Latent diffusion models for generative nowcasting and uncertainty quantification of precipitation fields

Deep generative modeling is able to generate highly realistic atmospheric fields, one prominent example being precipitation. So far, almost all studies have used generative adversarial networks (GANs) for this purpose, but recent progress in machine learning research has had a new class of methods called diffusion models replace GANs in many applications. Diffusion models have been often shown to be able to generate a wider variety of samples than GANs, suggesting that they might be able to better capture uncertainty in applications such as weather and climate where quantifying it is important.

In this presentation, we describe our research on using diffusion models for short-term prediction (nowcasting) of precipitation fields. We adapt the latent diffusion model used by Stable Diffusion (Rombach et al. 2022) to the this problem, predicting precipitation up to 3 hours ahead to the future at 5-min temporal resolution and 1-km horizontal resolution. Predictions can be produced as an ensemble where each member represent a possible future evolution of the precipitation field.

We show that our model:

• generates highly realistic precipitation fields that are consistent with the past precipitation used as input.
• outperforms the state-of-the-art GAN-based Deep Generative Models of Rainfall (DGMR) model by most relevant metrics.
• performs particularly well at representing the uncertainty of its own predictions, as shown by uncertainty quantification methods developed for ensemble forecast verification.

Therefore, it appears that diffusion models are indeed suitable for generative modeling of precipitation fields with highly realistic representation of uncertainty. Our model architecture also permits multiple inputs data sources to be combined, in particular allowing seamless generative predictions to be made by exploiting observations and numerical weather predictions.