WADEPre: A Wavelet-based Decomposition Model for Extreme Precipitation Nowcasting with Multi-Scale Learning

TL;DR

WADEPre tackles extreme precipitation nowcasting under heavy-tailed distributions by moving to the wavelet domain, where a dual-branch architecture decouples deterministic large-scale advection (A-Net) from stochastic high-frequency convection (D-Net). A physics-aware Refiner then fuses these components to enforce consistency, while a multi-scale curriculum stabilizes training. Empirical results on SEVIR and Shanghai Radar show state-of-the-art performance, particularly for extreme events, with strong structural fidelity (SSIM) and CSI at high thresholds. The approach offers a principled balance between spatial localization, physical coherence, and high-frequency detail preservation, enabling more reliable operational forecasts of hazardous convective systems.

Abstract

The heavy-tailed nature of precipitation intensity impedes precise precipitation nowcasting. Standard models that optimize pixel-wise losses are prone to regression-to-the-mean bias, which blurs extreme values. Existing Fourier-based methods also lack the spatial localization needed to resolve transient convective cells. To overcome these intrinsic limitations, we propose WADEPre, a wavelet-based decomposition model for extreme precipitation that transitions the modeling into the wavelet domain. By leveraging the Discrete Wavelet Transform for explicit decomposition, WADEPre employs a dual-branch architecture: an Approximation Network to model stable, low-frequency advection, isolating deterministic trends from statistical bias, and a spatially localized Detail Network to capture high-frequency stochastic convection, resolving transient singularities and preserving sharp boundaries. A subsequent Refiner module then dynamically reconstructs these decoupled multi-scale components into the final high-fidelity forecast. To address optimization instability, we introduce a multi-scale curriculum learning strategy that progressively shifts supervision from coarse scales to fine-grained details. Extensive experiments on the SEVIR and Shanghai Radar datasets demonstrate that WADEPre achieves state-of-the-art performance, yielding significant improvements in capturing extreme thresholds and maintaining structural fidelity. Our code is available at https://github.com/sonderlau/WADEPre.

Figures (11)

• Figure 1: Breaking the Forecasting Dilemma. (a) The heavy-tailed distribution creates an optimization trap that ignores rare extremes. (b) Existing paradigms fail: pixel-wise methods blur details, while Fourier methods introduce ghosting artifacts. (c) WADEPre resolves this via wavelet decomposition, achieving sharp and accurate extremes.
• Figure 2: Schematic overview of the WADEPre architecture. The input sequence is decomposed via DWT into approximation ($\boldsymbol{A}_\text{seq}$) and details ($\boldsymbol{D}_\text{seq}$) coefficients. These components are processed by the dedicated Approximation Network (Encoder-Mixer-Decoder) and Details Network (Multi-scale FPN), respectively. The predicted coefficients are reconstructed via IDWT and fused by the Refiner to generate the final forecast $\boldsymbol{Y}_\text{pred}$. The green capsules indicate the loss functions applied during training.
• Figure 3: Architecture of the Approximation Network (A-Net). Designed to model deterministic low-frequency advection. The network employs a Temporal Injector (via 3D Convolution) to extract inter-frame dynamics from the input sequence $\boldsymbol{A}_\text{seq}$. The core evolution is driven by stacked Spatio-Temporal Blocks (STBlocks), which capture synoptic-scale spatial dependencies without loss of resolution.
• Figure 4: Architecture of the Detail Network (D-Net). Designed to resolve stochastic high-frequency convection. The network projects the detail coefficients $\boldsymbol{D}_\text{seq}$ into latent feature spaces via Temporal MLPs. A Feature Pyramid Network (FPN) backbone facilitates bidirectional cross-scale energy transfer, while the proposed Iterative Detail Refinement (IDR) module rectifies spectral inconsistencies to preserve sharp, localized boundaries in the final prediction.
• Figure 5: Temporal evolution of forecast skill for extreme events on the SEVIR. The curves visualize frame-wise CSI scores at high thresholds (CSI-181 and CSI-219) from 10 to 60 minutes. WADEPre demonstrates better long-term robustness than baselines.