Latent diffusion models for generative precipitation nowcasting with accurate uncertainty quantification

TL;DR

This work addresses the challenge of short-term precipitation nowcasting with reliable uncertainty quantification. It introduces LDCast, a latent diffusion framework that uses a VAE to operate in latent space, a forecaster based on AFNOs, and a denoiser to generate ensembles of realistic precipitation fields conditioned on recent observations. Compared to DGMR (GAN-based) and PySTEPS, LDCast achieves higher probabilistic accuracy (CRPS) and substantially better uncertainty calibration (rank histograms near uniform), while maintaining competitive threshold-based forecast skill (FSS). The approach demonstrates robust performance across Swiss and German radar domains, highlights the benefits of latent-space diffusion for weather applications, and discusses practical considerations for training, sampling cost, and potential extensions to incorporate more predictors and physics constraints.

Abstract

Diffusion models have been widely adopted in image generation, producing higher-quality and more diverse samples than generative adversarial networks (GANs). We introduce a latent diffusion model (LDM) for precipitation nowcasting - short-term forecasting based on the latest observational data. The LDM is more stable and requires less computation to train than GANs, albeit with more computationally expensive generation. We benchmark it against the GAN-based Deep Generative Models of Rainfall (DGMR) and a statistical model, PySTEPS. The LDM produces more accurate precipitation predictions, while the comparisons are more mixed when predicting whether the precipitation exceeds predefined thresholds. The clearest advantage of the LDM is that it generates more diverse predictions than DGMR or PySTEPS. Rank distribution tests indicate that the distribution of samples from the LDM accurately reflects the uncertainty of the predictions. Thus, LDMs are promising for any applications where uncertainty quantification is important, such as weather and climate.

Figures (6)

• Figure 1: Sample cases of $256\ \mathrm{km} \times 256\ \mathrm{km}$ size comparing the precipitation rate observation and the prediction of the LDCast model. Time steps are produced by the model at $5\ \mathrm{min}$ resolution but they are visualized at $20\ \mathrm{min}$ intervals due to space constraints. The first ensemble member is shown in each case.
• Figure 2: CRPS (lower is better) for the LDCast model as a function of the forecast lead time, compared to the DGMR and PySTEPS benchmarks. The two top rows show CRPS for the absolute precipitation $R$ while the bottom rows show LogCRPS, i.e. CRPS for $\log_{10}(R)$. The three columns correspond to different amounts of averaging: no averaging ($1\ \mathrm{km}$ scale) for the first column, $8\ \mathrm{km} \times 8\ \mathrm{km}$ averaging for the second and $64\ \mathrm{km} \times 64\ \mathrm{km}$ for the third.
• Figure 3: Ensemble members of predicted precipitation at $90\ \mathrm{min}$ lead time. In each of four cases, the results from LDCast are shown on the first row on the left and the results from DGMR on the second row. The actual observed precipitation is shown for comparison on the right.
• Figure 4: Rank distributions for the LDCast, DGMR and PySTEPS models. The columns correspond to different averaging scales as with Fig. \ref{['fig:crps-leadtime']}. The numbers in the legend indicate the Kullback--Leibler divergence from the uniform distribution. The gray line in each plot indicates the ideal uniform distribution.
• Figure 5: FSS as a function of scale for LDCast, DGMR and PySTEPS. The three columns show the FSS for thresholds of $0.1\ \mathrm{mm\,h^{-1}}$, $1.0\ \mathrm{mm\,h^{-1}}$ and $10.0\ \mathrm{mm\,h^{-1}}$, respectively.