Extending Precipitation Nowcasting Horizons via Spectral Fusion of Radar Observations and Foundation Model Priors

The spectral decomposition approach is empirically well-motivated: Figure 3 demonstrates that reconstructing radar fields using Pangu-Weather phases preserves spatial structure, validating the phase-similarity assumption underlying the Pangu-Weather-guided Frequency Modulation (PFM). The ablation study (Table III) rigorously validates that each component contributes distinctively—PFM improves skill metrics (CSI/HSS) while Frequency Memory (FM) reduces spatial errors (MSE/MAE)—and their synergy is necessary for peak performance. The loss function combining spatial MSE and spectral $L_1$ terms ($\mathcal{L}=\mathbb{E}[\|X_{1:K}-\hat{X}_{1:K}\|_2^2] + \lambda\mathbb{E}[\|\mathcal{F}(X_{1:K})-\mathcal{F}(\hat{X}_{1:K})\|_1]$) is appropriate for preserving high-frequency echo structure.

The evaluation excludes recent diffusion-based SOTA methods such as Prediff (NeurIPS 2023) and DiffCast (CVPR 2024), which are cited as references [23] and [24] but conspicuously absent from Tables I and II, weakening the state-of-the-art claim. The Frequency Memory module requires a complex two-stage training procedure ("training proceeds in two stages") where the memory bank is populated with ground-truth phases in Phase 1 and fixed during Phase 2 matching; this introduces training instability risks and convergence challenges that are neither analyzed nor justified with ablations. Additionally, the method's dependency on Pangu-Weather inference adds substantial computational overhead—generating 20 frames of meteorological variables at 0.25° resolution—that is not quantified or compared against baseline inference costs.

While the paper compares against 12 baselines including NowcastNet (Nature 2023), the exclusion of latent diffusion models—currently the dominant paradigm for precipitation nowcasting—makes the comparative analysis incomplete. The reported percentage improvements (e.g., 6.84% CSI increase) lack confidence intervals, statistical significance testing, or variance estimates across multiple training seeds, making it difficult to assess whether gains are statistically robust or subject to initialization variance. The qualitative results (Figures 6–7) show favorable case studies but provide no failure case analysis or discussion of meteorological conditions where spectral fusion might underperform spatial approaches.

The authors provide a GitHub repository link and architectural specifications (4 convolutional encoder-decoder modules, hidden depth $L=6$, $128\times 128$ resolution), but omit critical training details including batch size, total training epochs, wall-clock time, and GPU memory requirements. The Pangu-Weather preprocessing pipeline—which regrids 0.25° global forecasts to $32\times 32$ via "spatial and temporal interpolation"—lacks specifics on interpolation methods (bilinear vs. bicubic) and temporal alignment strategies that could affect reproducibility. The two-stage Frequency Memory training protocol lacks algorithmic pseudocode, details on the convergence criteria for each phase, or the rationale for selecting $S=240$ memory slots for SEVIR versus $S=160$ for MeteoNet beyond empirical sweeps.