PostCast: Generalizable Postprocessing for Precipitation Nowcasting via Unsupervised Blurriness Modeling

TL;DR

An unsupervised postprocessing method is proposed to eliminate the blurriness in precipitation nowcasting without the requirement of training with the pairs of blurry predictions and corresponding ground truth and demonstrates the generality and superiority of this method.

Abstract

Precipitation nowcasting plays a pivotal role in socioeconomic sectors, especially in severe convective weather warnings. Although notable progress has been achieved by approaches mining the spatiotemporal correlations with deep learning, these methods still suffer severe blurriness as the lead time increases, which hampers accurate predictions for extreme precipitation. To alleviate blurriness, researchers explore generative methods conditioned on blurry predictions. However, the pairs of blurry predictions and corresponding ground truth need to be generated in advance, making the training pipeline cumbersome and limiting the generality of generative models within blur modes that appear in training data. By rethinking the blurriness in precipitation nowcasting as a blur kernel acting on predictions, we propose an unsupervised postprocessing method to eliminate the blurriness without the requirement of training with the pairs of blurry predictions and corresponding ground truth. Specifically, we utilize blurry predictions to guide the generation process of a pre-trained unconditional denoising diffusion probabilistic model (DDPM) to obtain high-fidelity predictions with eliminated blurriness. A zero-shot blur kernel estimation mechanism and an auto-scale denoise guidance strategy are introduced to adapt the unconditional DDPM to any blurriness modes varying from datasets and lead times in precipitation nowcasting. Extensive experiments are conducted on 7 precipitation radar datasets, demonstrating the generality and superiority of our method.

Figures (6)

• Figure 1: Left: Previous methods require two stages to generate predictions with local weather patterns, which generate (GT, blurry prediction) pairs in stage 1 and apply these pairs to supervise the training of conditional generative models in stage 2. Right: We propose to directly train an unconditional DDPM to convert blurry predictions into the distribution of ground truths.
• Figure 2: Visualization of applying our PostCast on 5 datasets at time step 12 when the spatiotemporal prediction model is TAU.
• Figure 3: Left: Visualizations on SEVIR. Right: Visulizations on HKO7. Both blurry predictions are given by model TAU. The lead times of predictions are 30 minutes, 60 minutes, and 90 minutes.
• Figure 4: (a) The distribution of the mean of blur kernel at reverse step $t=0$ on SEVIR dataset. (b) The variation in the mean of the blur kernel with reverse steps on SEVIR dataset.
• Figure 5: (a) Variations in the guidance scale can result in different intensities of precipitation. From left to right, the guidance scale values are 1.25, 1, and 0.75 times of our auto-scale gradient guidance strategy; (b) Different initial values of the blur kernel parameters affect the resultant precipitation intensity maps to varying degrees. As the initial values of parameters decrease from left to right, the precipitation intensities correspondingly diminish. The intermediate map in the model output represents the model’s standard parameter settings, and the results are closest to the ground truth.