IMPA-Net: Meteorology-Aware Multi-Scale Attention and Dynamic Loss for Extreme Convective Radar Nowcasting

Technical Summary: IMPA-Net for Extreme Convective Radar Nowcasting

1. Problem Statement

Short-range (0–2 hour) radar nowcasting is critical for severe weather warnings. However, existing deep learning (DL) models for deterministic nowcasting face three primary challenges:

• Regression-to-the-Mean: Standard pixel-wise loss functions (like MSE) penalize sharp, high-amplitude predictions, causing models to produce "smooth" forecasts that suppress the intense convective echoes most vital for hazard detection.
• Scale Mismatch: Convective systems exhibit complex dynamics across multiple scales (Meso-$\beta$ organization vs. Meso-$\gamma$ cell-scale processes), which simple architectures often fail to capture.
• Suboptimal Feature Fusion: Most models treat heterogeneous geophysical inputs (e.g., terrain, wind, precipitation) as independent channels, failing to explicitly model the local physical interactions between them.

2. Methodology

The authors propose IMPA-Net (Integrated Multi-scale Predictive Attention Network), a deterministic framework that integrates meteorological knowledge through a coordinated design at the input, architecture, and training levels.

A. Input Level: Spatial Mixer
To address the fusion problem, the model uses a parameter-free Spatial Mixer. It partitions the input into $2 \times 2$ cells (corresponding to a $\sim$2 km Meso-$\gamma$ neighborhood) and performs a deterministic geometric permutation of the four input channels (radar, surface precipitation, topography, and along-slope wind). This interleaves information from all fields at the local scale, providing a structured cross-field prior for the encoder without adding trainable parameters.

B. Architectural Level: IMPA Module
The "translator" of the Encoder-Translator-Decoder paradigm is the IMPA (Integrated Multi-scale Predictive Attention) module, which consists of four stages:

1. Multi-scale Analysis: Parallel depthwise convolutions with varying kernel sizes ($3\times3, 5\times5, 7\times7$) to capture different spatial scales.
2. Global Relation Modeling: A self-attention mechanism to capture long-range spatiotemporal dependencies.
3. Intensity Calibration: A learnable, channel-wise scaling parameter ($\gamma$) that adaptively amplifies features associated with high-intensity events.
4. Detail Restoration: A residual block to refine fine-scale structures and storm boundaries.

C. Training Level: Meteorologically-Aware Dynamic Loss (MAD-Loss)
The model is trained using a multi-objective loss function that employs a three-level dynamic weighting strategy:

1. Epoch-adaptive: Weights shift from prioritizing extreme values and gradients early in training to prioritizing structural and temporal consistency later.
2. Storm-aware: Amplifies the loss for frames containing high-intensity pixels (exceeding 40 dBZ).
3. Lead-time-aware: Increases the penalty for underestimating intensity in later forecast frames to counteract the natural decay of skill over time.
   Additionally, the extreme loss component uses asymmetric weighting, penalizing the underestimation of strong echoes more heavily than overestimation.

3. Key Contributions

• Coordinated Design: Unlike previous works that address only one aspect (e.g., just the loss or just the architecture), IMPA-Net integrates input fusion, multi-scale attention, and dynamic loss into a single framework.
• Meteorological Integration: Incorporates terrain and climatological wind priors through a structured spatial mixer rather than simple concatenation.
• Extreme-Value Preservation: Specifically targets the "smoothing" problem of DL models through learnable intensity calibration and asymmetric loss.

4. Results

The model was evaluated on a multi-source radar dataset from eastern China against seven baselines (including SimVP, PredRNNv2, and the extrapolation method pySTEPS).

• Superior Severe-Echo Skill: At the $\geq$45 dBZ threshold (severe convection), IMPA-Net achieved a Heidke Skill Score (HSS) of 0.143, a 190% relative improvement over the SimVP baseline (0.049).
• Temporal Robustness: While most DL models lose all skill by 48–72 minutes, IMPA-Net maintains non-zero skill through the entire 120-minute window, exhibiting the slowest degradation rate among all tested methods.
• Structural Fidelity: Spectral analysis (RAPSD) confirmed that IMPA-Net preserves energy across both Meso-$\beta$ and Meso-$\gamma$ bands more effectively than competing DL models, maintaining more realistic spatial variability.
• Balanced Performance: Compared to the extrapolation method pySTEPS (which has high skill but extremely high false alarms), IMPA-Net provides a more practical trade-off between detecting severe events and controlling false alarms.

5. Significance

IMPA-Net demonstrates that meteorology-informed model design can significantly overcome the inherent limitations of pure data-driven regression in weather forecasting. By explicitly designing for the physical characteristics of convection—such as multi-scale organization and intensity gradients—the framework moves deterministic nowcasting closer to operational utility for severe weather warnings. While the current version uses static environmental priors, it sets a foundation for future models that incorporate dynamic, time-varying atmospheric variables.