction of coherent rainfall patterns while remaining consistent with CML observations. In addition, we are currently developing an extension of this method that allows to combine the sparse CML data with dense gridded cloud observation from geostationary satellites (GEOs) to better constrain the rainfall field reconstruction in areas without good CML coverage. This work is based on our existing conditional generative adversarial network (cGAN) for quantitative precipitation estimation from GEOs data.
In this contribution we show skillful results of generative CML rainfall field reconstruction with a cGAN and a diffusion model. Compared to existing methods like Kriging the generated rainfall fields much better capture the real distribution of rainfall values. Furthermore, the generative approaches allow to create an ensemble of solutions. In addition we will show preliminary results of the generative model which combines sparse CML data and dense GEOs data.