DuoCast: Duo-Probabilistic Diffusion for Precipitation Nowcasting

1. Problem Statement

Precipitation nowcasting (short-term forecasting using radar data) is critical for agriculture, transportation, and disaster response. However, existing approaches face significant challenges:

• Numerical Weather Prediction (NWP): While physically grounded, NWP relies on solving complex partial differential equations, making it computationally expensive and too slow for real-time applications.
• Deterministic Deep Learning Models: These models (e.g., ConvLSTM, PhyDNet) predict mean values but often produce blurry outputs, failing to preserve fine-grained spatial details and high-frequency variability.
• Probabilistic Models: While they capture uncertainty, treating the entire system as fully stochastic can introduce excessive randomness, degrading forecast precision and temporal consistency.
• Hybrid Models: Recent attempts (e.g., DiffCast) combine deterministic and probabilistic approaches but struggle to preserve high-frequency information due to the limitations of Convolutional Neural Network (CNN) backbones in modeling high-frequency spectral components.

Core Challenge: Balancing global structural consistency (large-scale trends) with local detail preservation (micro-scale variability) under complex meteorological conditions.

2. Methodology: DuoCast Framework

The authors propose DuoCast, a dual-diffusion framework that decomposes the precipitation forecasting task into low-frequency and high-frequency components modeled in orthogonal latent subspaces.

A. Theoretical Foundation

The paper provides a theoretical proof that frequency decomposition reduces prediction error compared to standard single-branch U-Net diffusion models.

• Spectral Decay Limitation: The authors prove (via Lemma 1 and Theorem 1) that standard CNNs with bounded variation kernels exhibit polynomial Fourier decay, meaning they inherently struggle to generate or preserve high-frequency components.
• Irreducible Error: Theorem 2 demonstrates that a single CNN-based estimator cannot eliminate high-frequency errors.
• Orthogonal Decomposition: By splitting the target $Y$ into $Y_{low}$ (low-frequency) and $Y_{high}$ (high-frequency) using orthogonal projections, the framework allows two specialized models to approximate the target with zero irreducible error (Corollary 2).

B. Architecture Overview

DuoCast employs two distinct stochastic diffusion processes:

1. Low-Frequency Diffusion Model (Macro-Trends)

• Goal: Capture large-scale precipitation trends and weather front dynamics.
• Mechanism: Uses a convolutional backbone conditioned on weather front representations.
• Key Components:
  • Air Mass Spatial Modeling: Extracts spatial cues using three tailored convolutional layers:
    • ConvI: Large-kernel depth-wise convolutions for broad air mass context.
    • ConvII: Dilated convolutions for detecting occluded fronts and multi-scale interactions.
    • ConvIII: Pixel-wise temporal convolutions to encode evolution.
  • Weather Front Temporal Modeling: A lightweight cross-attention mechanism that relates current spatial features to historical structures (leading frame) and recent trends (previous frame) to model front evolution.
  • Intensity Awareness: Uses a thresholding operator to isolate high-intensity regions ($Y^\dagger$) to guide the modeling of strong air mass interactions.

2. High-Frequency Diffusion Model (Micro-Variability)

• Goal: Refine fine-scale details and localized precipitation intensities.
• Mechanism: Operates in the latent space using a self-attention-based architecture (similar to Stable Diffusion's U-Net but attention-focused).
• Conditioning: It takes the historical radar observations ($X$) and the coarse low-frequency forecast ($\hat{Y}_{low}$) as inputs.
• Advantage: Self-attention mechanisms dynamically reweight spatial features, overcoming the spectral limitations of CNNs to preserve high-frequency details.

C. Optimization

The models are trained using a two-stage schedule:

1. Train the low-frequency model first.
2. Jointly train both models.
   The total loss function combines the predictor loss, low-frequency diffusion loss, and high-frequency diffusion loss, weighted by hyperparameters $\lambda_1, \lambda_2, \lambda_3$.

3. Key Contributions

1. Dual-Diffusion Framework: Proposes DuoCast, which decomposes forecasting into frequency-specific subspaces, allowing specialized modeling of macro-trends and micro-variability.
2. Theoretical Proof: Provides a formal mathematical proof demonstrating that frequency decomposition reduces prediction error compared to standard single-branch diffusion models, specifically addressing the spectral limitations of CNNs.
3. Meteorology-Aware Design: Integrates specific meteorological priors (weather front dynamics and air mass interactions) into the low-frequency branch via specialized convolutional and attention mechanisms.
4. State-of-the-Art Performance: Demonstrates superior performance across four benchmark datasets, achieving significant gains in both spatial detail and temporal evolution.

4. Experimental Results

The model was evaluated on four radar datasets: SEVIR, MeteoNet, Shanghai Radar, and CIKM.

• Quantitative Performance:
  • DuoCast outperforms state-of-the-art baselines (including DiffCast, Prediff, AlphaPre, and NowcastNet).
  • Achieved 2–4% improvement in Critical Success Index (CSI) and 2–5% improvement in Heidke Skill Score (HSS) across datasets.
  • Showed particular strength in forecasting high-intensity precipitation (measured by CSI at high thresholds like CSI-181 and CSI-219).
• Qualitative Analysis:
  • Detail Preservation: Unlike deterministic models that produce blurry outputs, DuoCast preserves sharp edges and micro-scale variability.
  • Temporal Consistency: Unlike fully probabilistic models that introduce excessive randomness, DuoCast maintains consistent trend evolution (e.g., correctly predicting the expansion of a cold front or dissipation of a warm front).
  • Ablation Studies:
    • Removing the high-frequency model resulted in blurred predictions and loss of fine details.
    • Removing the weather front temporal modeling degraded temporal consistency.
    • The intensity-aware component ($Y^\dagger$) was crucial for predicting severe rainfall events, with its removal causing a ~27% drop in CSI-219.

5. Significance

• Bridging the Gap: DuoCast successfully bridges the gap between the structural consistency of deterministic models and the variability modeling of probabilistic approaches.
• Spectral Awareness: It addresses a fundamental limitation in deep learning for meteorology: the inability of standard CNNs to model high-frequency spectral components. By explicitly separating frequency bands, it overcomes the "blurry prediction" problem common in diffusion-based nowcasting.
• Practical Impact: The ability to accurately forecast both large-scale weather patterns and localized heavy rainfall events (flash floods) has direct implications for disaster response and public safety, as highlighted by the reference to the 2025 Texas flash floods in the introduction.

In summary, DuoCast represents a significant advancement in precipitation nowcasting by leveraging spectral decomposition theory and meteorological priors to achieve high-fidelity, probabilistic forecasts.