[论文解读] Latent diffusion models for generative precipitation nowcasting with accurate uncertainty quantification

Jussi Leinonen, Ulrich Hamann+3 more|arXiv (Cornell University)|Apr 25, 2023

Precipitation Measurement and Analysis被引用 25

一句话总结

本文提出 LDCast，一种用于集合降水现在预报的潜在扩散模型，能够提供准确的不确定性量化和多样化 Forecast，在跨数据集的 CRPS 上超越 DGMR 和 PySTEPS。

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.

研究动机与目标

• Motivate short-term precipitation nowcasting and the need for realistic uncertainty quantification for downstream hydrological applications.
• Develop a latent diffusion model (LDM) based nowcasting system (LDCast) that operates in a latent space to improve training stability and efficiency.
• Enable generation of ensemble forecasts to capture predictive uncertainty and compare with existing methods.
• Demonstrate robustness and generalization across in-domain Swiss data and out-of-domain German data.

提出的方法

• Propose LDCast, a conditional latent diffusion model comprising a forecaster stack based on AFNOs with temporal cross-attention, a 3D-convolutional denoiser, and a variational autoencoder (VAE) to map between pixel and latent spaces.
• Use a PLMS-based iterative sampler (50 steps) to generate 32 ensemble members conditioned on past precipitation.
• Train the VAE first, then train the forecaster and denoiser jointly with an L2 loss in the diffusion framework.
• Preprocess precipitation with a logarithmic transform to emphasize relative errors and apply Gaussian antialiasing to mitigate artifacts.
• Evaluate on Swiss radar-based data (training domain) and German radar data (out-of-domain) using CRPS, LogCRPS, rank distributions, and Fractions Skill Score (FSS).
• Compare against DGMR (conditional GAN) and PySTEPS (statistical nowcasting).

实验结果

研究问题

• RQ1 Can a latent diffusion model reliably generate diverse和 realistic ensemble nowcasts of precipitation from recent radar observations?
• RQ2 Does LDCast provide improved uncertainty quantification (e.g., rank distributions close to uniform) compared to GAN-based and statistical baselines?
• RQ3 How does LDCast perform in terms of accuracy (CRPS) and event-based skill (FSS) across different spatial scales and lead times, including out-of-domain data?

主要发现

• LDCast outperforms DGMR and PySTEPS in CRPS across Swiss and German datasets, with larger gains at longer lead times and when averaging over larger scales.
• DGMR tends to produce overconfident ensembles with less diversity, while LDCast yields more diverse forecasts.
• Rank distributions for LDCast are closest to the ideal uniform distribution, indicating more realistic uncertainty representation (lower KL divergence from uniform).
• FSS results show mixed improvements for threshold exceedance skills; LDCast is competitive with DGMR and generally better than PySTEPS for Swiss data, with some regional variation in the German dataset.
• LDCast requires substantial but feasible training time (about 53 hours on 8 Nvidia V100 GPUs for 128x128 then ~5 hours for fine-tuning to 256x256) and ~19 seconds to sample one ensemble member on a GPU, highlighting a trade-off between training efficiency and generation cost.
• In-domain Swiss data show stronger CRPS and FSS advantages than out-of-domain German data, though uncertainty quantification remains robust across domains.