Improved precipitation nowcasting using generative AI and data fusion
Nowcasts and forecasts of precipitation are essential to reduce the impacts of extreme rainfall on society. Precipitation nowcasting usually refers to short-term forecasts (up to a few hours) based on observations, such as KNMI’s calibrated radar product. State-of-the-art precipitation nowcasting uses generative AI models such as the Deep Generative Model of Radar (DGMR, Ravuri et al., 2021, Cambier van Nooten et al., 2023) or the diffusion-based LDcast (Leinonen et al., 2023). These generative models are desirable as they can produce spatially realistic and calibrated ensemble forecasts, which can communicate uncertainties to users. Additionally, recent architectures based on vision transformers (Bai et al. 2022) provide a promising avenue for building the models that incorporate more features for precipitation nowcasting.
In this project, our goal is to make skillful nowcasts for extreme rainfall events. One avenue of research for increasing nowcast skill in this area is through the inclusion of additional input variables, which can help the generative AI model learn the conditional distribution of extreme precipitation. KNMI has recently compiled a new data set containing such additional information. The new dataset contains many satellite-derived products that are known to be associated with rainfall, as well as forecasts from their high-resolution numerical weather prediction (NWP) model, Harmonie-Arome, that can explicitly resolve deep convection.
This project's aim would be to explore the added value of these potential input variables within deep learning approaches for precipitation nowcasting. One of our previous studies (Cambier van Nooten et al. 2023) showed that incorporation of temperature data helps in nowcasting extreme precipitation events, suggesting that adding more features is a promissing direction for this application. However, it is an open question as to which variables are most informative, which deep learning architectures can best leverage the new data, and what is the best way to include this additional information.
This project will be a follow up on previous work in our group.
The project will be supervised by Yuliya Shapovalova from Radboud University (and a team of researchers from KNMI).
Contact: Yuliya Shapovalova
References:
Cambier van Nooten, C., Schreurs, K., Wijnands, J.S., Leijnse, H., Schmeits, M., Whan, K. and Shapovalova, Y., 2023. Improving precipitation nowcasting for high-intensity events using deep generative models with balanced loss and temperature data: A case study in the Netherlands. Artificial Intelligence for the Earth Systems, 2(4), p.e230017.
Leinonen, J., Hamann, U., Nerini, D., Germann, U. and Franch, G., 2023. Latent diffusion models for generative precipitation nowcasting with accurate uncertainty quantification. arXiv preprint arXiv:2304.12891.
Ravuri, S., Lenc, K., Willson, M., Kangin, D., Lam, R., Mirowski, P., Fitzsimons, M., Athanassiadou, M., Kashem, S., Madge, S. and Prudden, R., 2021. Skilful precipitation nowcasting using deep generative models of radar. Nature, 597(7878), pp.672-677.
Bai, C., Sun, F., Zhang, J., Song, Y., & Chen, S. (2022). Rainformer: Features extraction balanced network for radar-based precipitation nowcasting. IEEE Geoscience and Remote Sensing Letters, 19, 1-5.