[Paper Review] CasCast: Skillful High-resolution Precipitation Nowcasting via Cascaded Modelling

Junchao Gong, Lei Bai+6 more|arXiv (Cornell University)|Feb 6, 2024

Precipitation Measurement and Analysis7 citations

TL;DR

CasCast is a cascaded framework that splits precipitation nowcasting into deterministic mesoscale prediction and probabilistic small-scale detail generation using a frame-wise guided diffusion transformer in a low-dimensional latent space, achieving strong performance on high-resolution radar datasets especially for regional extreme events.

ABSTRACT

Precipitation nowcasting based on radar data plays a crucial role in extreme weather prediction and has broad implications for disaster management. Despite progresses have been made based on deep learning, two key challenges of precipitation nowcasting are not well-solved: (i) the modeling of complex precipitation system evolutions with different scales, and (ii) accurate forecasts for extreme precipitation. In this work, we propose CasCast, a cascaded framework composed of a deterministic and a probabilistic part to decouple the predictions for mesoscale precipitation distributions and small-scale patterns. Then, we explore training the cascaded framework at the high resolution and conducting the probabilistic modeling in a low dimensional latent space with a frame-wise-guided diffusion transformer for enhancing the optimization of extreme events while reducing computational costs. Extensive experiments on three benchmark radar precipitation datasets show that CasCast achieves competitive performance. Especially, CasCast significantly surpasses the baseline (up to +91.8%) for regional extreme-precipitation nowcasting.

Motivation & Objective

• Motivate the need for skillful, high-resolution nowcasting across multiple precipitation scales (mesoscale deterministic motion and small-scale stochastic patterns).
• Propose a cascaded framework that decouples prediction of global precipitation distribution from generation of local, small-scale features.
• Enable high-resolution training while keeping computation tractable via latent-space diffusion modeling.
• Improve extreme-precipitation forecasts through frame-wise guided diffusion in a low-dimensional latent space.

Proposed method

• CasCast decouples forecasting into a deterministic component (predicts the mesoscale, high-determinism part) and a probabilistic component (generates small-scale patterns conditioned on the deterministic prediction).
• The probabilistic part uses diffusion models in a low-dimensional latent space to model small-scale variability, reducing training and sampling costs.
• A frame-wise guided diffusion transformer (CasFormer) is conditioned on a blurry deterministic prediction, with frame-wise encoding/decoding to ensure frame-to-frame correspondence and improve extreme-event optimization.
• Radar frames are compressed with a framewise autoencoder to enable high-resolution diffusion generation.
• CasFormer employs a frame-split mechanism and a sequence aggregator to produce latent vectors for diffusion denoising, integrating cross-attention with the observation latent to maintain spatiotemporal consistency.
• Training uses a diffusion-based denoising objective with classifier-free guidance and 1000-step noise schedule, 20 denoising steps for inference (DDIM).

Experimental results

Research questions

• RQ1 Can a cascaded modeling approach improve both mesoscale (deterministic) and small-scale (stochastic) precipitation nowcasting at high spatial resolution?
• RQ2 Does frame-wise guided diffusion in a latent space enhance extreme-precipitation forecasts compared to existing deterministic or probabilistic methods?
• RQ3 What is the impact of the deterministic backbone choice on cascaded forecasting performance and extreme-event accuracy?
• RQ4 How does the proposed CasCast perform across multiple radar datasets (SEVIR, HKO-7, MeteoNet) in terms of CRPS, CSI, and HSS?
• RQ5 What are the computational trade-offs of high-resolution training with latent-space diffusion versus full-pixel-space diffusion?

Key findings

• CasCast achieves competitive state-of-the-art results on three radar datasets, with notable gains in extreme regional precipitation nowcasting.
• The cascaded design yields substantial improvements in regional extreme metrics (CSI-181 and CSI-219) compared to baselines and some diffusion-only methods.
• Frame-wise guided CasFormer in a low-dimensional latent space provides improved optimization for extreme events while reducing computational cost.
• Ablation studies show large gains when replacing the deterministic backbone with stronger networks and validating the effectiveness of frame-wise guidance.
• Across SEVIR, HKO-7, and MeteoNet, CasCast demonstrates better distribution alignment (CRPS) and local pattern capture (CSI-M), as well as higher HSS and CSI scores in various pool settings.
• Compared to pure probabilistic models, CasCast’s two-stage setup reduces mode-collapse risk and enhances sharpness of local-structure predictions.