MFC-RFNet: A Multi-scale Guided Rectified Flow Network for Radar Sequence Prediction

TL;DR

MFC-RFNet tackles the key challenges of radar-based nowcasting by unifying rectified flow generation with scale-aware feature communication, shallow-feature alignment, and frequency-guided skip fusion. The approach combines a four-scale U-KAN backbone with a conditional encoder, enhanced by three modules: FCM for cross-scale communication, CGSTF for displacement-driven shallow alignment, and WGSC for adaptive, wavelet-guided skip fusion, along with VRWKV blocks for long-range context. Empirical results on SEVIR, MeteoNet, Shanghai, and CIKM show consistent improvements in CSI and HSS, particularly at high rain-rate thresholds and longer lead times, while maintaining competitive MSE and reasonable compute. The findings indicate that integrating RF training with multi-scale communication, spatial alignment, and frequency-aware fusion yields a robust, efficient framework for radar nowcasting with potential extensions to higher resolutions and multi-sensor data streams.

Abstract

Accurate and high-resolution precipitation nowcasting from radar echo sequences is crucial for disaster mitigation and economic planning, yet it remains a significant challenge. Key difficulties include modeling complex multi-scale evolution, correcting inter-frame feature misalignment caused by displacement, and efficiently capturing long-range spatiotemporal context without sacrificing spatial fidelity. To address these issues, we present the Multi-scale Feature Communication Rectified Flow (RF) Network (MFC-RFNet), a generative framework that integrates multi-scale communication with guided feature fusion. To enhance multi-scale fusion while retaining fine detail, a Wavelet-Guided Skip Connection (WGSC) preserves high-frequency components, and a Feature Communication Module (FCM) promotes bidirectional cross-scale interaction. To correct inter-frame displacement, a Condition-Guided Spatial Transform Fusion (CGSTF) learns spatial transforms from conditioning echoes to align shallow features. The backbone adopts rectified flow training to learn near-linear probability-flow trajectories, enabling few-step sampling with stable fidelity. Additionally, lightweight Vision-RWKV (RWKV) blocks are placed at the encoder tail, the bottleneck, and the first decoder layer to capture long-range spatiotemporal dependencies at low spatial resolutions with moderate compute. Evaluations on four public datasets (SEVIR, MeteoNet, Shanghai, and CIKM) demonstrate consistent improvements over strong baselines, yielding clearer echo morphology at higher rain-rate thresholds and sustained skill at longer lead times. These results suggest that the proposed synergy of RF training with scale-aware communication, spatial alignment, and frequency-aware fusion presents an effective and robust approach for radar-based nowcasting.

Figures (11)

• Figure 1: Pareto plot of parameters, compute, and performance on SEVIR. Vertical axis: CSI-M (%); horizontal axis: number of parameters (Million). The color bar encodes computational cost in GFLOPs—darker indicates higher cost.
• Figure 2: MFC-RFNet overall architecture. Top: the conditional encoder extracts multi-scale features; FCM performs cross-scale communication; CGSTF aligns shallow features; WGSC modulates skip fusion. Bottom: the RF generator updates $z_t$ with few ODE steps ($z_{t+\Delta t} = z_t + v_{\theta}\,\Delta t$) to produce $K$ future frames, with VRWKV placed in deep layers to supply long-range context.
• Figure 3: Feature Communication Module (FCM).Left: Multi-directional fusion—top–down (TD), bottom–up (BU), and lateral branches are built at every level; an SE head conditioned on $\mathbf{F}_i^{\text{lat}}$ produces weights $\boldsymbol{\alpha}_i$ that mix the three streams into $\widehat{\mathbf{F}}_i$. Middle: Pixel-wise cross-scale communication—all $\widehat{\mathbf{F}}_i$ are first projected with $1{\times}1$ convolutions and bilinearly aligned to a reference resolution $(H^\star{\times}W^\star)$, concatenated, and fed to the attention head $g$ to obtain per-pixel scale softmax weights $\mathbf{W}_{\text{att}}$. For each level $i$, features from the other scales are fused by a convolutional head and gated by $\mathbf{W}_{\text{att}}[i]$ to yield a per-level enhanced map $\mathbf{R}_i^\star$ at the reference resolution. Right: Gated residual enhancement—each $\mathbf{R}_i^\star$ is adapted by the mapping head $\varphi_i$ to $\mathbf{R}_i$, modulated by the sigmoid gate map $\mathbf{G}_i$, and fused residually with the original features $\mathbf{F}_i$ to produce $\mathbf{F}_i^{\text{out}}$.
• Figure 4: Condition-Guided Spatial Transform Fusion (CGSTF). Conditioning features produce an offset field $\mathbf{O}$ (bounded by $\tanh$) to perturb the base grid $(\mathrm{Grid}_x,\mathrm{Grid}_y)$ and obtain $\mathrm{Grid}_{\text{offset}}$. GridSample warps the main features to $\mathbf{F}_{\text{warp}}$, which is fused with the conditional features to yield $\mathbf{F}_{\text{out}}$.
• Figure 5: The Wavelet-Guided Skip Connection (WGSC) architecture. Conditional features are processed by the Wavelet Processor (top inset) via 2D DWT to yield a frequency-aware guidance map ($\mathbf{A}_{\text{wav}}$). Concurrently (main path, bottom), spatial (SA) and channel (CA) attention masks are derived from concatenated encoder and decoder features. These components, along with $\mathbf{A}_{\text{wav}}$, inform the synthesis of three distinct gates. Encoder features are modulated by SA/CA gates, while decoder features are modulated by a wavelet-derived gate. An adaptive fusion mechanism combines these modulated streams based on a learned weight $\omega$, followed by a final convolutional refinement step to produce the output.